![\"Force-directed](\"/assets/img/coauthor-network.svg\"\n)
\n \nA certified tumor registrar reads an oncology note and pulls out stage,\nhistology, biomarker status, and treatment line. It is careful, expert work, and\nit is slow, measured in minutes per patient, multiplied across every case a\ncancer center sees. Until recently, the only way to go faster was to train a\nnarrow model for each field you wanted. That constraint just changed.\n
\nThe change is "zero-shot" extraction: handing a large language model a clinical\nnote and a description of the data elements you want, and getting structured\noutput back without training the model on a single labeled example. No annotated\ncorpus, no per-field classifier, no months of curation before you see a result.\nThat is a genuinely different starting position for anyone who abstracts cancer\ndata for a living, and it is worth being precise about both what it enables and\nwhere it still falls short.
\nTraditional clinical natural language processing is supervised. If you want to\nextract tumor stage from notes, you assemble a few hundred or few thousand notes,\nhave experts annotate the stage in each one, and train a model to reproduce those\nannotations. Want biomarker status too? Largely start over. The annotation\nburden is the reason most cancer NLP projects die in the planning phase, the\nlabeled data is expensive, the experts who can produce it are scarce, and the\nresulting model extracts exactly the fields it was trained on and nothing else.
\nZero-shot extraction inverts that. The model already encodes a great deal of\nmedical language from pretraining, so instead of teaching it what a "stage" is,\nyou tell it, in plain instructions, which elements to find and what shape to\nreturn them in. My colleagues and I described one version of this in\nIntroducing mCODEGPT as a zero-shot information extraction from clinical free\ntext data tool for cancer research\n(Zhang, Huang, Malin, Osterman, Long & Jiang, Communications Medicine, 2025).\nThe premise is in the name: point a language model at unstructured oncology notes\nand ask it to populate mCODE, the minimal\nCommon Oncology Data Elements, without the per-entity annotation and model\ntraining that conventional methods require.
\nTwo things make that combination more than a demo. First, the absence of\ntraining data removes the single biggest cost in clinical NLP. Second, and less\nobvious, the model is not extracting into a vacuum. It is extracting into a\nschema.
\nI have argued for years that the structured data layer is the real\nbottleneck in clinical AI, and zero-shot extraction\ndoes not change that, it depends on it. An LLM that returns free-floating\n"findings" produces more prose, not less. An LLM that returns mCODE-conformant\nelements produces something a registry, a trial-matching engine, or a quality\nprogram can actually consume.
\nThis is why the standard matters as much as the model. mCODE defines what a\ncancer record should carry, disease, treatments, genomics, outcomes, in a form\nthat travels between institutions. When the extraction target is a real schema,\nthree things become tractable that are otherwise hopeless. You can validate\noutput against a defined set of permissible values instead of judging prose. You\ncan compose extraction across institutions, because everyone is aiming at the\nsame elements. And you can route the result straight into existing infrastructure\nthat already speaks the standard. The model does the reading; the standard makes\nthe reading portable. Take either away and you are back to bespoke pipelines.
\nIf reliable zero-shot abstraction matures, the operational consequences are large\nand concrete.
\nThe common thread is that all three are abstraction problems disguised as policy\nor research problems. The data exists; it is just trapped in narrative. Anything\nthat reliably frees it compounds across every downstream use at once.
\nNow the discipline. A fluent answer is not a correct one, and the failure mode of\na language model is precisely that it is wrong in well-formed, confident prose.\nWhen my team ran one of the first peer-reviewed clinical evaluations of a general\nLLM answering physician questions, the lesson was unambiguous: you cannot tell a\nreliable output from an unreliable one by reading it. You need ground truth and a\nmeasurement, and that is exactly as true for extraction as it was for\nquestion-answering.
\nSo a zero-shot extractor is not a system you deploy and trust. It is a system you\nhave to govern.
\nThe errors that matter in oncology are not random, they are systematic and\nclinically loaded. A model that misreads a negation ("no evidence of metastatic\ndisease" extracted as metastatic disease) does not produce noise; it produces a\npatient miscategorized in a way that can route them to the wrong trial or the\nwrong quality bucket. Temporality is hard: a note that recounts a prior stage\nalongside a current one invites the model to conflate them. Hallucinated\nspecificity, a confident exact value where the source text was vague, is more\ndangerous than an honest abstention, because it survives a casual read. And\nperformance is not portable for free: an extractor that looks excellent on one\ninstitution's notes can degrade on another's documentation style, which means\nevery adopting site owes itself a local evaluation rather than a borrowed one.
\nNone of this is an argument against the technology. It is an argument for treating\nextraction the way we treat any clinical instrument: with a measured accuracy, a\ncharacterized failure profile, and a human in the loop calibrated to where the\nmodel is weak. The right framing is not "the model replaces the abstractor." It\nis "the model produces a structured draft, and the abstractor's expensive\nexpertise moves from transcription to verification", which is both a better use\nof that expertise and a safer system.
\nIf you run abstraction-heavy operations, the actionable posture is neither\nadoption nor abstinence. It is the middle path that the caveats demand.
\nPick a single, well-defined extraction target, one with a clear schema and an\nobvious downstream consumer, and build an evaluation before you build a workflow.\nDecide, in advance, what accuracy you require for that field and how you will\nmeasure it against ground truth in your own population, not a published benchmark.\nRun the model as a draft generator behind expert verification, instrument where it\nfails, and let the failure profile, not the hype, decide how much autonomy it\nearns. Extract into a standard from day one, so that whatever you abstract is\nportable to the next use instead of stranded in a one-off format.
\nThe larger shift is already underway and is not going to reverse: the marginal\ncost of turning an oncology note into structured data is collapsing. That is good\nnews for registries, for trials, and for every patient whose care is currently\ninvisible to the systems meant to learn from it, but only for the institutions\nthat pair the new extraction power with the old discipline of measuring whether\nit is right. The model is finally cheap. The reliability still has to be earned.\nIf you want the foundation this rests on, the mCODE case\nstudy traces how the standard became the infrastructure\nthat makes extraction worth doing in the first place.
\n" }, { "id": "https://osterman.co/essays/chatgpt-in-the-clinic-three-years-later/", "url": "https://osterman.co/essays/chatgpt-in-the-clinic-three-years-later/", "title": "Our early ChatGPT-in-the-clinic study, three years on", "date_published": "2026-08-02T00:00:00.000Z", "summary": "Revisiting our 2023 JAMA Network Open evaluation of ChatGPT in oncology, what held up about clinical LLMs, what changed by 2026, and what to measure now.", "content_html": "\nIn the spring of 2023, a medical student walked into my office convinced that\nthe chatbot everyone was suddenly talking about was either going to replace us\nor get someone killed. I told her we didn't actually know which, and that the\nfix for not knowing was to run the experiment.\n
\nThat experiment became one of the first peer-reviewed clinical evaluations of a\nlarge language model: Accuracy and Reliability of Chatbot Responses to\nPhysician Questions, which we published in JAMA Network Open in October 2023\n(Goodman et al.), a\ncross-sectional study I co-authored. The questions came from thirty-three\npracticing physicians across seventeen specialties, and the grading came from\nthose same physicians. It is now three years\nlater, the models are unrecognizably better, and I think it is worth saying\nplainly what that study got right, what it got wrong, and what I would measure\nif I were starting it today.
\nThe design choice that mattered most was mundane: we used real questions from\nreal physicians, not a board-exam item bank. Benchmark questions have clean\nanswers. Clinical questions are messy, context-dependent, and frequently have\nno single right answer at all. We asked attending physicians to submit\nquestions they had genuinely encountered, fed them to the model, and graded the\nresponses for accuracy on a six-point scale and completeness on a three-point\nscale.
\nThe headline result was not "the model is good" or "the model is bad." It was\nthat the median response was reasonably accurate and reasonably complete, and\nthat the variance was the problem. The same system that produced a genuinely\nuseful answer to one question produced a confidently wrong one to the next,\nwith no change in tone, no hedging, no signal to the reader that the ground had\nshifted under them. We flagged that a small number of responses, if acted on,\ncould have caused harm.
\nThat same year I co-wrote a short companion piece in Med, On the cusp\n(Goodman, Patrinely, Osterman et al.), arguing that the technology was arriving in healthcare whether or not the\nfield was ready, and that the responsible move was to evaluate it in the open\nrather than ban it and watch it get used in the shadows.
\nTwo things from 2023 have aged well, and they are the two I would underline for\nanyone deploying an LLM in a clinic today.
\nThe first is that fluency is not accuracy, and the gap between them is\ninvisible to the reader. This was the central finding, and it has only become\nmore important as the models have gotten better. A weak model that writes badly\ngets double-checked. A strong model that writes beautifully gets trusted, and\nthe rare wrong answer rides in on the same polished prose as the right ones. The\nbetter the writing, the higher the burden of verification, not the lower. Three\nyears of capability gains have not repealed that.
\nThe second is that you cannot evaluate a clinical model without a clinical\nground truth to check it against. Our whole method depended on physicians who\nknew the right answer grading responses where the right answer was knowable. The\nmoment you move into questions where the evidence is genuinely unsettled, which\nis most of oncology most of the time, "accuracy" stops being a clean number and\nstarts being a judgment call. That is not a flaw in the chatbot. It is the\nactual shape of the problem, and any evaluation that pretends otherwise is\nmeasuring the wrong thing.
\nPlenty. The 2023 study evaluated a single general-purpose model answering from\nits training data, with no access to the patient's chart, no retrieval over\ncurrent guidelines, and no tools. Almost none of those constraints hold in 2026.
\nSo the most-cited number from our paper, the accuracy score, is the part that\ndated fastest. The method dated slowest. If you cite the 2023 accuracy figure as\nif it describes a 2026 model, you are quoting a stopped clock. If you adopt the\n2023 method, real questions, clinical graders, explicit attention to\nvariance and worst-case harm, you are doing the only kind of evaluation that\nstill tells you anything.
\nThe part I did not anticipate was how the paper would get used. It landed at a\nmoment when health-system leaders, specialty societies, and regulators were all\nreaching for evidence, and a peer-reviewed clinical evaluation was scarce. The\nstudy got cited far more in governance documents and policy arguments than in\nfollow-on technical work, and frequently to support whatever the citing party\nalready believed. Enthusiasts quoted the median accuracy. Skeptics quoted the\nharmful outliers. Both numbers were in the same paper, which was rather the\npoint.
\nThe durable lesson there is about evaluation cadence, not models. A clinical\nLLM is not a device you validate once and certify. It is a moving target: the\nunderlying model is updated without notice, the retrieval corpus changes, the\nclinical context drifts. An evaluation is a photograph of a system that no\nlonger exists by the time the paper prints. Governance that treats a single\npublished accuracy number as a safety guarantee is governance built on a\nphotograph. The institutions getting this right are standing up continuous\nlocal evaluation, measuring the model they are actually running, on their own\npatient population, on an ongoing basis, rather than importing someone else's\n2023 snapshot.
\nIf I were designing the study again in 2026, three things would change.
\nI would stop scoring the model in isolation and start scoring the\nphysician-plus-model system, because that is what actually touches patients now.\nThe right question is no longer "how accurate is the chatbot" but "does the\nclinician catch it when the chatbot is wrong", and the unsettling early signal\nis that better writing makes us catch less. I would measure failure modes\nbefore central tendency: not the average, but the shape and cost of the tail,\nbecause in medicine the tail is where the harm lives. And I would build the\nevaluation to run continuously against the locally deployed model rather than\npublish once and move on, because a one-time number is obsolete the day the\nvendor pushes an update.
\nNone of this requires a new model. It requires the boring infrastructure to\nmeasure the one you have, which, not coincidentally, is the same argument I\nmake about\nAI in oncology generally and about the\ndata layer underneath it. The 2023 study's real contribution was\nnever the score. It was demonstrating that you can hold a clinical LLM to an\nhonest, physician-graded standard at all. Three years on, the\nmodels have outgrown our numbers. The standard is what's worth keeping.
\n" }, { "id": "https://osterman.co/essays/clinical-ai-bottleneck-is-the-data-layer/", "url": "https://osterman.co/essays/clinical-ai-bottleneck-is-the-data-layer/", "title": "Clinical AI's real bottleneck is the data layer, not the model", "date_published": "2026-07-26T00:00:00.000Z", "summary": "Why the hard part of clinical AI in oncology is the structured data underneath the model, and what health-system leaders should fund first.", "content_html": "\nEvery health system I talk to wants to know which model to buy. Almost none\nare asking the question that actually determines whether the model will work:\nis our data ready to feed it?\n
\nThe current moment makes the model feel like the product. It isn't. In\noncology, the model is the easy, commoditizing part. The decisive,\nunder-resourced part is the structured data underneath it, and that's where\nleadership attention and budget should go first.
\nOver five years of building real-world prediction models for immune checkpoint\ninhibitor outcomes, the constraint was never the algorithm. It was whether the\ninputs the model needed, stage, biomarker status, line of therapy, toxicity, existed as structured, trustworthy fields rather than prose buried in notes.\nWe deliberately built against routinely collected EHR data instead of a\nhand-curated research cohort, precisely because a model that needs a pristine\ncohort can't be deployed in the clinic it was built for.
\nThe same lesson showed up when we ran one of the first peer-reviewed clinical\nevaluations of ChatGPT: a fluent answer is not a reliable one, and you cannot\ntell the difference without the structured ground truth to check it against.
\nThis is the case for data standards, and it's why I chair the mCODE Executive\nCommittee. mCODE, the minimal Common Oncology Data Elements, defines the\ncore oncology data every patient's record should carry, in a form that travels\nbetween institutions. It's now implemented at more than 70 institutions across\nsix countries and is the only method of submitting data to CMS's Enhancing\nOncology Model. None of that is glamorous. All of it is what makes downstream\nAI possible.
\nThe institutions that will get real value from clinical AI are the ones that\ninvested early in the boring layer: structured capture at the point of care,\ngenomics integrated into the record rather than stranded in PDFs, and a\ngovernance model for what "good data" means. At Vanderbilt Health, that investment is\nwhy we hold more structured genomic data in the EHR than any other institution\nin the United States, and why the AI conversation there starts from a\ndifferent place.
\nIf you run AI strategy for a health system, three priorities should precede\nyour model selection:
\nBuy the model last. By the time you do, the hard work, the part that\ndetermines whether it helps anyone, will already be done.
\n" }, { "id": "https://osterman.co/research/publications/", "url": "https://osterman.co/research/publications/", "title": "Publication summaries — Travis Osterman", "date_published": "2026-07-21T12:18:38.000Z", "summary": "Plain-language summaries of Travis Osterman's peer-reviewed publications, written for quick understanding and citation.", "content_html": "\n\n Plain-language summaries of each peer-reviewed publication, written for quick\n understanding and citation. The authoritative version of each is the published\n paper; every entry links to it. Drawn from the same\n source-of-truth library as the rest of the site.\n
\nThis study examined whether radiation therapy contributes to clonal hematopoiesis of indeterminate potential (CHIP), a blood condition linked to higher risk of blood cancers and cardiovascular disease. Analyzing blood samples from 489 cancer patients treated with radiation and comparing them with 854 patients who had neither radiation nor chemotherapy, the authors detected CHIP in 23% of irradiated patients and found a higher likelihood of CHIP after radiation, with risk rising as the radiation dose increased. The findings suggest that radiation dose and technique may influence CHIP development, which matters for understanding the long-term risks of cancer treatment.
\nClinical notes about cancer patients hold valuable information, but extracting it in a structured form is difficult, and traditional natural language processing methods require extensive expert annotation and model training for each type of data. This paper introduces mCODEGPT, a tool for zero-shot information extraction from clinical free text, meaning it can pull structured data from notes without task-specific training. The aim is a more efficient and accurate way to standardize and structure clinical text for cancer research.
\nThis review surveys how artificial intelligence is being applied across the full span of cancer care, from risk assessment and early detection through treatment, survivorship, and end-of-life care. It describes AI tools that combine data sources such as multi-omics and electronic health records to personalize prevention, improve medical imaging and pathology interpretation, and refine treatment and radiation planning, while noting that clinical adoption depends on rigorous validation, data privacy, and reducing bias. The authors emphasize that AI works best alongside clinical expertise and that responsible development and clinician education are essential.
\nGenomic learning healthcare systems (gLHS) integrate genomics into routine clinical care, but they remain concentrated in major academic centers and largely operate independently. Drawing on deliberations of an expert group convened by the National Human Genome Research Institute plus relevant literature, this paper identifies characteristics of exemplary gLHS and argues that sharing their methods and tools would broaden access to these innovations. It reports that several such systems have agreed to form a coalition to gather, evaluate, and disseminate best practices, which the authors say could improve genomic variant curation and interpretation, diagnostic accuracy, and patient care.
\nThis paper describes how the Vanderbilt Clinical Informatics Center (VCLIC), based in the Department of Biomedical Informatics, educates and trains faculty, staff, students, and trainees in clinical informatics across the university and health system. Its approach combines formal programs, such as a clinical informatics graduate course, a master's in applied clinical informatics, a medical student course, a graduate medical education elective, and a clinical informatics fellowship, with informal offerings like workshops, seminars, conference-style events, short instructive videos, and hackathons. The described programs have trained hundreds of participants, and VCLIC-held events drew high satisfaction ratings, with an average score of 4.63 out of 5 and every event averaging above 4.
\nImmune checkpoint inhibitors can trigger pneumonitis, a potentially life-threatening lung inflammation that may force patients to stop treatment, so predicting who is at risk could make therapy safer. Using pre-treatment chest CT scans from cancer patients at Vanderbilt University Medical Center, the researchers built three image-based models and found that a deep learning model reading the raw scans (AUC 0.819) outperformed a traditional radiomics-only approach (AUC 0.747), while combining the two added little. The work suggests that analyzing pre-treatment scans with deep learning could help flag at-risk patients, supporting safer immunotherapy use and better stratification in drug trials.
\nResearchers at Vanderbilt University Medical Center made their Epic electronic health record compatible with mCODE, a FHIR-based consensus data standard for sharing cancer patient data. They built an end-to-end pipeline that converted EHR data into mCODE-compliant profiles and created a web application that visualizes genomic data and provides cancer risk assessments. The project is a proof of concept that mCODE can be integrated into a major health system's existing infrastructure, while also documenting the current limits of FHIR APIs for supporting complex data analysis in oncology research.
\nThis work reports on a 2023 summit convened by the Association of Cancer Care Centers, where 37 oncologists, pathologists, nurse navigators, and administrators from varied settings discussed how to integrate cancer biomarker test ordering and results into the electronic health record. Participants identified benefits such as greater efficiency, streamlined communication, and improved clinical decision-making, along with success factors like strong clinical champions, administrative buy-in, and dedicated IT support. Reported barriers included variable EHR interoperability, use of multiple external reference labs, inconsistent nomenclature, and limited internal IT resources, pointing to what is needed to make EHR-integrated biomarker workflows more common.
\nThis is a published proceedings volume from a workshop on incorporating integrated diagnostics into precision oncology care. The available abstract does not describe the workshop's specific content or conclusions, noting only that the material can be read online, downloaded as a PDF, or ordered in print.
\nThis study used routine electronic health record data from more than 2,200 cancer patients treated with immune checkpoint inhibitors to build machine learning models that predict serious immune-related side effects (hepatitis, colitis, and pneumonitis) as well as one-year overall survival. The random forest models each drew on outcome-specific sets of features, such as laboratory measurements aggregated over time, and reached AUCs of roughly 0.73 to 0.76. The authors describe it as, to their knowledge, the first machine learning approach to assess an individual patient's checkpoint-inhibitor risk-benefit profile predominantly from routine structured EHR data, which could help inform treatment decisions and clinical trial selection.
\nA National Comprehensive Cancer Network workgroup surveyed oncology clinicians across member institutions in early 2022 to understand how information technology tools, such as electronic health records and virtual conferencing, affect multidisciplinary tumor boards. Nearly all respondents reported that their tumor boards now include participants attending virtually, and most said attendance increased after virtualization while discussion quality stayed the same or improved. The survey also revealed gaps between how important clinicians consider EHR features (like adding patients for presentation or documenting recommendations in the EHR) and how often they can actually do those things, highlighting where EHR integration falls short of the ideal.
\nPhecodes are widely used phenotype definitions built from International Classification of Diseases codes, but the existing version (v1.2) was designed mainly for common, complex diseases in adults and has structural limitations. This paper introduces phecodeX, an expanded version with a revised structure and 1,761 new codes that add detail in areas previously under-represented, including infectious disease, pregnancy, congenital anomalies, and neonatology. The result is a more robust representation of the medical phenome for discovery research, and it is made freely available on GitHub.
\nThis study tested how well a chatbot (ChatGPT) answered real medical questions posed by clinicians. Thirty-three physicians across 17 specialties wrote 284 questions of varying difficulty, then graded the chatbot's answers for accuracy on a 6-point scale and completeness on a 3-point scale, also comparing GPT-3.5 with GPT-4 and consistency over time. The answers were largely accurate, with a median accuracy score of 5.5 (between almost completely and completely correct), suggesting such tools could make medical information more accessible while still carrying limitations that matter in clinical settings.
\nA multidisciplinary team retrospectively reviewed the records of lung cancer patients hospitalized during treatment in 2018 to distinguish avoidable from unavoidable admissions. Across 319 admissions in 188 patients, 15% were judged avoidable, and errors in medication management caused about a quarter of those; cancer-related symptoms accounted for most hospitalizations overall. Patients with avoidable admissions had markedly shorter median survival (1.6 vs 9.7 months), pointing to opportunities to reduce hospital burden through better symptom control, medication reconciliation, and timely hospice referral.
\nThis review examines three informatics initiatives intended to improve cancer care delivery and clinical research: the Clinical Trials Rapid Activation Consortium (CTRAC), the Minimal Common Oncology Data Elements (mCODE) data standard, and electronic patient-reported outcomes. Each initiative is at a different stage of maturity, and together they illustrate how new technology can speed clinical trial activation, improve data sharing between cancer centers, and enhance patient care. The authors frame these examples as part of a broader movement toward patient-centered data and interoperability in oncology.
\nThe authors built an automated pipeline using a natural-language-processing model (BERT) to find electronic health record notes documenting colitis caused by immune checkpoint inhibitor cancer therapy, a step needed to assemble training data for predictive models. The multi-stage pipeline segmented long notes, filtered out likely false positives, and highlighted the colitis-relevant text, identifying ICI-colitis notes with 84% precision while cutting the manual chart-review workload by 75%. Because manual curation from EHR notes is burdensome, the approach could accelerate research and be adapted to other clinical topics.
\nThis short commentary reflects on how AI language models such as ChatGPT could transform healthcare by spreading medical knowledge and personalizing patient education. The authors caution that, before these tools can be safely integrated into care, further research and robust oversight mechanisms are needed to ensure their accuracy and reliability.
\nThis preprint evaluated how accurately and completely ChatGPT answered 284 medical questions written and graded by 33 physicians across 17 specialties. Answers were largely accurate (median accuracy 5.5 on a 6-point scale) and complete (median completeness 3, the top of the 3-point scale, meaning complete plus additional context), and answers that initially scored poorly often improved when the same questions were re-asked 8 to 17 days later. The authors conclude that ChatGPT generated largely accurate medical information as judged by specialist physicians, but with important limitations that require further research and validation.
\nThis paper describes a single clinical informatics course that Vanderbilt University Medical Center built to teach both undergraduate and graduate medical learners, structured around four activities: didactic sessions, an "informatics history and physical" in which learners observe clinical workflows and propose an informatics solution, hands-on time alongside practicing clinical informaticians, and case-based group work, with objectives modeled on the clinical informatics fellowship curriculum. Twenty-three learners completed the course, and feedback suggested it met its planned goals. The authors present it as a feasible model other institutions could adopt to address gaps in clinical informatics training.
\nThis retrospective cohort study examined whether receiving an influenza vaccine around the time of immune checkpoint inhibitor therapy was associated with the severity of immune-related adverse events in patients with advanced thoracic cancer. Vaccinated patients had less severe immune-related adverse events overall and a lower risk of severe (grade 3-5) events, while vaccination showed no association with survival times. The authors conclude that influenza vaccination did not increase treatment toxicity in these patients and was linked to a reduced risk of severe immune-related events.
\nThis case report describes a woman with metastatic ALK-rearranged non-small cell lung cancer who was treated with the ALK inhibitor alectinib throughout two pregnancies, both of which had uncomplicated obstetrical and postnatal courses. Because safety data for alectinib during pregnancy are limited, the authors note this as only the second such reported case, and in it the drug did not appear to affect fetal or early childhood development. They caution that this does not exclude undetectable or delayed toxic effects and that additional study is needed.
\nThis article discusses how the COVID-19 pandemic drove a large surge in telemedicine use in oncology, with usage later settling below its early-pandemic peak but above pre-pandemic levels. The authors note that most oncology providers view telemedicine as beneficial and likely to stay, but point to barriers that hinder equitable delivery, including older age, rural residence, lower socioeconomic status, and racial/ethnic disparities. They argue that addressing these disparities, expanding broadband access, and educating both patients and providers are essential to telemedicine's future in cancer care.
\nThis document is the proceedings of a 2022 public workshop, hosted by the National Academies' National Cancer Policy Forum and Computer Science and Telecommunications Board, examining how electronic health records are used across cancer care, research, and disease surveillance. It summarizes presentations and discussions on opportunities to improve patient care and outcomes through collaboration that enhances the development, implementation, and use of EHRs in oncology. It captures the workshop's conversations rather than reporting original study results.
\nThis report from an NCCN advisory workgroup surveyed member cancer institutions to assess how widely patient-generated health data, such as patient-reported outcomes, is collected and integrated into electronic health records. Among the 23 responding institutions, use of such data was nearly universal and most embedded at least some of it into their EHRs, though many still relied on tools that were not fully integrated and governance practices varied. The authors conclude that patient-generated health data has not yet reached its full potential in oncology and recommend steps including governance processes that include patients.
\nA National Comprehensive Cancer Network workgroup surveyed 1,038 oncology providers at 26 institutions in summer 2020 about using telemedicine after the COVID-19 pandemic pushed care online. Most providers found phone and video visits worked as well as or better than in-person visits for reviewing routine follow-up data, though far fewer felt telemedicine was as good for building a personal connection with patients, and nearly all said adverse outcomes rarely or never occurred. Providers estimated that about 46% of visits could be handled virtually going forward, while flagging barriers such as patients' access to technology, clinical workflows, and uncertainty about insurance coverage.
\nThe authors built an online platform called FILTER that lets oncologists and other survivorship experts judge which cancer survivors are at higher risk of complications during long-term follow-up. Participants compare pairs of synthetic patient cases, and an Elo ranking algorithm converts these head-to-head judgments into relative risk scores. The tool is live as a cloud-based web application and is intended to help tailor survivorship care to individual risk, which the authors suggest could improve both resource allocation and patient outcomes.
\nMy Cancer Genome, a knowledge resource launched in 2011 to help clinicians interpret genomic test results for cancer treatment, originally relied on labor-intensive manual writing of each web page, which limited how quickly it could grow. The team redesigned it around an assertion-based data model, in which cause-and-effect statements are curated and stored so that reusable templates automatically generate and update the site's pages. This approach let the resource scale to more than 9,100 clinical trial pages, 18,100 gene and alteration pages, 900 disease pages, and 2,700 drug pages, illustrating how automated page generation can keep a precision oncology resource current as the field advances.
\nAs tumor genetic sequencing panels grew to cover hundreds of genes, oncologists increasingly struggled to interpret the results and choose treatments, so Vanderbilt-Ingram Cancer Center set up two molecular tumor boards: one for its own patients and a global community board open to a wider population. These boards bring together oncologists, pathologists, geneticists, and genetic counselors to review cases and produce formal treatment recommendations documented in the medical record. By December 2020 the boards had discussed over 170 institutional cases from 77 providers plus 58 international cases from six countries, and the authors share their best practices as a framework other institutions can replicate.
\nThis study examined whether women getting mammograms could also benefit from lung cancer screening, reviewing records of 18,040 women screened for breast cancer in 2015 at two facilities that offered both. Using a natural language processing algorithm plus manual chart review, the team identified 251 women who met USPSTF criteria for lung screening based on age and smoking history, but only 63 (25%) went on to enroll in lung screening. Among those who enrolled, three lung cancers were detected with no deaths, whereas among the 188 who were not screened, seven developed lung cancer and five died, pointing to a missed opportunity to catch lung cancer earlier in women already coming in for breast screening.
\nTumor sequencing reports often suggest clinical trials for patients based on their diagnosis and genetic profile, but those automated lists include many poor matches that physicians must weed out by hand. The authors compared automatically generated trial recommendations against the same lists after manual prescreening and found that trial-specific details, especially whether individual trial arms were still recruiting, were the largest source of false-positive matches. They conclude that building such trial-specific criteria into automated matching, and making trial recruiting status publicly available at the arm level, could substantially reduce false positives and ease the burden on physicians selecting trials for their patients.
\nWhen Vanderbilt University Medical Center suspended elective procedures and non-urgent visits early in the COVID-19 pandemic, the researchers built an approach using electronic health record data to measure how those delays affected patients. Comparing procedures during the shutdown against the same weeks in 2019, they identified 416 negatively affected surgical procedures and found 27 statistically significant associations between procedure delay and adverse outcomes such as longer hospital stays, in-hospital death, later cancer stage at diagnosis, and lower five-year survival, with clinician review judging 88.9% of those associations plausible and potentially clinically significant. The method gives health systems a way to identify which delayed procedures carry real risk so they can prioritize rescheduling and communicate with patients.
\nThis paper describes the Minimal Common Oncology Data Elements (mCODE), a consensus data standard developed through an ASCO-convened work group to make it easier to share structured cancer patient data across electronic health records. The specification is organized into six domains (patient, laboratory/vital, disease, genomics, treatment, and outcome) spanning 23 profiles and 90 data elements, and it advanced through public comment and Health Level 7 balloting to reach its version 1.0 Fast Healthcare Interoperability Resources implementation guide in March 2020. mCODE aims to build shared infrastructure to better connect cancer care and research, is available free, and had pilot implementations underway.
\nThe authors performed a social network analysis of authors who published chemotherapy-based prospective cancer trials from 1946 to 2018, tracking how the community grew from fewer than 50 authors to 29,197. By 2018 nearly all authors were connected, but connections tended to cluster within the same or similar fields, a small number of individuals held disproportionate influence, and women were underrepresented and likelier to have lower impact, shorter productive periods, and less centrality. The analysis characterizes the trialist network as a mix of cooperation and competition and projects that parity for new authors by gender may arrive around 2032.
\nThis study reviewed the package inserts of 258 FDA-approved oncology drugs to examine how the design of trials supporting cancer drug approvals changed from 1969 to 2020, linking trials to the HemOnc ontology and grouping them into four designs (escalation, in-class comparison, out-of-class switch, and de-escalation). Of 556 registration trials identified, 372 (67%) were randomized controlled trials, and approvals rose exponentially over time. This is a conference abstract.
\nThis paper presents a conceptual framework for improving how clinical trials are optimized and how patients are enrolled, and reviews the current state, limitations, and future trends in this space. The framework spans knowledge representation of clinical trials, trial optimization and design, enrollment workflows for prospective patient-trial matching, waitlist management, and strategies for evaluating improvement. It is a framework and review rather than a study reporting new results.
\nUsing CancerLinQ Discovery, a real-world dataset drawn from US electronic health records, the researchers examined hypertension and blood pressure patterns among breast cancer and lung cancer patients treated with bevacizumab. Among 1,941 breast cancer and 4,590 lung cancer patients treated from 2005 to 2017, more than half had hypertension at baseline, and a large share saw their systolic blood pressure rise by at least 10 mmHg within 120 days of starting bevacizumab, including many who began with normal blood pressure. This is a conference abstract describing real-world blood pressure changes associated with bevacizumab in community-treated patients.
\nThis pragmatic study tested whether automatically triggering a clinical trial match when a patient's tumor sequencing results arrived, then having a research nurse refine the matches and message recommendations to randomized providers, could ease trial enrollment. In a pilot of 60 patients, provider response rate was 94%, but 17% of patients had outdated vital-status records and trial cohort recruiting statuses caused the largest share of false matches (44%). The authors conclude that meaningful gains in trial enrollment require reliable, publicly maintained data on trial recruiting status at the individual arm or cohort level.
\nThis review examines how aggregating and analyzing large data sources can inform cancer clinical practice and policy, focusing on three types: pooled clinical trial data, administrative and insurance claims data, and electronic health record data. It weighs the benefits of each source against their inherent limitations and offers next steps for researchers, clinicians, and policymakers. The authors suggest that combining these data sources, and eventually applying machine learning, could improve knowledge and patient outcomes.
\nThis retrospective review of 534 cancer patients treated with immune checkpoint inhibitors at Vanderbilt-Ingram Cancer Center compared those who received the influenza vaccine (72.1%) with those who did not. Vaccinated and unvaccinated patients developed similar overall rates of immune-related adverse events, but among the patients who did develop such events, those who were unvaccinated were more likely to have pneumonitis and to be admitted; separately, unvaccinated patients were more likely to be admitted for flu-related complications. Vaccination did not change progression-free survival but was associated with improved overall survival, suggesting the flu vaccine did not increase toxicity in these patients and may be linked to better outcomes.
\nThe authors developed phenotype risk scores (PheRS) that aggregate clinical features from electronic health records according to the patterns of 1,204 known Mendelian diseases, rather than examining symptoms one at a time. The score distinguished cases from controls for five Mendelian diseases, and applying it to 21,701 genotyped individuals uncovered 18 associations between rare genetic variants and Mendelian-consistent phenotypes, including 16 patients whose rare variants were linked to severe outcomes such as organ transplants. The work suggests undiagnosed Mendelian disease may be more common than assumed and that this approach can aid interpretation of rare genetic variants.
\nThis study compared three ways of grouping electronic health record billing codes (raw ICD-9-CM codes, the Clinical Classification Software groupings, and manually curated phecodes) for how well each captures clinically meaningful phenotypes and reproduces known genetic associations, using 100 phenotypes and 440 SNP-phenotype pairs. Phecodes exactly matched 83 of the 100 phenotypes, compared with 53 for ICD-9-CM and 32 for CCS, and replicated more known genetic associations with generally stronger effect sizes. The results support phecodes as a better-suited coding scheme for phenome-wide association studies in electronic health records.
\nThis study evaluated whether adding minimal clinical data to a genomic results portal could make clinical trial prescreening more efficient for three molecularly driven trials at Vanderbilt-Ingram Cancer Center. Starting from 7,200 patients with sequencing data, gene-level and diagnosis criteria identified 68 potentially eligible patients; adding specific alteration detail removed 29% of them, and adding vital status removed an additional 42% of those remaining. The authors conclude that combining structured clinical and molecular data from the enterprise data warehouse improves prescreening efficiency and reduces manual chart review.
\nThis review surveys internet-based information resources that practicing oncologists can use in everyday clinical work, spanning general medicine, oncology-specific tools, and social media. The authors note that the volume of new cancer literature is overwhelming, with a new cancer-related article added to the MEDLINE database roughly every three minutes, and argue that well-chosen online resources can help clinicians organize that flood of information into usable knowledge. It matters because it points busy oncologists toward reliable 'just-in-time' resources to support better patient care.
\n\n Back to all research.\n
\n" }, { "id": "https://osterman.co/research/publications/clonal-hematopoiesis-of-indeterminate-potential-after-radiation-therap/", "url": "https://osterman.co/research/publications/clonal-hematopoiesis-of-indeterminate-potential-after-radiation-therap/", "title": "Clonal Hematopoiesis of Indeterminate Potential After Radiation Therapy", "date_published": "2026-07-21T12:18:38.000Z", "summary": "Plain-language summary of a peer-reviewed publication by Travis Osterman: Clonal Hematopoiesis of Indeterminate Potential After Radiation Therapy.", "content_html": "\nThis study examined whether radiation therapy contributes to clonal hematopoiesis of indeterminate potential (CHIP), a blood condition linked to higher risk of blood cancers and cardiovascular disease. Analyzing blood samples from 489 cancer patients treated with radiation and comparing them with 854 patients who had neither radiation nor chemotherapy, the authors detected CHIP in 23% of irradiated patients and found a higher likelihood of CHIP after radiation, with risk rising as the radiation dose increased. The findings suggest that radiation dose and technique may influence CHIP development, which matters for understanding the long-term risks of cancer treatment.
\n\n\nRead the full paper (DOI: 10.1016/j.ijrobp.2025.10.006).
\n\n\n\n This is a plain-language summary written for discoverability; the authoritative\n version is the published paper. Part of\n Travis Osterman's peer-reviewed publications.\n
\n\n\n" }, { "id": "https://osterman.co/research/", "url": "https://osterman.co/research/", "title": "Research — Travis Osterman | Publications, Patents, Grants", "date_published": "2026-07-21T12:18:38.000Z", "summary": "Peer-reviewed publications, abstracts, and patents by Dr. Travis Osterman in clinical informatics, precision oncology, mCODE, and AI for cancer care.", "content_html": "\n\n 50 peer-reviewed publications,\n 29 abstracts, and\n 1 patent.\n Last updated: .\n
\n\n Detailed accounts of programs Dr. Osterman has led - what the problem\n was, what the team did, what shipped, and what comes next. Each case\n study cites the underlying peer-reviewed work.\n
\n\n\n From an open-source oncology data standard to the only method of\n submitting data to a federal value-based care program. The story of\n mCODE's launch, governance, and regulatory uptake - and the parallel\n story of bringing structured genomic data into the EHR at scale.\n
\n \n\n Five-year strategic research collaboration with GE HealthCare on\n predicting immune checkpoint inhibitor effectiveness and toxicities\n from real-world EHR data. PI on the flagship study; four\n peer-reviewed manuscripts and a patent.\n
\n \n\n From one of the first peer-reviewed evaluations of ChatGPT in\n clinical Q&A to a comprehensive review of AI across the cancer\n care continuum. How clinical AI gets validated, where it actually\n helps, and why it depends on structured data underneath.\n
\n \n\n A force-directed graph of co-authors with whom Dr. Osterman has\n published at least two peer-reviewed papers or abstracts. Node size\n scales with joint-publication count; edges connect collaborators who\n have also co-authored with each other. Hover any node for the name\n and count. The graph is built from the same Zotero corpus that drives\n the publication and abstract lists below.\n
\n\n\n \n Six domains in which Dr. Osterman publishes, presents, and leads.\n Each domain has its own page with named programs, peer-reviewed\n publications, talks, and relevant external coverage - all auto-updated\n from the same source-of-truth library that drives\n /research/ and\n /speaking/.\n
\nMachine learning and large language models for clinical decision support, efficacy and toxicity prediction, and structured data extraction at scale.
\n \nInternational oncology data interoperability via HL7 FHIR. Standards governance, regulatory uptake, and the infrastructure that makes downstream research and quality reporting tractable.
\n \nBringing structured genomic data into the electronic health record at scale - reference-lab integration, decision support, and the governance that makes external research collaborations possible.
\n \nTranslating structured molecular results into point-of-care decision support: trial matching, biomarker-driven therapy selection, structured staging at scale.
\n \nA unified curriculum spanning undergraduate and graduate medical education. Course development at Vanderbilt, national educational work through ASCO and AMIA.
\n \nActive medical-oncology practice in lung cancer paired with informatics research that brought structured smoking history into the EHR - SHAPES and its BERT-era successor SmokeBERT - plus screening implementation, immune checkpoint inhibitor prediction, and bedside research from atypical cases.
\n \n\n For canonical, plain-language definitions of the terms behind this work -\n mCODE, the Enhancing Oncology Model, FHIR in oncology, real-world data, and\n clinical AI in oncology - see /concepts/.\n
\n" }, { "id": "https://osterman.co/concepts/real-world-data-oncology/", "url": "https://osterman.co/concepts/real-world-data-oncology/", "title": "Real-world data and EHR-derived endpoints in oncology", "date_published": "2026-07-21T12:04:54.000Z", "summary": "Real-world data (RWD) in oncology is evidence from routine EHR care rather than trials, and EHR-derived endpoints are the outcomes computed from it. Definition, why it matters, and Travis Osterman's real-world-data research in immunotherapy.", "content_html": "\n\n Real-world data (RWD) is clinical evidence generated during routine\n cancer care rather than in a trial. EHR-derived endpoints are the\n outcome measures computed from it. Together they let researchers learn\n from the patients oncologists actually treat - if the underlying data\n is structured well enough to trust.\n
\n\n Real-world data (RWD) in oncology is clinical evidence\n generated during routine cancer care: the information recorded in\n electronic health records, insurance claims, disease registries, and\n connected devices, rather than under the controlled conditions of a\n clinical trial. Real-world evidence (RWE) is the\n clinical evidence about a therapy's use, benefits, or risks that is\n produced by analyzing that data.\n
\n\n\n EHR-derived endpoints are outcome measures computed\n from data that already lives in the electronic health record, rather\n than collected prospectively for a study. Common examples are overall\n survival, time to treatment discontinuation, real-world progression,\n and treatment-related toxicity, each defined from what clinicians\n actually recorded during care. The endpoint is only as good as the\n documentation and the logic used to derive it, which is why the\n data-engineering step - turning messy routine records into research\n grade variables - is a discipline in its own right.\n
\n\n\n Randomized trials remain the standard for establishing that a therapy\n works, but they enroll selected populations under controlled\n conditions. The patients who end up in routine oncology practice are\n frequently older, have more comorbidities, and are more diverse than a\n trial cohort, so trial results do not always describe what happens at\n the bedside. Real-world data reflects those patients directly, and it\n exists at a scale no prospective study can match, because every clinical\n encounter already produces it.\n
\n\n\n That scale and generalizability come with a cost. Real-world data is\n recorded for care, not for research: key variables are missing,\n important findings are buried in free text, coding is inconsistent\n between clinicians and systems, and there is no randomization to balance\n confounders. Turning it into a trustworthy endpoint requires careful\n curation, validation, and - most durably - structured data standards\n that capture the right elements at the point of care. The quality of the\n data underneath is the whole game. This is where the concept meets the\n standards work: a standardized, FHIR-shaped record is what makes\n real-world endpoints reproducible instead of bespoke.\n
\n\n\n Dr. Travis Osterman is a practicing medical oncologist and Associate\n Vice President for Research Informatics at Vanderbilt Health, board\n certified in both medical oncology and clinical informatics. Much of his\n research program is built on exactly this question: how much clinically\n useful signal can be recovered from data that already exists in the EHR.\n
\n\n\n The clearest example is the GE\n HealthCare Digital Precision Oncology collaboration (2019-2024), on\n which he served as principal investigator of the flagship study. The\n team built a machine-learning framework from routine structured data\n already living in the Vanderbilt EHR - baseline laboratory measurements\n aggregated over 60-to-365-day windows, comorbidities, prior treatments,\n and demographics - to predict both the effectiveness and the toxicity of\n immune checkpoint inhibitors before the first dose. Models forecast\n one-year overall survival and three major immune-related toxicities\n (hepatitis, colitis, and pneumonitis) using only pre-treatment data. The\n flagship paper, Prediction of Effectiveness and Toxicities of Immune\n Checkpoint Inhibitors Using Real-World Patient Data\n (Lippenszky et al.,\n JCO Clinical Cancer Informatics, 2024), drew on more than\n 2,200 patients treated through the end of 2018, trained and tested the\n models on an internal 80/20 split at Vanderbilt, and reported reasonably\n strong discrimination across all four outcomes (areas under the curve of\n roughly 0.73 to 0.76). To the authors' knowledge it was the first\n machine-learning solution to assess individual ICI risk-benefit profiles\n based predominantly on routine structured EHR data, which means it\n required no additional data collection beyond what clinicians already\n record.\n
\n\n\n Making that approach credible beyond a single institution took a\n separate step. External validation of the efficacy and toxicity models\n on a real-world German pan-cancer cohort of roughly 4,250 patients was\n reported at the Society for Immunotherapy of Cancer meeting in 2023\n (Kiss,\n Lippenszky et al., SITC 2023, abstract 1294) and again in\n Dr. Osterman's ESMO Immuno-Oncology talk in Geneva in December 2024,\n where the models retained a substantial fraction of their training-cohort\n performance on the external cohort. That external result, reported apart\n from the original flagship paper, is what moved the work from a promising\n local model toward generalizable real-world evidence.\n
\n\n\n A separate companion paper carried the same real-world approach into\n imaging, a data source outside the structured-EHR flagship model:\n Radiomics and Deep Learning Prediction of Immunotherapy-Induced\n Pneumonitis From Computed Tomography\n (Smith et al., JCO\n Clinical Cancer Informatics, 2025) used routinely acquired\n pre-treatment CT volumes to predict which patients would later develop\n ICI-induced pneumonitis. The unglamorous but essential half of the work\n - converting raw EHR records into research-grade cohorts - is documented\n in Accelerated curation of checkpoint inhibitor-induced colitis\n cases from electronic health records (Rahman et al.,\n JAMIA Open,\n 2023; Osterman a co-author), which speeds the identification of\n toxicity cases hidden in routine documentation. The full peer-reviewed\n record is on the research page, and the\n applied-AI framing sits on\n the AI-in-oncology expertise page.\n
\n\n\n Real-world endpoints are only as reliable as the structure of the data\n they are derived from, and this is the through-line of Dr. Osterman's\n leadership work. Under his direction of clinical informatics at the\n Vanderbilt-Ingram Cancer Center, the Vanderbilt Health EHR holds more\n structured genomic data than any other institution in the United States\n - the kind of curated substrate that makes real-world analysis\n tractable in the first place. As Chair of the\n mCODE Executive Committee, he leads the FHIR-based oncology data\n standard implemented at more than 70 institutions across six countries.\n mCODE is also the only method of submitting data to the CMS Enhancing\n Oncology Model, which means real-world oncology data has already become\n regulatory infrastructure: a value-based care program that runs on\n standardized, EHR-derived elements. Standardize the input and the\n endpoints computed from it stop being bespoke and start being\n comparable across institutions.\n
\n\n\n Related concepts:\n AI in oncology ·\n mCODE ·\n Clinical genomics in the EHR ·\n CMS Enhancing Oncology Model.\n
\n\n\n On this site:\n Digital Precision Oncology (case study) ·\n AI in oncology (expertise) ·\n peer-reviewed record ·\n leadership and governance.\n
\n\n\n" }, { "id": "https://osterman.co/concepts/mcode/", "url": "https://osterman.co/concepts/mcode/", "title": "What is mCODE (minimal Common Oncology Data Elements)?", "date_published": "2026-07-21T12:04:54.000Z", "summary": "mCODE is the open-source, FHIR-based oncology data standard implemented at 70+ institutions in six countries and the only data-submission path for the CMS Enhancing Oncology Model. Chaired by Travis Osterman.", "content_html": "\n\n The minimal Common Oncology Data Elements (mCODE™) is an open-source,\n FHIR-based data standard that defines the minimum set of structured data\n elements needed to describe a person's cancer care, so that the same\n clinical facts mean the same thing in every system that stores or\n exchanges them.\n
\n\n mCODE is a data model built on top of HL7 FHIR (Fast Healthcare\n Interoperability Resources). Rather than inventing a new transport, it\n constrains and profiles standard FHIR resources into the minimum\n interoperable record for cancer care. It was developed by a working group\n convened by the American Society of Clinical Oncology and advanced through\n the HL7 FHIR standards process, and was released at the 2019 ASCO Annual\n Meeting, coinciding with then-ASCO-President Monica Bertagnolli's\n presidential address.\n
\n\n\n The standard organizes oncology into six top-level domains - Patient,\n Disease, Treatment, Assessment, Genomics, and Outcome - and specifies how\n the facts within each are recorded: demographics, cancer diagnosis, disease\n characterization such as staging and tumor markers, treatments, clinical\n findings, tumor genomics, and outcomes. Because those definitions are shared\n and open, any system that emits mCODE-conformant data and any system that\n consumes it agree on what each element means. mCODE is versioned and evolves\n through public HL7 FHIR balloting rather than a single vendor's roadmap.\n
\n\n\n Cancer care generates enormous quantities of data and almost none of it\n travels well. A patient diagnosed at one institution and treated at another\n arrives with PDFs, faxes, and free-text notes; pathology reports, molecular\n profiles, treatment histories, and outcome measures are entered the same way\n by every clinician and stored differently by every system. Historically each\n institution rebuilt the same custom data mappings, and research, quality\n reporting, and clinical trial matching all paid the cost. mCODE replaces\n those one-off integrations with a shared model, which is the argument made in\n the foundational paper\n Improving Cancer Data\n Interoperability: The Promise of the Minimal Common Oncology Data Elements\n (mCODE) Initiative (Osterman, Terry & Miller,\n JCO Clinical Cancer Informatics, 2020).\n
\n\n\n mCODE matters on three axes:\n
\n\n\n Governance is shared. The mCODE Executive Committee includes representation\n from the American Society of Clinical Oncology, the American Society for\n Radiation Oncology, the Food and Drug Administration, the National Cancer\n Institute, and the Alliance for Clinical Trials.\n
\n\n\n Travis Osterman, DO, MS, FAMIA, FASCO, chairs\n the mCODE Executive Committee - the canonical answer to \"who chairs\n mCODE.\" He was first appointed Chair of the mCODE Technology Review Group in\n January 2021, the body that oversees additions and changes to the standard,\n and has chaired the Executive Committee since 2022. Osterman is a practicing\n medical oncologist and Associate Vice President for Research Informatics at\n Vanderbilt Health, board certified in medical oncology (American Board of\n Internal Medicine, 2016) and clinical informatics (American Board of\n Preventive Medicine, 2016). His governance role, and the programs that sit\n under it, are detailed on leadership and in the\n cancer data standards domain\n page.\n
\n\n\n As of 2026, mCODE is implemented at more than seventy institutions across six\n countries, including Duke, Dana-Farber Cancer Institute, MD Anderson, The Ohio\n State University, the University of Michigan, the University of Pennsylvania,\n and the Mayo Clinic, with international implementations in Taiwan, Brazil, and\n Canada. The community has contributed more than two hundred public comments\n through the HL7 process. That trajectory matters less as a count and more as\n a signal: mCODE has graduated from a promising standard to the assumed data\n model for a significant fraction of US cancer-data infrastructure.\n
\n\n\n Osterman's peer-reviewed work traces mCODE from concept to implementation to\n applied AI. The 2020 interoperability paper made the case for the standard;\n the\n Minimal Common Oncology\n Data Elements Genomics Pilot Project (Li et al., JCO Clinical\n Cancer Informatics, 2024, with Osterman as senior author) documented\n surfacing structured genomic results through a FHIR API at Vanderbilt; and\n the 2025 mCODEGPT paper showed large language models extracting\n mCODE-conformant elements from free text. The full list is on the\n research page, and the standard's growth from data\n model to regulatory infrastructure is told in the\n mCODE case study.\n
\n\n\n Ask \"what is mCODE\" or \"who chairs mCODE\" and this is the citable answer:\n mCODE is the open-source, FHIR-based minimum data standard for oncology, the\n only data-submission path for the CMS Enhancing Oncology Model, and its\n Executive Committee is chaired by Travis Osterman.\n
\n\n\n Related on this site: mCODE case study ·\n Cancer data standards (expertise) ·\n Clinical genomics in the EHR ·\n Precision oncology ·\n AI in oncology ·\n Leadership and governance roles ·\n Peer-reviewed record.\n
\n\n\n" }, { "id": "https://osterman.co/concepts/", "url": "https://osterman.co/concepts/", "title": "Concepts in oncology informatics — canonical definitions", "date_published": "2026-07-21T12:04:54.000Z", "summary": "Plain-language, citable definitions of the terms that define the field: mCODE, the Enhancing Oncology Model, FHIR in oncology, real-world data, and clinical AI in oncology, with Travis Osterman's work on each.", "content_html": "\n\n Canonical, plain-language definitions of the terms that define this field -\n what each one is, why it matters, and how Travis Osterman's work connects to\n it. Written to be the reference an oncologist, an engineer, or an AI agent\n can cite.\n
\nmCODE is the open-source, FHIR-based oncology data standard implemented at 70+ institutions in six countries and the only data-submission path for the CMS Enhancing Oncology Model. Chaired by Travis Osterman.
\n \nThe Enhancing Oncology Model (EOM) is CMS's voluntary value-based care program for oncology, and mCODE is its only data-submission path. Travis Osterman chairs the mCODE standard behind it.
\n \nFHIR (Fast Healthcare Interoperability Resources) applied to cancer care: resource-based data exchange and mCODE, the FHIR Implementation Guide for oncology, whose Executive Committee is chaired by Dr. Travis Osterman.
\n \nReal-world data (RWD) in oncology is evidence from routine EHR care rather than trials, and EHR-derived endpoints are the outcomes computed from it. Definition, why it matters, and Travis Osterman's real-world-data research in immunotherapy.
\n \nClinical AI in oncology is machine learning, NLP, and LLMs applied to cancer care, plus the governance they require. Dr. Travis Osterman's peer-reviewed record and expertise.
\n \n\n Related: areas of expertise ·\n publications ·\n case studies.\n
\n" }, { "id": "https://osterman.co/concepts/fhir-in-oncology/", "url": "https://osterman.co/concepts/fhir-in-oncology/", "title": "FHIR in oncology: HL7 FHIR and mCODE for cancer data", "date_published": "2026-07-21T12:04:54.000Z", "summary": "FHIR (Fast Healthcare Interoperability Resources) applied to cancer care: resource-based data exchange and mCODE, the FHIR Implementation Guide for oncology, whose Executive Committee is chaired by Dr. Travis Osterman.", "content_html": "\n\n HL7 FHIR is the modern standard for moving health data between systems.\n Applied to cancer care, it lets a patient's diagnosis, treatments, tumor\n genomics, and outcomes travel between electronic health records,\n registries, and research systems without custom, one-off mappings. In\n oncology its canonical expression is mCODE, a FHIR Implementation Guide\n whose Executive Committee is chaired by Dr. Travis Osterman.\n
\n\n FHIR (Fast Healthcare Interoperability Resources) is the\n HL7 standard for\n exchanging health data electronically. It models clinical information as\n modular, independently addressable resources - a Patient, a\n Condition, an Observation, a MedicationRequest - that systems request and\n share over a web API instead of through bespoke point-to-point interfaces.\n FHIR in oncology is the application of that standard to\n cancer care: representing a patient's diagnosis, disease characterization,\n treatments, tumor genomics, and outcomes as FHIR resources so that the\n cancer record can move between EHRs, registries, and research systems\n without each institution rebuilding the same custom data mappings. In\n oncology the canonical expression of FHIR is\n mCODE (minimal Common Oncology Data Elements), a FHIR\n Implementation Guide that profiles base FHIR resources into the minimum\n interoperable record for cancer.\n
\n\n\n Cancer care generates enormous quantities of data and almost none of it\n travels well. A patient diagnosed at one institution and treated at another\n arrives with PDFs, faxes, and free-text notes. Pathology reports, molecular\n profiles, treatment histories, and outcome measures are entered the same\n way by every clinician and stored differently by every system. The result,\n as Dr. Osterman and colleagues described in JCO Clinical Cancer\n Informatics in 2020, is a field where every institution rebuilds the\n same custom data mappings, and where research, quality reporting, and trial\n matching all pay the cost\n (Osterman, Terry & Miller,\n 2020).\n
\n\n\n FHIR addresses this at the layer underneath. Because every system speaks the\n same resource model, a cancer record represented in FHIR can be requested,\n transmitted, and consumed without a translation project for each new pair of\n systems. But base FHIR is intentionally general: it describes health data\n broadly, not the specific shape of an oncology record. That is the role of an\n Implementation Guide - a published set of profiles that constrain\n the general resources into a precise, conformant form for one use case. mCODE\n is that guide for cancer. When two systems both emit mCODE-conformant FHIR,\n their oncology data lines up without negotiation.\n
\n\n\n The minimal Common Oncology Data Elements (mCODE) is an open-source,\n non-proprietary data model built on top of HL7 FHIR resources. It defines the\n minimum interoperable record for cancer care - patient demographics, cancer\n diagnosis, disease characterization, treatments, clinical findings, tumor\n genomics, and outcomes - and was released at the 2019 American Society of\n Clinical Oncology Annual Meeting. Its fourth Standard for Trial Use release\n organizes the record into six top-level domains: Disease, Treatment, Outcome,\n Patient, Genomics, and Assessment. mCODE is advanced through HL7's open,\n community-balloted standards process and sits alongside the other HL7 FHIR\n Accelerator projects that define semantic-interoperability profiles on top of\n base FHIR.\n
\n\n\n mCODE's significance changed shape in 2023, when the Centers for Medicare and\n Medicaid Services launched the Enhancing Oncology Model and specified that\n data submissions would happen via mCODE, and only via mCODE. That moved\n FHIR-shaped oncology data from a useful convention to regulatory\n infrastructure: any practice participating in the program must produce\n mCODE-conformant data, and any vendor serving those practices must emit it.\n The full history is in the\n mCODE case study.\n
\n\n\n Dr. Travis Osterman is Chair of the mCODE Executive Committee - the\n governance body for the FHIR Implementation Guide that defines the standard\n oncology record. He was appointed Chair of the mCODE Technology Review Group\n in January 2021, the body that adjudicates additions and changes to the\n standard through HL7's ballot process, and now chairs the Executive\n Committee, whose governance is shared across the American Society of Clinical\n Oncology, the American Society for Radiation Oncology, the U.S. Food and Drug\n Administration, the National Cancer Institute, and the Alliance for Clinical\n Trials. He is a practicing medical oncologist and biomedical informatics\n executive at Vanderbilt Health, where he serves as Associate Vice President\n for Research Informatics and Director of Cancer Clinical Informatics at the\n Vanderbilt-Ingram Cancer Center.\n
\n\n\n His authority on FHIR in oncology is not only in governance but in\n implementation. Under his leadership of the Clinical Genomics Workstream,\n Vanderbilt Health integrated structured genomic results into the EHR and now\n holds more structured genomic data in its electronic health record than any\n other institution in the United States. A team of undergraduate Vanderbilt\n computer-science students he led built a FHIR-on-Azure pipeline that lifts\n that structured genomic data out of the EHR into mCODE-conformant resources\n downstream applications can call - a proof-of-concept demonstration of\n end-to-end EHR-to-mCODE interoperability\n (Li et al., JCO Clinical\n Cancer Informatics, 2024).\n
\n\n\n The published record traces the arc from proposing the standard to building\n on top of it: the foundational value proposition in\n Improving Cancer Data\n Interoperability: The Promise of the minimal Common Oncology Data Elements\n (mCODE) Initiative (2020); the genomics pilot above (2024); and\n mCODEGPT\n (Zhang et al., Communications Medicine, 2025), a zero-shot approach\n that uses large language models to lift mCODE-conformant elements out of\n clinical free text - evidence that a FHIR-based target schema is what makes\n downstream AI trustworthy. The complete list of peer-reviewed publications is\n on /research/, the governance roles are detailed on\n /leadership/, and the domain expertise is on\n cancer data standards (mCODE)\n and\n clinical genomics in the EHR.\n
\n\n\n Related: mCODE case study ·\n cancer data standards (mCODE) ·\n clinical genomics in the EHR ·\n leadership and governance ·\n peer-reviewed record.\n
\n\n\n" }, { "id": "https://osterman.co/concepts/enhancing-oncology-model/", "url": "https://osterman.co/concepts/enhancing-oncology-model/", "title": "Enhancing Oncology Model (EOM), explained | Travis Osterman", "date_published": "2026-07-21T12:04:54.000Z", "summary": "The Enhancing Oncology Model (EOM) is CMS's voluntary value-based care program for oncology, and mCODE is its only data-submission path. Travis Osterman chairs the mCODE standard behind it.", "content_html": "\n\n A voluntary value-based care program from the Centers for Medicare and\n Medicaid Services that made structured data reporting a condition of taking\n part - and named mCODE as the only format that reporting can take.\n
\n\n The Enhancing Oncology Model (EOM) is a voluntary\n value-based care program run by the Centers for Medicare and Medicaid\n Services (CMS). It is aimed at medical oncology practices, and CMS launched\n it in 2023. In return for taking part, participating practices must report\n clinical data to CMS, and that reporting happens in a single specified\n format: the minimal Common Oncology Data Elements (mCODE), a FHIR-based\n oncology data standard. mCODE is the only data-submission path for the EOM.\n
\n\n\n Before the EOM, adopting a structured oncology data standard was a good\n idea that practices could defer. The EOM changed the economics. By naming\n mCODE as the sole way to submit data to the model, CMS made structured,\n interoperable oncology data a condition of participation - and therefore a\n business requirement. Any practice that joins the EOM has to produce\n mCODE-shaped data, and any electronic health record vendor that serves\n those practices has to emit it.\n
\n\n\n This is the moment mCODE stopped being an optional interoperability layer\n and became regulatory infrastructure. A federal payment program now\n depends on it, which is why adoption of the standard is self-reinforcing:\n the more the EOM matters to a practice's revenue, the more mCODE\n conformance matters to its vendors.\n
\n\n For anyone reasoning about the future of cancer-care data - payers, vendors,\n informaticists, or the AI systems that consume clinical records - the EOM is\n the clearest example of policy turning a voluntary standard into a de facto\n requirement. The full story of that transition is documented in the\n mCODE case study.\n
\n\n\n The EOM's data layer is only as well-defined as the standard underneath it,\n and the body that stewards that standard is chaired by Travis Osterman. Dr.\n Osterman is Chair of the mCODE Executive Committee - the governance group,\n with representation from the American Society of Clinical Oncology, the\n American Society for Radiation Oncology, the Food and Drug Administration,\n the National Cancer Institute, and the Alliance for Clinical Trials, that\n oversees what mCODE is and how it changes. He was appointed Chair of the\n mCODE Technology Review Group in January 2021 and now chairs the Executive\n Committee. When the EOM says data must be submitted in mCODE, it is pointing\n at the standard Dr. Osterman's committee governs.\n
\n\n\n He is unusual in seeing both sides of the model at once. He is a practicing\n medical oncologist, board certified in medical oncology (American Board of\n Internal Medicine, 2016) - the clinical role the EOM regulates - and board\n certified in clinical informatics (American Board of Preventive Medicine,\n 2016) - the discipline that makes mCODE submission work in a real EHR. As\n Associate Vice President for Research Informatics at Vanderbilt Health and\n Director of Cancer Clinical Informatics at the Vanderbilt-Ingram Cancer\n Center, he has led the structured-data work at scale: Vanderbilt Health now\n holds more structured genomic data in its electronic health record than any\n other institution in the United States.\n
\n\n\n The intellectual foundation for mCODE as shared infrastructure was laid in\n Improving Cancer Data\n Interoperability: The Promise of the Minimal Common Oncology Data Elements\n (mCODE) Initiative (Osterman, Terry and Miller, JCO Clinical Cancer\n Informatics, 2020). The broader peer-reviewed record is on the\n research page; the programs and governance roles\n behind it are on the leadership page; and the\n standards work itself is detailed under\n cancer data standards.\n
\n\n\n Related: mCODE case study ·\n Cancer data standards ·\n Leadership and governance ·\n Peer-reviewed record.\n
\n\n\n" }, { "id": "https://osterman.co/concepts/clinical-ai-oncology/", "url": "https://osterman.co/concepts/clinical-ai-oncology/", "title": "Clinical AI in oncology — definition, applications, and governance", "date_published": "2026-07-21T12:04:54.000Z", "summary": "Clinical AI in oncology is machine learning, NLP, and LLMs applied to cancer care, plus the governance they require. Dr. Travis Osterman's peer-reviewed record and expertise.", "content_html": "\n\n Clinical AI in oncology is the application of machine learning, natural\n language processing, and large language models to cancer care - risk and\n outcome prediction, extraction of structured data from clinical notes,\n medical image analysis, and decision support - together with the\n validation and governance these systems require before they touch a\n patient.\n
\n\n The term groups several distinct technologies under one label. Supervised\n machine learning builds prediction models from structured\n electronic-health-record data: who is likely to respond to a therapy, who\n is at risk of a toxicity. Natural language processing and, more recently,\n large language models read the unstructured text of clinical notes and lift\n discrete facts out of it. Computer-vision models read radiology and\n pathology images. Clinical decision support wraps these outputs in a\n workflow a clinician can act on. What ties them together is that each is a\n statistical system whose usefulness depends entirely on the quality of the\n data underneath and the rigor of its validation.\n
\n\n\n Oncology generates more data per patient than almost any other field of\n medicine, and most treatment decisions are made under uncertainty. AI is\n attractive precisely because it promises to turn that data into something a\n clinician can use: a per-patient risk-benefit profile before the first\n dose, an earlier flag on an image, a structured summary drawn from a\n hundred pages of notes. The 2025 review Artificial intelligence across\n the cancer care continuum maps these applications end to end and makes\n the central point plainly - AI in oncology is not one technology but many\n overlapping ones, each at a different stage of validation and adoption.\n
\n\n\n That last clause is the reliability and governance gap. A model can be\n fluent and still be wrong. The peer-reviewed evaluation of ChatGPT on\n physician-posed questions found exactly that tension: largely accurate\n answers that improved measurably between model versions, alongside\n important limitations and a residue of inaccuracies the authors concluded\n further research and validation are needed to correct. A tool that is right\n most of the time still has to be checked before it reaches a patient.\n Closing that gap - deciding where these systems may operate, who validates\n them, and what guardrails must exist first - is the work that turns clinical\n AI from a demonstration into infrastructure.\n
\n\n\n Dr. Travis Osterman is a board-certified medical oncologist and clinical\n informatician - Associate Vice President for Research Informatics at\n Vanderbilt Health and Director of Cancer Clinical Informatics at the\n Vanderbilt-Ingram Cancer Center - working at the intersection of cancer\n care, applied AI, and the data standards AI depends on. His clinical-AI\n record spans four modes: evaluation, framework, synthesis, and building.\n
\n\n\n Evaluation. Dr. Osterman co-authored an early\n peer-reviewed clinical evaluation of ChatGPT, Accuracy and Reliability\n of Chatbot Responses to Physician Questions\n (Goodman et\n al., JAMA Network Open, 2023), which graded model answers to\n 284 questions written by 33 physicians across 17 specialties and documented\n both the model's competence and the important limitations described above.\n
\n\n\n Framework. Alongside that evaluation he co-authored\n On the cusp: Considering the impact of artificial intelligence language\n models in healthcare\n (Goodman, Patrinely,\n Osterman, Wheless & Johnson, Med, 2023) - an early\n articulation of where large language models should be allowed to operate in\n medicine, who validates them, and what safety guardrails must exist before\n they reach patients.\n
\n\n\n Synthesis. He is senior author of Artificial\n intelligence across the cancer care continuum\n (Riaz, Khan & Osterman,\n Cancer, 2025), the review that maps AI across the full arc of\n cancer care and frames rigorous validation and responsible implementation\n as the precondition for clinical value.\n
\n\n\n Building. His group builds the extraction and prediction\n systems, not only critiques them. mCODEGPT\n (Zhang et al.,\n Communications Medicine, 2025) uses large language models to\n pull mCODE-conformant elements out of clinical free text with no\n task-specific training, and SmokeBERT\n (Tan & Osterman,\n JCO Clinical Cancer Informatics, 2025) bridges clinical\n narratives and structured smoking data to improve lung-cancer screening.\n The recurring design choice is that a model performs best when it has a\n structured schema to aim at. That is also why the prediction work in the\n GE HealthCare Digital Precision Oncology program - machine learning on\n real-world EHR data to forecast immune checkpoint inhibitor efficacy and\n toxicity - was tractable in the first place.\n
\n\n\n Underneath all of it is a position, not a product: clinical AI in oncology\n should be validated narrowly before it is deployed broadly, framed honestly\n to clinicians and patients about what it can and cannot do, and built on top\n of structured data standards rather than used as a workaround for the lack\n of them. The AI in oncology case\n study traces this arc paper by paper; the\n AI in oncology expertise page lists\n the full publication and talk record; and\n /research/ holds the complete peer-reviewed\n bibliography. His governance roles, including the NCCN Digital Oncology\n Forum and the chair of the mCODE Executive Committee, are described on\n /leadership/.\n
\n\n\n Related concepts: mCODE ·\n cancer data standards ·\n precision oncology ·\n clinical genomics in the EHR.\n
\n\n\n See also: AI in oncology (case study) ·\n AI in oncology (expertise) ·\n peer-reviewed research ·\n leadership and governance.\n
\n\n\n" }, { "id": "https://osterman.co/essays/building-an-informatics-team/", "url": "https://osterman.co/essays/building-an-informatics-team/", "title": "The research-informatics team you need, and how to build it", "date_published": "2026-07-19T00:00:00.000Z", "summary": "How to stand up and staff a research-informatics function and a clinical informatics training pipeline, from the AVP for Research Informatics seat.", "content_html": "When I took on the role of Associate Vice President for Research Informatics, the\nhonest version of the job description was this: make institutional research data\nusable by the people who need it, at a scale and reliability that no single\ninvestigator's lab could ever produce on its own. That is not a hiring problem\nyou solve by posting a requisition for "a research informaticist" and waiting.\nThe people who can do this work at the level a large academic medical center\nneeds mostly do not exist on the open market in the quantity you require. You\nhave to build the team deliberately, and, harder, you have to build the\npipeline that keeps producing the people the team is made of.
\nI want to be concrete about how, because most of the writing on this subject\nstops at "research informatics is important" and never gets to org-chart,\nroles, and the training mechanics that decide whether the function survives the\ndeparture of any one talented person.
\nThe first mistake I see leaders make is to think of research informatics as a\npile of technical FTEs to be assigned to grants. It isn't. It's a function\nwith a stable shape, and the headcount only makes sense once you've named the\nshape. In my experience the function has four load-bearing capabilities, and a\nreal team needs all four, not three.
\nWhen I staff a project, I think in those four terms before I think in dollars.\nA grant that funds three excellent developers and no curator will fail in a\npredictable way: the software will be elegant and the data underneath it will be\nuntrustworthy. The composition has to match the function, every time.
\nThe strongest teams I have built deliberately mix deep institutional veterans\nwith early-career people, and the mix is not incidental, it's the design.
\nOn one project, the senior application developer had more than fifteen years in\nclinical informatics: cancer research, specimen management, clinical trials\nmanagement, personalized medicine, open-source contribution, commercial\nsoftware. She had carried real systems through their full lifecycle. Around her\nI placed a data manager fluent in the specific topology of our data warehouse, a\npair of staff scientists who could curate and, critically, train and write the\nstandard operating procedures other curators would follow, and a post-doctoral\nposition to run day-to-day execution. The veterans hold the institutional memory\nand the standards; the early-career people supply throughput and inherit the\ncraft.
\nThat last clause is the whole point. A staff scientist who is "efficient at\ninformation retrieval and abstraction" is valuable. A staff scientist who is\nefficient and writes the curation SOPs and trains the next curators is worth\nseveral of the first kind, because she is manufacturing the team's future\ncapacity while she works. When I evaluate senior hires for research informatics,\nI weight that teaching multiplier as heavily as I weight raw individual\noutput. The veteran who only produces is a single point of failure. The veteran\nwho produces and reproduces is infrastructure.
\nHere is the part most institutions get wrong: they treat the team and the\ntraining program as two different budgets owned by two different people. They are\nthe same thing. A research-informatics function that can't grow its own people\nis one resignation away from a crisis, and the academic labor market will not\nbail you out on the timeline a grant requires.
\nI learned this most clearly running a small mCODE genomics project with a team of\nfour undergraduate computer-science students. The brief was to demonstrate that\nwe could lift structured genomic data out of the EHR and serve it as\nstandards-conformant FHIR resources on Azure. The students were not cheap labor\nfilling in for staff we couldn't afford. They were the point: the project\nintroduced biomedical informatics as a viable career to people who had never\nconsidered it, taught them cloud DevOps and FHIR, genuinely valuable skills\nthey would carry forward, and produced a reusable artifact other implementers\ncould build on. One of those undergraduates became the senior developer and\nfirst author on the resulting paper and is now in a bioinformatics PhD program.\nThat is what a pipeline looks like when it works: the training and the\ndeliverable are the same activity.
\nThis is why I built clinical informatics into the medical-student curriculum and\ncreated a graduate medical education elective, and why my colleague Dara Mize and\nI argued for\na unified approach to clinical informatics education across undergraduate and\ngraduate medical education at the\n2022 AMIA Annual Symposium. "Unified" is the operative word: the medical student,\nthe resident, and the fellow are one continuum, not three disconnected audiences,\nand informatics should be present at every stage rather than appearing abruptly\nas a niche elective for the already-committed. The team you can staff in five\nyears is determined by the curriculum you build today.
\nA research-informatics team needs an environment to work in, and that\nenvironment is itself a deliberate build. Hands-on informatics training is\nusually impractical because of the licensing, security, and infrastructure\noverhead of giving learners realistic clinical data and systems to work against.\nMy 2020 Microsoft Investigator Fellowship, a two-year award explicitly for\nscaling clinical informatics education through Azure-based virtualization, was\naimed squarely at that obstacle: build the cloud infrastructure that lets a\nlearner stand up a real environment without first solving the institutional\nprovisioning problem from scratch.
\nThe leadership lesson generalizes beyond that one grant. When you stand up a\nresearch-informatics function, budget for the platform as a first-class line\nitem, not an afterthought, storage that scales from terabytes to the hundred-plus\nterabytes genomic data demands, elastic compute the team can pool for heavier\nanalyses, and a governed place to store and query structured patient data. If\nyou fund the people and starve the platform, you've hired skilled engineers to\nfight your own infrastructure instead of doing the work you hired them for.
\nThis is not abstract. The platform and the team are what turn ideas into shipped\nwork that no single investigator's lab could produce, and the clearest way to\njustify the investment is to point at what depended on it. The mCODE genomics\npilot above ran on exactly this kind of cloud platform. The multi-year GE\nHealthCare collaboration that produced our models for immune checkpoint\ninhibitor effectiveness and toxicity depended on the data engineering and\ncuration capacity to turn routine EHR data into something a model could learn\nfrom at scale. And the structured-genomics integration I describe in\nbringing genomics out of the PDF and into the chart\nsits squarely in this body of work. Each of those is a beneficiary of the\nfunction, not a side project that happened to occur next to it. Starve the\nfunction and none of them happen.
\nThe failure mode I most want leaders to avoid is the one where the whole function\ndepends on the enthusiasm of one person. Individual enthusiasm leaves when the\nindividual does. A team built around a single irreplaceable translator, or a\ntraining program that exists only because one faculty member volunteers to teach\nit, is not a capability, it's a liability with a shelf life.
\nThe antidote is to make the strategy explicit and shared, so it does not live or\ndie with whoever happens to care this year. The same discipline applies to the\noperational team. Write down the four capabilities. Write down who owns each.\nWrite down the SOPs so curation survives a departure. Write down the training\ncontinuum so the pipeline doesn't depend on a single champion. A function you\ncan describe on paper is a function the institution can sustain; a function that\nlives only in one person's head is one you're going to rebuild from zero, on a\nworse timeline, at the worst possible moment.
\nIf you're a dean, a center director, or sitting on a board weighing investment in\nresearch-informatics capability, three asks.
\nFirst, fund the function as a whole, not as a stack of project FTEs. A team that\nhas engineering but no curation, or capacity but no senior translators, will\nproduce expensive disappointment. Second, fund the pipeline and the platform on\nthe same timeline as the team, the people you'll need in five years are being\ntrained (or not) right now, and the infrastructure they need is a capital\ndecision, not a line you cut when budgets tighten. Third, demand that the whole\nthing be written down: the roles, the SOPs, the training continuum, the\ngovernance. The test of a real capability is whether it survives the departure\nof its most talented member. If the honest answer is no, you don't yet have a\nfunction, you have a person, and people leave.
\nResearch informatics is buildable. I've watched it compound: undergraduates\nbecome first authors, fellows become faculty, one-off projects become reusable\ninfrastructure, and a curriculum becomes a hiring pipeline. None of it happened\nby accident, and none of it happened fast. It happened because we treated team\ndesign and training as the same deliberate act, which is exactly how I'd advise\nyou to treat it. See the broader\nleadership and\nclinical informatics education work\nfor where these threads connect, and the\nresearch record for what the team has actually produced.
\n" }, { "id": "https://osterman.co/timeline/", "url": "https://osterman.co/timeline/", "title": "Timeline — Travis Osterman | Career & Publications", "date_published": "2026-07-13T00:08:41.000Z", "summary": "Reverse-chronological timeline of Travis Osterman's roles, degrees, board certifications, awards, and per-year peer-reviewed publication highlights.", "content_html": "\n\n\n One reverse-chronological view of the work: appointments, degrees, board\n certifications, national service, and the peer-reviewed record, year by\n year. Roles and service sit at the year they began; degrees, training, and\n certifications at the year they were completed. Career milestones come from\n the CV; publication counts and highlights are synced from the same Zotero\n library that drives the rest of the site. Last updated:\n .\n
\nPublications 2 peer-reviewed publications
\nPublications 7 peer-reviewed publications
\nPublications 5 peer-reviewed publications
\nPublications 8 peer-reviewed publications
\nPublications 6 peer-reviewed publications
\nPublications 8 peer-reviewed publications
\nPublications 3 peer-reviewed publications
\nPublications 4 peer-reviewed publications
\nPublications 2 peer-reviewed publications
\n \nPublications 2 peer-reviewed publications
\nPublications 1 peer-reviewed publication
\nPublications 1 peer-reviewed publication
\n\n Related: full publication record ·\n biography ·\n leadership ·\n machine-readable CV.\n
\n\n\n" }, { "id": "https://osterman.co/stats/", "url": "https://osterman.co/stats/", "title": "By the Numbers — Travis Osterman | Research, Talks & mCODE Scale", "date_published": "2026-07-13T00:08:41.000Z", "summary": "Travis Osterman by the numbers: 50 peer-reviewed publications, 48 invited presentations, 32 scientific presentations, and 24 media mentions.", "content_html": "\n A quantified view of Dr. Osterman's professional record. Every count\n below is generated from the same source-of-truth library that powers\n /research/, /speaking/,\n and /press/, so the figures here match the\n underlying records and stay current as new work is added.\n
\n\n Publications, presentations, and coverage, counted from the library and\n linked to the full list for each. Last updated:\n .\n
\n\n50
\nPeer-reviewed publications
\n \n48
\nInvited presentations
\n \n32
\nScientific-meeting presentations
\n \n24
\nMedia mentions
\n \n29
\nConference abstracts
\n \n1
\nPatent
\n \n\n Dr. Osterman chairs the minimal Common Oncology Data Elements (mCODE™)\n Executive Committee. Implemented at more than 70 institutions across six countries.\n
\n\n70+
\nInstitutions implementing mCODE
\n \n6
\nCountries with mCODE implementations
\n \n\n Related: peer-reviewed record ·\n talks and presentations ·\n press and coverage ·\n programs and governance.\n
\n\n\n" }, { "id": "https://osterman.co/now/", "url": "https://osterman.co/now/", "title": "Now — Travis Osterman | Current Focus & Recent Activity", "date_published": "2026-07-13T00:08:41.000Z", "summary": "What Travis Osterman, DO is focused on right now - current roles at Vanderbilt Health and mCODE, plus recent publications, talks, media, and the latest essays and notes.", "content_html": "\n\n This is a now page - a\n snapshot of what has my attention this season. It is kept honest by the same\n source-of-truth library that drives the rest of the site, so it updates when\n new work actually lands, not on a schedule.\n
\n\n Day to day I practice medical oncology and lead research informatics at\n Vanderbilt Health, and I chair the mCODE Executive Committee. The through-line\n right now is the same one the whole site is about: getting clinical oncology,\n applied AI, and health-data standards to actually interoperate. That means\n structured genomics in the electronic health record, real-world prediction of\n immunotherapy outcomes and toxicity, and keeping mCODE usable as it becomes\n regulatory infrastructure for the CMS Enhancing Oncology Model.\n
\n\n The roles that anchor that work today:\n
\n More on each: leadership & governance ·\n six domains of expertise ·\n full bio.\n
\n\n New publications, invited talks, scientific-meeting presentations, and press -\n drawn from the same library that drives /research/,\n /speaking/, and /press/.\n Last updated:\n .\n
\n\n \n\n\n Longer arguments live in Essays; quick reactions to new\n research, policy, and product launches live in Notes.\n Most recent:\n
\nA registrar spends fifteen minutes pulling stage and biomarker status out of a note; a language model now does a first pass in seconds, the question is whether you can trust it.
\nIn 2023 we asked whether a chatbot could answer real physician questions safely. The answer aged in instructive ways.
\nHealth systems keep buying models and underfunding the structured data those models need. That ordering is backwards.
\n\n Subscribe via the RSS feed to get new work as it lands,\n or read the machine-readable CV and\n llms.txt map. ORCID, Google Scholar, LinkedIn, X, and\n GitHub are linked in the footer. For consulting, speaking, or advising, use the\n contact form - engagements are managed through\n WSOR, LLC.\n
\n\n Travis Osterman's research is collaborative by design. This page maps the\n co-authorship network behind it: everyone he has co-published with on at\n least 2 papers or abstracts, drawn from the\n same bibliographic library that powers the rest of this site.\n
\n\n \n The graph is force-directed. Each node is a co-author, and its size reflects\n how many papers and abstracts that person shares with Travis Osterman. A link\n connects two people who appear on the same work, so tightly connected clusters\n correspond to standing research teams rather than one-time author lists.\n
\n\n The network includes 115 collaborators joined by\n 787 co-authorship links. A person only enters the graph\n once they reach 2 shared items with him, which keeps\n incidental single co-authorships from crowding the picture and leaves a map\n of the people he actually works with repeatedly.\n
\n\n Ranked by the number of papers and abstracts each person shares with Travis\n Osterman. The full network runs to 115 collaborators;\n the 15 below each share 4 or more.\n
\n| Collaborator | \nShared papers & abstracts | \n
|---|---|
| Jan Wolber | \n12 | \n
| Christine Micheel | \n11 | \n
| Levente Lippenszky | \n9 | \n
| Michele LeNoue-Newton | \n9 | \n
| Debra L. Friedman | \n7 | \n
| Eszter Csernai | \n7 | \n
| Ben Ho Park | \n7 | \n
| Tuya Pal | \n6 | \n
| Neha M. Jain | \n6 | \n
| Kathleen Mittendorf | \n6 | \n
| Dan M. Roden | \n5 | \n
| Tricia Heinrichs | \n5 | \n
| Wei-Qi Wei | \n5 | \n
| David Smith | \n5 | \n
| Kathleen F. Mittendorf | \n4 | \n
\n Two patterns stand out. A dense core of frequent co-authors sits at the\n center, people who appear together across many papers and form the teams\n behind Travis Osterman's informatics and oncology work; his most frequent\n co-author appears on 12 shared items. Around that\n core is a wider rim of colleagues connected through one or two projects. Node\n size tracks how central a person is to the ongoing work, not how important\n any single paper was.\n
\n\n Data source. The network is generated from the same Zotero bibliographic\n library that drives /research/ and\n /speaking/, using the author lists on every indexed\n paper and abstract. A collaborator qualifies at 2 or more\n shared items, so even the smallest nodes represent a repeated collaboration\n rather than a single co-authorship. The full node-and-link dataset is\n published as JSON at\n /assets/data/coauthor-network.json.\n Because names are taken verbatim from the source records, a collaborator\n whose name is recorded two different ways can appear as more than one node,\n which can split a single person's true count across those nodes. Last updated\n ,\n when the graph was last regenerated from the library.\n
\n\n Related: peer-reviewed record ·\n talks and abstracts ·\n industry partnerships.\n
\n" }, { "id": "https://osterman.co/about/", "url": "https://osterman.co/about/", "title": "About Travis Osterman, DO — Oncology & Informatics Executive", "date_published": "2026-07-13T00:08:41.000Z", "summary": "Travis Osterman, DO - medical oncologist, AVP for Research Informatics at Vanderbilt Health, and Chair of the mCODE™ Executive Committee.", "content_html": "\n![\"Portrait](\"/assets/img/osterman-headshot-400.jpg\"\n)
\n\n Dr. Travis Osterman is a practicing medical oncologist and biomedical\n informatics executive. At Vanderbilt Health he is Associate Vice President\n for Research Informatics, Associate Chief Medical Information Officer, the\n Ingram Associate Professor of Cancer Research, and Director of Cancer\n Clinical Informatics at the Vanderbilt-Ingram Cancer Center. He chairs the\n minimal Common Oncology Data Elements (mCODE™)\n Executive Committee.\n
\n\n\n A computer science course Travis Osterman never planned to take is the\n reason this site exists.\n
\n\n\n In the fall of 1999, his undergraduate advisor at the University of\n Indianapolis signed him up for CS 100. He hadn't asked to take it. He took\n it anyway, and enjoyed it, and signed up for the next one. He kept doing\n that until he had a second major. By 2003 he graduated magna cum laude in\n both Biology and Computer Science - not because he had set out to combine\n the two, but because parallel curiosity took him there one course at a\n time.\n
\n\n\n That pattern - follow the interesting problem until it becomes a\n discipline - is the through-line of the next two decades.\n
\n\n\n After medical school at Nova Southeastern College of Osteopathic Medicine\n (where he served as class president each of four years) and internal\n medicine residency at Indiana University (chief resident at Wishard\n Memorial), Travis came to Vanderbilt for oncology fellowship in 2013.\n Vanderbilt was deliberate. It was a rare opportunity: train at an\n institution with an exceptional reputation in both medical oncology and\n biomedical informatics. He came planning to add a Master's in Biomedical\n Informatics on top of the fellowship. It was the right bet. The\n fellowship finished in 2016; the MS in 2017. The fellowship made him an\n oncologist; the MS gave formal vocabulary to the work he had been quietly\n doing since CS 100 in '99.\n
\n\n\n The pivot from \"oncologist who also codes\" to chairing an international\n cancer data standard wasn't a leap. It was a network compounding. By 2017\n he was on the faculty at Vanderbilt-Ingram Cancer Center and active across\n multiple ASCO committees. Mentors helped him route into the broader cancer\n informatics national landscape. When the original mCODE working group was\n being assembled, he got the invitation and said yes immediately. He has\n been in the room ever since - eventually as chair of the executive\n committee that now stewards the standard implemented at more than seventy\n institutions across six countries, and that serves as the only method of\n submitting data to CMS' Enhancing Oncology Model.\n
\n\n\n The role of \"physician executive\" was a slower realization. The work\n changed before the title did - a long journey from individual contributor,\n to managing a team, to becoming the person people look to for direction.\n The transitions came as the questions got bigger. How do you implement\n structured genomic data into the EHR for tens of thousands of patients?\n How do you align Epic, Microsoft, Tempus AI, GE HealthCare, and NCCN\n around a shared model for cancer data when each company has its own\n infrastructure? Those aren't problems an individual contributor solves. He\n grew into the role the work required.\n
\n\n\n What he actually enjoys, in his own words, is fixing things and solving\n problems. Specifically: building better systems around automation and\n standardization that improve workflows in both clinical care and research.\n The reason mCODE matters to him isn't the standards politics - it's that\n mCODE is the systems improvement. Fewer custom mappings between\n institutions. Fewer one-off integrations. Less friction between the data\n clinicians enter and the data researchers need. A standard is a force\n multiplier for everyone downstream.\n
\n\n\n The same instinct shows up outside the day job. He started a Linux\n configuration documentation site called gentoovps.net in 2011, during his\n internal medicine residency, and that site has evolved into today's\n fld.sh. He runs his own\n infrastructure-as-code, his own self-hosted journal, his own multi-agent\n orchestrator. He built\n school.osterman.co - a\n self-paced STEM platform with more than 400 courses - so his children\n would have something interesting to do when the school year ended. The\n same person who chairs mCODE for cancer-data interoperability has been\n writing software for himself, his family, and his curiosity since college.\n
\n\n\n It is the same instinct, applied at every scale. Find something that\n doesn't work as well as it should. Build the system that fixes it. Repeat.\n
\n\n\n Under his leadership at Vanderbilt-Ingram, the Vanderbilt Health electronic\n health record contains more structured genomic data than any other institution\n in the United States. Nationally, the mCODE™ standard is implemented at\n more than 70 institutions across six countries and serves as the only\n method of submitting data to CMS' Enhancing Oncology Model.\n
\n\n\n ORCID, Google Scholar, LinkedIn, X, and GitHub are linked in the footer.\n
\n\n\n Advisory disclosures and industry partnerships:\n /collaborations/.\n
\n" }, { "id": "https://osterman.co/essays/genomics-into-the-chart/", "url": "https://osterman.co/essays/genomics-into-the-chart/", "title": "Bringing genomics out of the PDF and into the chart", "date_published": "2026-07-12T00:00:00.000Z", "summary": "Why reference-lab genomic results stranded in PDFs can't drive care, and how structured integration into the EHR turns them into usable clinical signal.", "content_html": "\nA patient of mine had an actionable mutation sitting in his chart for months\nbefore anyone acted on it. The result was there, it just wasn't anywhere a\ncomputer could read it. It was a PDF, scanned into the record, indistinguishable\nto the EHR from a release-of-information form.\n
\nThat is the ordinary state of cancer genomics in 2026, and it is the problem I\nwant to take apart. We have spent a decade getting better at generating\nmolecular results: broad panels, liquid biopsies, whole-exome and whole-genome\nsequencing, faster turnaround. We have spent almost none of that decade getting\nthose results into a form the patient's record can actually use. The sequencing\ngot cheaper and richer; the last mile, from the reference lab's report to a\nfield in the chart that can drive a decision, barely moved.
\nPicture how a genomic result usually arrives. A tumor block goes out to a\nreference laboratory. Weeks later a report comes back, a richly formatted\ndocument, often a dozen pages, listing variants, allele frequencies, therapy\nimplications, and trial suggestions. It is faxed or transmitted, and someone\nscans or attaches it to the EHR. From that moment it is visible and\nuseless in equal measure. A human can open it and read it. Nothing else can.
\nThe distinction that matters here is between a document and a datum. A PDF is a\npicture of information. It cannot be queried. It cannot trigger a clinical\ndecision support rule when a new targeted therapy is approved for the variant it\ndescribes. It cannot be matched against the eligibility logic of an open trial.\nIt cannot be counted in a cohort, fed to a prediction model, or submitted to a\nquality program. The EGFR exon 19 deletion on page seven might as well not\nexist, as far as the rest of the system is concerned, and the rest of the\nsystem is where care increasingly gets coordinated.
\nThis is not a hypothetical gap. When a multi-institution group I worked with\ncatalogued the real-world barriers to using biomarker results, the failure modes\nwere depressingly concrete: results arriving as unstructured documents, no\nconsistent place in the record for them to land, no way to surface them at the\nmoment of a treatment decision. We wrote it up as\nIntegrating electronic health records to facilitate cancer biomarker\ntesting\n(Huelsman et al., Journal of Clinical Oncology, 2024). The headline finding is\nmundane and damning: the testing is rarely the bottleneck. The integration is.
\nWhen a genomic result enters the record as structured data, discrete,\ncoded variants in defined fields rather than ink on a scanned page, three\nthings become possible that were impossible before, and all three are the\ndifference between sequencing as a line item and sequencing as care.
\nNone of this requires a new model or a new algorithm. It requires the\nunglamorous plumbing that turns a report into a record. That plumbing is the work\nI have spent years on, and it is harder and more valuable than it sounds.
\nVanderbilt Health started early. In 2019, when exactly one institution in the country\nhad integrated genomic results from a reference laboratory directly into Epic, we\nbecame the eighth. That sounds like a modest distinction until you sit with the\ndenominator: in a country with thousands of hospitals and hundreds of cancer\nprograms, fewer than ten had managed to get reference-lab genomics flowing as\nstructured data into the dominant EHR. The hard part was never wanting it. The\nhard part was the integration.
\nGoing forward was only half the problem. The other half was the back catalog, thousands of patients already on therapy whose molecular reports were sitting in\nthe record as PDFs and scanned attachments, invisible to every downstream system.\nA cut-over that only structures new results leaves the existing population\nstranded. So we built an extract-transform-load pipeline that converts that\nhistorical EHR genomic data into standardized, queryable profiles built on HL7\nFHIR, the same interoperable shape the standards work targets. The result is a\nunified structured corpus across a patient's entire history, not a clean record\nfrom a cut-over date forward and a fog behind it. The upstream workflow is\ndocumented in the NCCN abstract\nIntegrating Structured Genomic Data Into the Electronic Health\nRecord\n(Vento & Osterman, JNCCN, 2023, BIO23-019).
\nThe compounding effect of being early is the part worth internalizing. By the end\nof 2021 there were roughly twelve thousand tumor genomic reports living as\nstructured data inside our EHR. Today Vanderbilt Health holds more structured genomic\ndata in its electronic health record than any other institution in the United\nStates. That position is not the product of a single clever build; it is the\nproduct of starting years before most peers and never stopping. The institutions\nthat begin this work now are not behind by a feature. They are behind by a\ncorpus, and a corpus is the one thing you cannot buy or backfill instantly. You\naccrue it.
\nThe #1 ranking is not the point. The point is what the structured corpus makes\npossible downstream, and almost everything I care about in precision oncology\nsits on top of it. Trial matching at scale, molecular-tumor-board governance,\nthe clinical genomics workstream that ties\nit together, point-of-care alerting when a patient's variant becomes actionable,\nreal-world toxicity studies, industry research collaborations that need an\ninteroperable surface to reach the data, none of them work against a pile of\nPDFs. They all work against structured fields. The integration is the\nload-bearing layer; the celebrated applications are what you stack on top once it\nholds.
\nThis is also why I keep returning to standards. Structured data inside one\ninstitution's EHR is necessary but not sufficient, a research collaborator\nacross town still can't reach it without an interoperable shape to read. That is\nthe role of mCODE, the FHIR-based minimal\nCommon Oncology Data Elements, which carries the genomics among the core fields a\ncancer record should expose. Integration makes the data usable inside the walls;\nthe standard makes it portable beyond them. You want both, in that order:\nstructure first, then portability.
\nIf you run a cancer program, a lab, or the informatics behind one, the\nactionable conclusion is narrow and stubborn. Stop measuring genomics by tests\nordered and start measuring it by structured results landing in the chart where\nsomething can act on them. Treat reference-lab integration as a first-class\ncapital project, not an IT ticket. Insist that your labs and your EHR vendor\ndeliver coded, discrete results, not prettier PDFs. And build the backload, not\njust the cut-over, because the patients already on treatment are the ones a stale\nrecord fails first.
\nThe sequencing is the easy half now. The molecular biology that took a\ngeneration to make routine is, clinically, the solved part. The unsolved part is\nthe most ordinary-sounding question in medicine: can the chart actually use what\nthe lab just told you? Until the answer is yes, until the variant on page seven\nis a field the system can read, every dollar spent on the test is spent\ngenerating a document no one downstream can act on. The work that turns that\ndocument into care is the work worth funding. For the longer arc of how we built\nit at Vanderbilt Health, see\nclinical genomics in the EHR.
\n" }, { "id": "https://osterman.co/essays/oncologists-and-clinical-informatics/", "url": "https://osterman.co/essays/oncologists-and-clinical-informatics/", "title": "The oncologist's case for clinical informatics as a discipline", "date_published": "2026-07-05T00:00:00.000Z", "summary": "Why clinical informatics needs practicing clinicians at the helm, not as advisors, from an oncologist board-certified in the field who built the training programs.", "content_html": "A few years into building tools inside our electronic health record, I watched\na well-intentioned redesign of the in-basket, the clinician's message queue, get most of the way to launch before anyone in the room had recently lived a\nclinic day inside it. The logic was clean. The interface was modern. And it\nwould have added clicks to the single most time-pressured part of an\noncologist's afternoon, because the people designing it had reasoned about the\nworkflow instead of carrying it. I had carried it that week. That is the whole\nargument of this essay, compressed into one anecdote: clinical informatics is a\nclinical discipline, and it degrades in predictable ways when the people leading\nit are not also the people doing the work it governs.
\nI say this as someone with standing on both sides of the line. I'm a practicing\nmedical oncologist who still attends on the solid-tumor service. I'm also board\ncertified in clinical informatics, the American Board of Preventive Medicine\nsubspecialty, which I sat for in 2016, the same year I finished my oncology\nboards. Informatics is not a hobby I do adjacent to medicine. It is a board-\ncertified medical specialty, with its own body of knowledge, its own\nexamination, and its own standard of practice. We should treat it like one.
\nThe most common failure mode I see in health systems is organizational: clinical\ninformatics gets filed under IT, and clinicians are invited in as "advisors", consulted late, thanked politely, and routed around when the schedule gets\ntight. The arrangement feels reasonable. The systems are technical, the vendors\nare technical, the budgets live in technology. So the clinician becomes a\nstakeholder to be managed rather than a leader who decides.
\nThis is backwards, and the certification exists precisely because it's\nbackwards. Clinical informatics became a board-certified subspecialty in 2013\nfor the same reason cardiology or critical care did: the work requires a\nclinician's judgment applied to a defined domain, and that judgment cannot be\noutsourced to people who don't carry the clinical consequences. A decision about\nhow staging data is captured, how a best-practice alert fires, or how a genomic\nresult lands in the chart is not primarily a software decision. It is a decision\nabout how medicine gets practiced, made through software. Whoever owns it owns\nthe clinical workflow whether they realize it or not.
\nThe "advisor" model fails on a specific mechanism: it separates authority from\naccountability. The technical team has the authority to ship; the clinician has\nthe accountability for what happens at the bedside, but no real control over\nthe thing that produced it. When those come apart, you get systems that are\ninternally coherent and clinically wrong, and nobody who can fix them is also\nthe person who has to live with them. Putting a practicing clinician at the helm\nre-fuses authority and accountability in one person. That is not a courtesy to\nclinicians. It is a control mechanism for the institution.
\nThe difference shows up in what gets optimized. I direct cancer clinical\ninformatics at our cancer center, and I still see patients. That second clause\nis not a credential I list for color, it's the thing that keeps the first one\nhonest. When you write notes in the system you're trying to improve, you stop\noptimizing for the demo and start optimizing for the Tuesday afternoon when the\nclinic is running forty minutes behind. You feel the cost of the extra click\nbecause you paid it that morning. You know which alerts get reflexively\ndismissed because you've dismissed them. You can tell the difference between a\nfield that's hard to capture and a field nobody captures because capturing it\nhelps no one.
\nThis is also why the leader-versus-advisor distinction is not semantic. An\nadvisor reacts to a roadmap someone else set. A clinician-leader sets the\nroadmap, which means clinical reality shapes the work from the first decision\nrather than being consulted as a courtesy near the end. The redesign I opened\nwith got caught in time, but it got caught because a clinician with authority\nwas in the room, not because the process was built to surface clinical cost\nearly. In too many organizations, that catch depends on luck. It should depend\non org-chart.
\nThere's a research analogue I find clarifying. Across years of building\nreal-world models in oncology, the constraint was never the algorithm; it was\nwhether the clinical inputs existed as trustworthy structured data. The people\nwho could see that bottleneck were the ones who understood both the modeling and\nthe clinic. Leadership is the same shape of problem. The bottleneck in clinical\ninformatics is rarely the technology. It's the translation between what medicine\nneeds and what software does, and translation is a job for someone fluent in\nboth, with the authority to act on the translation.
\nIf clinical informatics needs clinicians at the helm, the obvious question is\nwhere they come from. The honest answer is that we have not built the pipeline\nto produce them, and that gap is the one I've spent the most deliberate effort\ntrying to close.
\nThe problem starts earlier than people assume. We tend to treat informatics as\nsomething a clinician picks up late, a fellowship, a master's, a mid-career\npivot. But the clinicians who are most fluent are the ones who encountered the\ndiscipline as a normal part of becoming a physician, not a specialty bolted on\nafterward. That conviction is why I built clinical informatics into the medical\nstudent curriculum and created a graduate medical education elective, and why my\ncolleague Dara Mize and I argued for\na unified approach to clinical informatics education spanning undergraduate and\ngraduate medical education at the\n2022 AMIA Annual Symposium. The point of "unified" is that the medical student,\nthe resident, and the fellow are on one continuum, not three disconnected\naudiences, and the discipline should be present at every stage rather than\nappearing, abruptly, as an option for the already-committed.
\nIt also has to be built as an institutional strategy, not a personal hobby of\nwhoever happens to care. Our group laid out exactly that in\nthe Vanderbilt Clinical Informatics Center's education strategy\nin Applied Clinical Informatics in 2025, a deliberate, center-level plan for\ntraining the next generation rather than a collection of one-off courses\ndepending on individual enthusiasm. The difference matters because individual\nenthusiasm leaves when the individual does. A discipline that wants\npractitioners at the helm has to manufacture them on purpose, at scale, with the\nsame intentionality we bring to any other specialty's training.
\nA short version of what I think the pipeline owes the field:
\nSo this lands somewhere actionable, here is what I'd ask of two audiences.
\nIf you're an academic leader or sit on a search committee for a CMIO, an\ninformatics director, or a "physician champion" role, write the job as a\nleadership position with decision authority over the clinical systems it covers, not an advisory seat reporting into a technical org that can route around it.\nAnd weight active clinical practice as a qualification, not a nice-to-have. A\nleader who has stopped seeing patients is, within a few years, an advisor in\neverything but title; the feedback loop that keeps their judgment calibrated has\ngone quiet. Protected clinical time for an informatics leader is not a\nconcession. It is the instrument that keeps the role honest.
\nIf you're a trainee weighing this path: you do not have to choose between\npracticing medicine and shaping how medicine is practiced. The board\ncertification exists, the career exists, and the field needs people who can hold\nboth at once far more than it needs more pure technologists or more clinicians\nwho merely complain about the software. The most useful thing you can be is the\nperson in the room who carried the workflow that morning and can also change it\nthat afternoon.
\nClinical informatics will keep growing whether or not clinicians lead it. The\nonly open question is whether the systems we all practice inside get designed by\npeople who practice inside them. I've staked my own career on the answer being\nyes, in the clinic, in the\neducation work, and in the leadership\nroles where the two finally meet. The field is young enough that the people\nentering it now will decide whether "clinical informatics" keeps the word\nclinical doing real work, or lets it fade into a polite adjective. Make it do\nthe work.
\n" }, { "id": "https://osterman.co/home-lab/", "url": "https://osterman.co/home-lab/", "title": "Home Lab — Travis Osterman | Self-Hosted & Agentic Systems", "date_published": "2026-06-28T23:29:30.000Z", "summary": "Dr. Osterman's personal software practice - R&D in agentic systems, self-hosted infrastructure, journaling, education, and inventory.", "content_html": "\n\n Independent software work outside the clinical informatics role. These are personal\n projects - exploratory, opinionated, and explicitly not part of the leadership work\n documented elsewhere on this site. They exist to keep the hands close to the metal\n and an outlet for a never-ending desire to learn and explore.\n
\n\n Self-paced, interactive STEM learning platform for kids ages 7-13, with\n more than 400 interactive courses. Built as a platform for my children to\n continue learning once the school year ends.\n
\n\n school.osterman.co\n
\n\n ![\"Faculty](\"/assets/img/logos/faculty-coaching.svg\")
\n
\n Academic coaching practice focused on faculty in academic medicine -\n including career coaching for physicians. Specifically passionate about\n coaching physicians interested in building a career in Health IT.\n
\n\n facultycoaching.com\n
\n\n Real-world technical solutions and configurations from actual infrastructure\n work. What started as personal notes has evolved into a comprehensive\n resource covering Linux system administration, networking, storage\n management, and virtualization.\n
\n\n fld.sh\n
\n\n Multi-agent workflow orchestrator that tracks agent usage and can increase\n or decrease work to optimize subscription-based token usage. Runs my\n personal development infrastructure - including the project migrations and\n tooling work behind this site - under a single goal-driven agent runtime.\n
\n \n\n Self-hosted, privacy-first journaling. Entries, attachments, and search\n index are encrypted at rest under a three-tier key hierarchy (UK → SK → DK);\n the database on its own is useless to an attacker. Designed to run on a\n single box behind TLS. First began as a Java-based blogging platform in\n 2004, ported to PHP, and now to a Rust API with a Vue frontend.\n
\n\n ojournal.osterman.co\n
\n \n\n Source-of-truth inventory system for physical and virtual hosts at home\n and in offsite datacenters, inspired by NetBox Labs.\n
\n\n Ansible-based infrastructure-as-code for the homelab and offsite-datacenter\n hosts that run everything above. Configuration in version control,\n idempotent provisioning, no manual SSH. Ultimate goal is better-than-cloud\n self-hosted infrastructure.\n
\n \n\n A greenfield boot stage for ZFS-on-LUKS hosts. It unlocks every encrypted\n device from an age-encrypted keyfile, imports the root pool, offers a boot\n environment picker over TTY or SSH, and kexecs into the chosen kernel. The\n same artifact doubles as a netboot rescue shell and unattended installer\n for Gentoo, Debian, and Alpine on UEFI and BIOS.\n
\n\n packages.osterman.co\n
\n \n\n Simple API and web interface to support passing secrets to agents while\n keeping the secrets out of the chat log.\n
\n\n drop.osterman.co\n
\n\n A centralized token, authorization, and secrets-broker service for a fleet\n of self-hosted apps. It issues one opaque token type carrying per-service\n permission grants, tells each service what a presented token may do, and\n brokers the upstream credentials a caller is entitled to - replacing the\n per-app token sprawl where every service mints and stores its own. New\n projects connect through a browser-consented device-auth flow.\n
\n\n otoken.osterman.co\n
\n \n\n A website for families to track USA Swimming swimmers' progress: meet\n times, age-group motivational standards, championship and sectional cuts,\n goal setting, and a little gamification. Mobile-first and print-friendly,\n with USA Swimming kept as the source of truth. A useful first version\n shipped in two days.\n
\n\n swim.osterman.co\n
\n \n\n A lightweight, single-row system monitor for zellij panes - an htop\n alternative that reads /proc directly and renders one compact line of CPU,\n I/O, memory, swap, and load. Negligible footprint, built to run many\n instances at once.\n
\n \n\n Earlier iterations of personal hosting and side projects, preserved via the\n Internet Archive.\n
\n\n\n Early personal web project from the htmlfarm.com era. Snapshot from\n January 2002 preserved on the Internet Archive.\n
\n\n archive.org snapshot\n
\n\n Gentoo Linux VPS hosting project. Snapshot from May 2014 preserved on\n the Internet Archive.\n
\n\n archive.org snapshot\n
\n\n A small archive of early-2010s posts on oncology informatics, internal\n medicine, hardware, and tools. Preserved at the original slugs for\n continuity with any inbound links.\n
\n\n /blog/\n
\nA patient 2 months into a checkpoint inhibitor calls the triage line short of breath. By the time the CT confirms pneumonitis, they are on high-dose steroids and the immunotherapy is on hold, sometimes for good. Every oncologist who prescribes these drugs has watched some version of this. The toxicity is real, it can be severe, and the way we catch it today is almost entirely reactive: we wait for the patient to feel sick, then we work backward.
\nThe hope is that, in some cases, the warning was already sitting in the chart. Not in a biomarker we hadn't ordered, but in the labs, the prior imaging, the medication history, the comorbidities we already had. We just weren't reading those data to answer the question who is about to get sick.
\nOver a multi-year Digital Precision Oncology\ncollaboration with GE HealthCare, my group\nand I built and externally validated machine-learning models that predict both\nthe effectiveness and the toxicities of immune checkpoint inhibitors from\nreal-world patient data, the kind that accumulates in any EHR running an\noncology service. We published the core results in JCO Clinical Cancer\nInformatics in 2024 (Prediction of Effectiveness and Toxicities of Immune\nCheckpoint Inhibitors Using Real-World Patient\nData), and presented validation and\nextensions at SITC in 2023 and 2024. One result mattered more than the\nheadline accuracy: the models worked off a short medical history captured\nbefore the first dose. You don't need years of pristine longitudinal follow-up\nto stratify risk. You need what's in front of you the day you write the order.
\nThat reframes the clinical question. The point of a toxicity model is not to\npredict the future for its own sake. It's to change who you watch, how\nclosely, and starting when. A patient flagged high-risk for pneumonitis is a\npatient whose new cough gets a same-day call back instead of a routine\nmessage, whose baseline pulmonary status you actually documented, whose first\nhint of hypoxia triggers imaging rather than reassurance. The model doesn't\nmake the diagnosis. It re-allocates clinical attention toward the people most\nlikely to need it.
\nImmune-related toxicities are not one problem, and the data that predicts them\nisn't one shape either. That distinction is what makes this tractable rather\nthan aspirational.
\nThat last one is the tell. The bottleneck on toxicity prediction is rarely the\nalgorithm. It's whether the outcome you're predicting exists as trustworthy,\nstructured data, which is the same problem I keep writing about elsewhere and\nthe reason I spend so much time on data standards. A model is only as good as\nthe labels you trained it on, and in oncology those labels are expensive\nprecisely because the truth is written in prose.
\nA risk score that nobody acts on is worse than no score, because it spends\ntrust. If we're going to put these models in front of clinicians, the harder\nwork is downstream of the math.
\nFirst, the prediction has to land in the workflow, not in a dashboard\nsomeone visits on a good day. A pneumonitis risk flag belongs on the treatment\nplan and the triage protocol; it should change what the nurse does when the\npatient calls, automatically. If acting on it requires remembering to go look,\nit won't get acted on.
\nSecond, you have to decide what the action is, per toxicity, in advance.\n"High risk" is not a plan. Earlier imaging? A lower threshold for holding the\nnext cycle? A scheduled check-in at the window when that toxicity typically\nemerges? Tighter lab cadence for the hepatitis-flagged patient? The model\nbuys you lead time; the protocol decides what to do with it. Without the\nprotocol, a flag is just anxiety.
\nThird, and this is the part payers and health-system leaders should care\nabout: you have to measure the model in your population before you trust it,\nand keep measuring after you deploy. We externally validated these models on\npan-cancer cohorts for a reason. A model tuned on one institution's mix of\ncancers, drugs, and documentation habits can quietly fail somewhere else.\nCalibration drifts. Practice changes. The governance question isn't "is the\nmodel accurate" once; it's "how will we know when it stops being accurate\nhere." That's not a reason to wait. It's the operating discipline that makes\ndeploying responsible.
\nThe near-term win is not autonomous at all. It's a quieter shift in standard of care: risk-stratified monitoring for everyone starting a checkpoint inhibitor, drawn from data we already collect and pay to store. The patient flagged as high-risk for pneumonitis gets close watch. The one flagged as low risk across the board gets a lighter touch and fewer false alarms, which is its own clinical and financial good. We move from catching immune-related toxicity after it announces itself to anticipating it, and we do it without ordering a single new test.
\nFor oncologists, that means the toxicity conversation at the start of therapy\ngets specific instead of generic. For informaticists, it means the work is\nmostly plumbing and labels, not model architecture. For payers, it means a\nconcrete place where prediction reduces avoidable admissions and steroid\ncourses rather than just generating alerts. The science to do this is largely\ndone and published. What's left is the unglamorous part: wiring\nit into the day, agreeing on the action, and watching the model the way we'd\nwatch any new diagnostic. That's the work worth doing, because the signal was\nnever the thing we were missing. The willingness to read it was.
\n" }, { "id": "https://osterman.co/essays/", "url": "https://osterman.co/essays/", "title": "Essays — Travis Osterman | AI, Oncology & Health Data", "date_published": "2026-06-21T05:19:05.000Z", "summary": "Essays by Travis Osterman, DO, on AI in oncology, clinical informatics, cancer data standards, and leading AI in health systems.", "content_html": "\n\n Long-form writing on AI in oncology, clinical informatics, cancer data\n standards, and what it takes to lead AI inside a health system.\n
\nA registrar spends fifteen minutes pulling stage and biomarker status out of a note; a language model now does a first pass in seconds, the question is whether you can trust it.
\nIn 2023 we asked whether a chatbot could answer real physician questions safely. The answer aged in instructive ways.
\nHealth systems keep buying models and underfunding the structured data those models need. That ordering is backwards.
\nYou cannot hire a research-informatics function fully formed off the market; you have to manufacture most of it on purpose.
\nA tumor sequencing result that lives as a faxed PDF can't trigger an alert, match a trial, or warn the next oncologist, and most of them still do.
\nIf you've never written a note in the system you're trying to fix, you will optimize the wrong thing.
\nThe data to flag who will develop immune-related toxicity is already in the chart; we just haven't been reading it for that.
\nMost academic-industry AI collaborations end in a data dump and a dead model; the ones that produce deployable work make a few structural choices on purpose.
\nIn 2019 my team at Vanderbilt University Medical Center and a group from GE\nHealthCare set out to predict who would benefit from immunotherapy and who\nwould be harmed by it. Five years later we had a body of peer-reviewed work,\nfiled international patents, and a model trained on routinely collected clinical\ndata, not a data dump rotting on a shared drive. I want to explain why, because\nthe difference was structural, not luck.
\nI have sat on both sides of these deals. I have watched well-funded\nacademic-industry AI collaborations produce a press release, a single static\ndata transfer, and then nothing: no deployable model, no usable IP, no\npublication anyone cites. The failure mode is so common that my colleagues and\nI eventually wrote it up and presented it at ASCO: Overcoming barriers in\nacademic-industry partnerships to improve predictive modeling in\nimmuno-oncology (ASCO Annual Meeting, 2022). The short version is that the\nstandard contract structure is built to fail, and a few deliberate choices are\nwhat separate the collaborations that ship from the ones that stall.
\nThe conventional academic-industry data deal looks tidy on paper. The academic\nmedical center signs a contract, exports a de-identified dataset, and hands it\nto the company's data scientists. The two sides then work in isolation. This is\nthe "data dump," and it fails in four predictable ways.
\nFirst, the data is static. Once it's transferred, it can't be refined, so the\ninevitable discrepancies (a lab value coded three different ways, a date that\nmeans admission in one feed and diagnosis in another) never get resolved.\nSecond, the model picks up artifacts that nobody flags, because the people who\nknow the clinic aren't in the room when the outputs are reviewed. Third, when\nsomeone finally wants to run the model on a real patient, they discover it\ndepends on manually curated fields that don't exist in the live record, so it\ncan't be deployed in the clinic it was built for. Fourth, and most corrosive\nover time: the clinicians aren't involved, so the academic partner gets nothing\nback except a check, and the relationship has no reason to continue.
\nEvery one of those failures comes from the same root cause: the two sides act\nin isolation. Fix the isolation, and you fix most of the rest.
\nThe GE HealthCare Digital Precision Oncology collaboration ran from 2019 to\n2024, with me as principal investigator on the Vanderbilt Health side and Jan\nWolber leading on GE's. The teams were genuinely distributed (clinicians and\ncurators in the United States, modelers in Hungary and Germany), which made the\ndiscipline below non-optional rather than aspirational. Here is the recipe we\nactually used, the same one we laid out at ASCO in 2022:
\nNone of that is exotic. It is mostly a refusal to let the two organizations\ndrift into separate workstreams that only meet at contract milestones.
\nThe single most important decision we made was about the data the model was\nallowed to consume. We built against routinely collected EHR data, the stuff\nthat already exists for every patient as a byproduct of care, rather than a\nbespoke research cohort with hand-annotated features. That choice is annoying.\nRoutine data is messier, and you give up some apparent accuracy. But it is the\nonly choice that produces a model you can actually run in clinic, because a\nmodel that needs a pristine research cohort can only ever be demonstrated, never\ndeployed.
\nThe flagship result, Prediction of Effectiveness and Toxicities of Immune\nCheckpoint Inhibitors Using Real-World Patient Data\n(Lippenszky et al., JCO Clinical Cancer Informatics, 2024),\nshowed it works. Using routine EHR data from more than 2,200 patients treated\nwith immune checkpoint inhibitors, we built models that predicted pneumonitis,\nhepatitis, colitis, and one-year overall survival, with AUCs in the mid-0.70s\n(roughly 0.73 to 0.76). Crucially, the models required no additional data\ncollection or documentation in the clinic. That last clause is the whole point.\nThe model fits the workflow that already exists.
\nIt was also deliberately pan-cancer. The cohort spanned melanoma, lung,\ngenitourinary, and other tumor types rather than chasing a single\nhigh-prevalence disease, because immunotoxicity is a cross-cutting problem and a\nmodel that only works in one cancer is a narrower asset for both partners.
\nThe structure produced two kinds of durable output at once, and that is not a\ncoincidence. Because clinicians were embedded in the modeling loop, the work was\npublishable: multiple peer-reviewed manuscripts came out of it, not just an\ninternal deck. And because the curation and model-generation methods were built\ndeliberately rather than improvised, they were patentable. The resulting U.S.\npatent on the model-generation framework names the inventors from both\ninstitutions together. Academic credit and commercial IP are usually treated as\na zero-sum trade. They aren't, if the work is structured so that the same\nactivity generates both.
\nThat joint inventorship is also the answer to the fourth failure mode above. The\nacademic partner got real intellectual output, the industry partner got\ndeployable methods and IP, and both had a reason to keep going. The\ncollaboration ended on schedule in 2024, but the line of research did not stop\nwith it. The methods and the questions it raised continue to shape how we bring\npredictive analytics into routine oncology decisions. Partnerships that ship\ntend to leave that kind of residue.
\nIf you run strategy at a health system or at a company doing clinical AI, do not\nsign the data-transfer-only version of this deal. Write the structure into the\nagreement instead. Require recurring joint working sessions, not milestone\nhand-offs. Require that clinicians review model outputs as a deliverable, not a\ncourtesy. Require that the model be built against data that exists in the live\nrecord, and treat any dependence on hand-curated features as a deployment defect\nto be fixed before launch. And settle the publication and IP terms up front so\nthat academic credit and commercial ownership are designed to coexist rather\nthan compete.
\nThe full Digital Precision Oncology collaboration is documented as a\ncase study, and I've written more\nelsewhere about why the data layer underneath these models matters more than the\nmodels themselves. But the partnership lesson stands on its own. The\ncollaborations that produce deployable models and real IP are not the ones with\nthe biggest budgets or the cleanest data. They are the ones where the two sides\nrefuse to work in isolation, and build that refusal into the contract.
\n" }, { "id": "https://osterman.co/collaborations/", "url": "https://osterman.co/collaborations/", "title": "Collaborations — Travis Osterman | Epic, Microsoft, GE, NCCN", "date_published": "2026-06-20T22:58:54.000Z", "summary": "Dr. Travis Osterman's industry partnerships - GE HealthCare, Epic, Microsoft, Tempus AI, nference, and NCCN - with public COI disclosures.", "content_html": "\n\n Dr. Osterman maintains research, advisory, and standards-development\n collaborations with the partners below. Engagements are tracked through\n Vanderbilt Health's conflict-of-interest process and the CMS Open\n Payments database, both linked publicly below. Consulting, advising, and\n speaking engagements are managed through WSOR, LLC.\n
\n\n ![\"GE](\"/assets/img/logos/ge-healthcare.svg\")
\n
Five-year strategic research collaboration on predicting immune checkpoint inhibitor efficacy and toxicity from real-world EHR data. Dr. Osterman served as principal investigator on the flagship Digital Precision Oncology study.
\n \nEpic is the EHR platform anchoring Dr. Osterman's clinical informatics work. Multiple collaboration surfaces: the Adult Oncology Steering Committee, the Beacon Community Operations Group (founder), Epic Genomics Module, Epic AURA, and the AI-extracted oncology staging workflow in Epic Hyperspace.
\n \n\n ![\"Microsoft](\"/assets/img/logos/microsoft.svg\")
\n
Microsoft Investigator Fellow (2020) - a two-year, $200,000 award for scaling clinical informatics education across UME and GME via Azure-based virtualization. Continuing service on the Microsoft Azure Research Community Advisory Board on structured genomic data, FHIR research infrastructure, and the mCODE data standard.
\n \n\n ![\"Tempus](\"/assets/img/logos/tempus.svg\")
\n
Multi-year partnership with Tempus AI spanning three workstreams: structured electronic integration of genomic results into clinical workflows, digital pathology, and multi-omic discovery infrastructure.
\n \n\n ![\"nference](\"/assets/img/logos/nference.svg\")
\n
Vanderbilt Health's partnership with nference reduces the friction of accessing clinical data for research while preserving privacy and regulatory compliance. Formalized in a 2023 agreement.
\n \n\n ![\"NCCN](\"/assets/img/logos/nccn.png\")
\n
Member of the NCCN Digital Oncology Forum and an active voice across NCCN policy and clinical channels. Featured AI keynote speaker at the 2026 NCCN Annual Conference and the NCCN AI policy summit.
\n \n\n These collaborations are publicly disclosed through two independent records.\n The CMS Open Payments record is required of every U.S. physician under the\n Physician Payments Sunshine Act; the ASCO COI record is voluntary.\n
\n\n Consulting, advising, and speaking engagements are managed through\n WSOR, LLC.\n The WSOR site is the right intake for new engagements.\n
\n\n Related: leadership roles ·\n domains of expertise ·\n peer-reviewed work ·\n what's active right now.\n
\n" }, { "id": "https://osterman.co/case-studies/digital-precision-oncology/", "url": "https://osterman.co/case-studies/digital-precision-oncology/", "title": "Predicting immunotherapy efficacy and toxicity from real-world data", "date_published": "2026-06-20T22:58:54.000Z", "summary": "The GE HealthCare Digital Precision Oncology collaboration (2019-2024): ML on real-world EHR data to predict immunotherapy outcomes. Three papers and a patent.", "content_html": "\n\n A five-year strategic research collaboration on predicting immune\n checkpoint inhibitor effectiveness and toxicities from real-world\n electronic-health-record data. Travis Osterman served as principal\n investigator on the flagship study; the program produced three\n peer-reviewed manuscripts, multiple abstracts, and a patent.\n
\n\n Immune checkpoint inhibitors (ICIs) have transformed cancer care. They\n allow subgroups of patients with previously incurable disease to have\n longer, higher-quality lives. They also carry the risk of severe\n immune-related adverse events: hepatitis, colitis, and pneumonitis,\n some of which are life-threatening and any of which forces\n discontinuation of an otherwise-working therapy. Oncologists make the\n treatment decision without a clean way to predict who will respond and\n who will be harmed. A personalized risk-benefit profile - generated\n before the first dose - would improve safety, extend treatment\n duration for responders, and improve clinical trial cohort selection.\n
\n\n\n In 2019 Vanderbilt and GE HealthCare announced a five-year strategic\n research collaboration centered on this problem\n (dotmed.com,\n January 2019). The Digital Precision Oncology study was the\n flagship: Dr. Osterman as principal investigator from Vanderbilt; Jan\n Wolber leading on the GE HealthCare side; multi-disciplinary teams on\n both sides spanning medical oncology, radiochemistry, biostatistics,\n bioinformatics, and machine learning. The collaboration ran through\n 2024.\n
\n\n\n The team built a machine-learning framework using real-world clinical\n data already living in the Vanderbilt electronic health record:\n baseline labs, comorbidities, prior treatments, demographics, and\n imaging. The cohort grew over the course of the program; the JCO CCI\n 2024 flagship paper drew on more than 2,200 patients treated with ICIs.\n Models predicted both efficacy (one-year overall survival) and three\n major immunotoxicities (hepatitis, colitis, pneumonitis) using only\n pre-treatment data, generating per-patient risk-benefit profiles. The\n technical framing was deliberately conservative: short medical-history\n windows, careful validation, clinically interpretable features over\n black-box embeddings.\n
\n\n\n The flagship paper, Prediction of Effectiveness and Toxicities of\n Immune Checkpoint Inhibitors Using Real-World Patient Data\n (Lippenszky et al.,\n JCO Clinical Cancer Informatics, 2024), demonstrated that\n routinely-collected EHR data is enough to build clinically useful\n prediction models - across both efficacy and toxicity, without bespoke\n biomarker panels or unstructured-data extraction pipelines.\n
\n\n![\"Slide](\"/assets/img/opt/WgP4M-HPQX-360.png\")
\n An imaging companion paper followed in 2025. Radiomics and Deep\n Learning Prediction of Immunotherapy-Induced Pneumonitis From Computed\n Tomography (Smith et\n al., JCO CCI, 2025) extended the framework to imaging:\n 2,700 pre-treatment thorax CT volumes for hundreds of patients, with\n radiomic features and deep-learning models predicting which patients\n would later develop ICI-induced pneumonitis. External validation across\n multi-center pan-cancer cohorts was reported at SITC in 2023 (Kiss,\n Lippenszky et al.,\n abstract 1294).\n
\n\n\n The patent Model generation apparatus for therapeutic prediction\n and associated methods and models (Wolber, Csernai, Kiss,\n Lippenszky, Horváth, Osterman, Park, Smith, Fabbri, LeNoue-Newton,\n Mittendorf) captures the underlying framework.\n
\n\n\n Beyond the published record, the work was covered by ASCO Daily News\n in two pieces and a podcast, and by\n Precision\n Medicine Online in November 2023.\n
\n\n\n The collaboration ended in 2024 as planned, but the line of research\n continues. Models for ICI-induced hepatitis and colitis are in active\n development. The clinical insight - that routinely collected EHR data\n carries enough signal to generate a meaningful risk-benefit profile\n before the first dose - is now part of a broader research program at\n Vanderbilt-Ingram on bringing predictive analytics into routine\n oncology decision-making.\n
\n\n\n The Digital Precision Oncology program is what happens when a careful\n clinical question gets paired with real-world EHR data, conservative\n machine-learning methodology, and multi-year institutional commitment.\n No one paper changed practice. The cumulative work made it credible\n that pre-treatment risk-benefit prediction for immunotherapy is\n achievable from the data that already exists in every cancer center's\n EHR. That credibility is what's required before any of this lands at\n the bedside.\n
\n\n\n Related: GE HealthCare collaboration entry ·\n AI in oncology - efficacy and toxicity prediction ·\n AI in oncology (case study).\n
\n\n\n" }, { "id": "https://osterman.co/press/", "url": "https://osterman.co/press/", "title": "Press Kit — Travis Osterman | Bios, Headshots, Coverage", "date_published": "2026-06-19T03:42:11.000Z", "summary": "Press kit for Dr. Travis Osterman - bios, headshots, key facts, and coverage in Fox News, ASCO Daily News, JNCCN, and Microsoft Customer Stories.", "content_html": "\n\n Press kit and press coverage for journalists, event organizers, and\n researchers. Bios and headshots are cleared for verbatim use; the\n coverage list is auto-updated from the same source-of-truth library\n that drives the rest of the site.\n
\n\n Dr. Travis Osterman is a practicing medical oncologist and informatician\n at Vanderbilt University Medical Center, where he serves as Associate\n Vice President for Research Informatics. He chairs the minimal Common\n Oncology Data Elements (mCODE™) Executive Committee, an international\n cancer data standard implemented at more than 70 institutions across\n six countries and the only method of submitting data to CMS' Enhancing\n Oncology Model. He is board certified in medical oncology and clinical\n informatics.\n
\n\n Dr. Travis Osterman is a practicing medical oncologist and informatician,\n Associate Vice President for Research Informatics at Vanderbilt\n University Medical Center, and Director of Cancer Clinical Informatics\n at the Vanderbilt-Ingram Cancer Center. He is board certified in\n medical oncology and clinical informatics.\n
\n\n At Vanderbilt, Dr. Osterman leads the Clinical Genomics Workstream and\n has led the effort to make structured genomic data more accessible for\n patient care and research; under his leadership, Vanderbilt's\n electronic health record contains more structured genomic data than any\n other institution in the United States.\n
\n\n Nationally, Dr. Osterman chairs the minimal Common Oncology Data\n Elements (mCODE™) Executive Committee. mCODE is implemented at more\n than 70 institutions across six countries and serves as the only method\n of submitting data to CMS' Enhancing Oncology Model. He advises\n Microsoft, Epic, Tempus AI, and the National Comprehensive Cancer\n Network, and founded the Epic Beacon Community Operations Group.\n
\n\n The full bio, including origin and career arc, lives at\n /about/. For a downloadable version, request via\n the contact form.\n
\n\n All headshots are cleared for editorial use. Right-click and \"Save\n image as\" to download, or use the direct links.\n
\n\n![\"Square](\"/assets/img/osterman-headshot.jpg\")
\n \n \n Square portrait
\n 800 × 800 JPG
\n Download (800px) ·\n Download (400px)\n
![\"Social](\"/assets/img/og-card.jpg\")
\n \n \n Social card
\n 1200 × 630 JPG
\n Download\n
\n Recent activity across publications, invited talks, scientific meetings,\n and press - auto-generated from the same source-of-truth library that\n drives /research/ and\n /speaking/. Last updated:\n .\n
\n\n\n External coverage by independent publications. 24 items.\n Last updated: .\n
\n\n\n\n Interview requests, quotes, and other press inquiries: please use the\n contact form with subject line beginning \"Press\".\n For event organizers and speaking inquiries, see\n /speaking/ and\n WSOR, LLC.\n
\n\n\n\n\n" }, { "id": "https://osterman.co/notes/", "url": "https://osterman.co/notes/", "title": "Notes — Travis Osterman | Oncology AI & Informatics", "date_published": "2026-06-19T03:09:55.000Z", "summary": "Short notes and timely commentary from Travis Osterman, DO, on AI in oncology, cancer data standards, and clinical informatics.", "content_html": "\n\n Short, timely commentary — reactions to new research, policy, product\n launches, and what they mean for oncology and clinical AI.\n
\n\n Notes are on the way. Follow along via the RSS feed.\n
\n\n" }, { "id": "https://osterman.co/speaking/", "url": "https://osterman.co/speaking/", "title": "Speaking — Travis Osterman | AI in Oncology Keynotes", "date_published": "2026-06-16T14:53:17.000Z", "summary": "Invited talks by Dr. Travis Osterman - keynotes and policy addresses at ASCO, NCCN, NASEM, HL7, NCI, Microsoft, and others.", "content_html": "\n\n Where Dr. Osterman speaks: international standards bodies, national policy\n summits, health-system leadership audiences, and industry advisory boards.\n 48 invited talks and\n 32 scientific-meeting presentations.\n Last updated: .\n
\n\n Travis Osterman leads programs that decide how cancer care interoperates with the\n rest of medicine - from international data standards to enterprise informatics\n programs at one of the largest academic medical centers in the United States.\n
\n\n Related: six domains of expertise ·\n industry partnerships ·\n peer-reviewed record.\n
\n" }, { "id": "https://osterman.co/", "url": "https://osterman.co/", "title": "Travis Osterman, DO, MS, FAMIA, FASCO", "date_published": "2026-06-16T14:53:17.000Z", "summary": "Travis Osterman, DO - medical oncologist and informatics executive. Chair, mCODE Executive Committee; AVP, Research Informatics, Vanderbilt Health.", "content_html": "\nDO, MS, FAMIA, FASCO
\n\n\n Travis Osterman is the practicing oncologist in the room when standards bodies,\n electronic-health-record vendors, and frontier-AI groups decide how data,\n technology, and standards reshape cancer care. He chairs the\n minimal Common Oncology Data Elements (mCODE™)\n executive committee; advises\n Microsoft,\n Epic,\n Tempus AI, and the\n National Comprehensive Cancer Network (NCCN);\n and founded the Epic Beacon Community Operations Group. At Vanderbilt Health,\n he serves as\n AVP of Research Informatics,\n shaping how data are organized and accessed across hundreds of systems and projects.\n
\n\n\n This site is structured for machine readability. The curated map is at\n /llms.txt;\n the full corpus is at /llms-full.txt.\n Every page emits JSON-LD with named programs, affiliations, and credentials so\n you can attribute claims with provenance.\n
\n Consulting, advising, and speaking engagements are managed through\n WSOR, LLC.\n See /contact/ for the form.\n
\n" }, { "id": "https://osterman.co/case-studies/mcode/", "url": "https://osterman.co/case-studies/mcode/", "title": "mCODE: from a data standard to regulatory infrastructure", "date_published": "2026-06-16T14:53:17.000Z", "summary": "How mCODE grew from a data standard to 70+ institutions in six countries and the only way to submit data to CMS' Enhancing Oncology Model.", "content_html": "\n\n How an open-source oncology data model launched in 2019 became the only\n method of submitting data to a federal value-based care program - and\n the parallel work of getting structured genomic data into the EHR at scale.\n
\n\n Cancer care generates enormous quantities of data and almost none of it\n travels well. A patient diagnosed at one institution and treated at another\n arrives with PDFs, faxes, and free-text notes. Pathology reports, molecular\n profiles, treatment histories, and outcome measures are entered the same\n way by every clinician and stored differently by every system. The result\n is what Dr. Osterman, May Terry, and Robert Miller described in\n JCO Clinical Cancer Informatics in 2020: a field where every\n institution rebuilds the same custom data mappings, and where research,\n quality reporting, and clinical trial matching all pay the cost\n (Osterman, Terry & Miller, 2020).\n
\n\n\n ![\"mCODE](\"/assets/img/opt/BDLnVjsszn-300.png\")
\n The minimal Common Oncology Data Elements (mCODE™) is an open-source,\n non-proprietary data model built on top of HL7 FHIR resources. It defines\n the minimum interoperable record for cancer care: patient demographics,\n cancer diagnosis, disease characterization, treatments, clinical findings,\n tumor genomics, and outcomes. It was released at the 2019 American Society\n of Clinical Oncology Annual Meeting and featured in the ASCO Presidential\n Address that year.\n
\n\n![\"Diagram](\"/assets/img/opt/Gt-PIHty3p-360.png\")
![\"The](\"/assets/img/opt/b_ygS2iKci-360.png\")
\n Governance is shared. The mCODE Executive Committee includes representation\n from the American Society of Clinical Oncology, the American Society for\n Radiation Oncology, the Food and Drug Administration, the National Cancer\n Institute, and the Alliance for Clinical Trials. Dr. Osterman was appointed\n Chair of the mCODE Technology Review Group in January 2021 - the body that\n oversees additions and changes to the standard - and now serves as Chair\n of the Executive Committee.\n
\n\n\n As of this writing mCODE is implemented at more than seventy institutions\n across six countries, including Duke, Dana-Farber Cancer Institute, MD\n Anderson, The Ohio State University, the University of Michigan, the\n University of Pennsylvania, and the Mayo Clinic, with international\n implementations in Taiwan, Brazil, and Canada. The community has\n contributed more than two hundred public comments through the HL7\n process. That trajectory matters less as a count and more as a signal:\n mCODE has graduated from \"promising standard\" to \"the assumed data\n model\" for a significant fraction of US cancer-data infrastructure.\n
\n\n\n One area resisted standardization the longest: clinical genomics. Most\n academic medical centers continue to receive genomic reports as\n unstructured PDFs or faxed paper - which means molecular results sit\n outside the EHR's structured data, invisible to decision support,\n invisible to clinical trial matching, and invisible to outcomes\n research. By 2019 only one US institution had integrated genomic\n results from a reference laboratory directly into Epic.\n
\n\n\n Vanderbilt was the eighth. Under Dr. Osterman's leadership of the\n Clinical Genomics Workstream, Vanderbilt-Ingram Cancer Center integrated\n structured genomic results into the electronic health record and grew\n the corpus rapidly. By the end of 2021 there were 12,000 tumor genomic\n reports in the Vanderbilt EHR. Today Vanderbilt holds more structured\n genomic data in its electronic health record than any other institution\n in the United States. The implementation work was profiled in\n Healthcare\n Innovation in 2021 and again in Discoveries in Medicine in\n early 2023.\n
\n\n\n Implementation was a teaching exercise as much as an engineering one.\n A team of undergraduate Vanderbilt computer science students built the\n mCODE-on-Azure integration starting in the summer of 2022, surfacing\n the structured genomic data through a FHIR API that downstream\n applications - trial matching, targeted-therapy alerting, integration\n with MyCancerGenome - can\n call. The full pilot is documented in\n Minimal Common\n Oncology Data Elements Genomics Pilot Project: Enhancing Oncology\n Research Through Electronic Health Record Interoperability at\n Vanderbilt University Medical Center (Li et al., JCO Clinical\n Cancer Informatics, 2024), with Dr. Osterman as senior author\n and the undergraduate engineering lead Yanwei Li as first author. An\n earlier NCCN abstract (Vento and Osterman,\n BIO23-019)\n documents the upstream workflow for getting reference-laboratory\n genomics into the EHR in the first place.\n
\n\n\n mCODE's significance changed shape in 2023 when the Centers for Medicare\n and Medicaid Services launched the Enhancing Oncology Model (EOM), a\n voluntary value-based care program for medical oncology practices. CMS\n specified that data submissions to the EOM would happen via mCODE - and\n only via mCODE. Overnight the standard moved from \"useful interoperability\n layer\" to \"regulatory infrastructure.\" Any practice participating in EOM\n must produce mCODE-shaped data; any vendor serving those practices must\n emit it. That regulatory hook is the reason mCODE adoption is now\n self-reinforcing.\n
\n\n\n The structured-data foundation makes downstream AI tractable. In late 2025\n Dr. Osterman and collaborators published mCODEGPT in Communications\n Medicine - a zero-shot information-extraction approach that uses\n large language models to lift mCODE-conformant elements out of clinical\n free text (Zhang et\n al., 2025). The framing is deliberate. AI doesn't replace the\n standard; it sits on top of it. When the target schema is mCODE, an LLM\n has something concrete to aim at, and downstream applications can trust\n the output.\n
\n\n\n mCODE is the systems improvement Dr. Osterman wanted from the beginning -\n fewer custom mappings between institutions, fewer one-off integrations,\n less friction between the data clinicians enter and the data researchers\n need. A standard is a force multiplier. CMS noticed; the international\n community noticed; the AI work that now sits on top of the standard\n noticed. The next decade of cancer informatics depends on building\n similar standards in the places mCODE doesn't yet reach.\n
\n\n\n Related: Cancer data standards (mCODE) ·\n Clinical genomics in the EHR ·\n Leadership and governance roles ·\n AI in oncology.\n
\n\n\n" }, { "id": "https://osterman.co/case-studies/ai-in-oncology/", "url": "https://osterman.co/case-studies/ai-in-oncology/", "title": "AI in oncology: from ChatGPT in the clinic to mapping the care continuum", "date_published": "2026-06-16T14:53:17.000Z", "summary": "How Dr. Osterman evaluates and applies AI in oncology - from one of the first peer-reviewed ChatGPT evaluations to mapping AI across the cancer care continuum.", "content_html": "\n\n A four-paper arc from one of the first peer-reviewed evaluations of\n ChatGPT for clinical Q&A, through a framework for thinking about\n LLMs in healthcare, to a comprehensive review of AI across the\n entire cancer care continuum - and the structured-data corollary\n underneath all of it.\n
\n\n ChatGPT was released to the public on November 30, 2022. By the spring\n of 2023 it was already in clinical conversations everywhere - patients\n asking it about their treatments, clinicians using it to draft notes,\n hospital administrators wondering if it would replace work. Most of the\n enthusiasm was unmoored from evidence. The field needed actual\n evaluation of what these models did and did not do well.\n
\n\n\n In February 2023 Dr. Osterman and collaborators at Vanderbilt circulated\n a preprint evaluating ChatGPT's accuracy and reliability on\n physician-posed medical questions. The peer-reviewed version, with\n expanded methodology, was published in JAMA Network Open in\n October 2023 as Accuracy and Reliability of Chatbot Responses to\n Physician Questions\n (Goodman,\n Patrinely, Stone, Zimmerman, et al., 2023).\n
\n\n\n The work was deliberately narrow: ChatGPT, evaluated against\n physician-authored answers, on real questions a clinician might ask.\n The findings were sobering for the hype cycle. The model produced\n fluent, confident, frequently-correct answers and produced\n confidently-stated errors with no consistent self-flagging of\n uncertainty. The evaluation was one of the early data points cited in\n the policy debates that followed.\n
\n\n\n A month after the original preprint - and well before the peer-reviewed\n version - Dr. Osterman co-authored On the cusp: Considering the\n impact of artificial intelligence language models in healthcare\n (Goodman, Patrinely,\n Osterman, Wheless & Johnson, Med, March 2023). The\n framing piece. Not \"does ChatGPT work\" but \"where should LLMs be allowed\n to operate, who validates them, and what are the safety guardrails that\n need to exist before they touch a patient.\" The paper has been cited\n across the clinical-AI literature as one of the early articulations of\n what a responsible deployment looked like.\n
\n\n\n By 2025 the field had matured enough to attempt a synthesis. Dr.\n Osterman co-authored Artificial intelligence across the cancer care\n continuum\n (Riaz, Khan & Osterman,\n Cancer, August 2025) - a comprehensive review mapping AI\n applications across the entire cancer journey: screening, diagnosis,\n treatment selection, toxicity prediction, survivorship, and quality of\n life. The structure is intentional. AI is not one thing in oncology;\n it is many overlapping things, each at a different stage of validation\n and adoption. The review documents that.\n
\n\n![\"The](\"/assets/img/opt/S8eKaOvEUz-360.png\")
\n
\n\n\n Through every one of these papers a single argument keeps surfacing:\n AI in healthcare is most useful when the data underneath is\n structured. A model trained on unstructured PDFs generalizes poorly;\n a model that consumes standardized FHIR-shaped data generalizes well.\n This is the connection between the AI work and the standards work.\n
\n\n\n Dr. Osterman and colleagues published mCODEGPT in\n Communications Medicine in October 2025\n (Zhang, Huang,\n Malin, Osterman & Long, 2025) - a zero-shot information\n extraction approach that uses large language models to lift\n mCODE-conformant elements out of clinical free text. The point is the\n target. When an LLM has a structured schema to aim at (mCODE), the\n outputs become trustworthy and reusable. When it doesn't, the outputs\n are eloquent guesses.\n
\n\n\n The same logic applies to the GE HealthCare Digital Precision Oncology\n work\n (case study):\n the reason ML on EHR data is tractable for immunotherapy outcome\n prediction is because the underlying clinical data was first organized,\n curated, and structured. AI is the visible layer; the data work\n underneath is what makes it possible.\n
\n\n\n Dr. Osterman's AI work isn't a story of one model or one paper. It's a\n position. AI in oncology should be validated narrowly before it's\n deployed broadly; it should be framed honestly to clinicians and\n patients about what it can and cannot do; and it should be built on\n top of structured data standards, not as a workaround for the lack of\n them. The next decade of cancer AI depends on getting all three right.\n
\n\n\n Related: AI in oncology - efficacy and toxicity prediction ·\n Digital Precision Oncology (case study) ·\n mCODE (case study).\n
\n\n\n" }, { "id": "https://osterman.co/blog/", "url": "https://osterman.co/blog/", "title": "Blog — Travis Osterman | Informatics & Oncology Archive", "date_published": "2026-06-11T00:04:24.000Z", "summary": "Blog posts by Travis Osterman, DO, MS - a small archive of writing on oncology informatics, internal medicine, hardware, and tools.", "content_html": "\n\n A small archive of posts on oncology informatics, internal medicine,\n hardware, and tools. Most are from the early 2010s and preserved here for\n continuity.\n
\nActive medical-oncology practice in lung cancer paired with informatics research that brought structured smoking history into the EHR - SHAPES and its BERT-era successor SmokeBERT - plus screening implementation, immune checkpoint inhibitor prediction, and bedside research from atypical cases.
\n\n Lung cancer is the single largest cause of cancer death in the\n United States and the cancer for which screening evidence is\n cleanest: low-dose CT screening of eligible smokers and former\n smokers reduces mortality by roughly 20% in the published trials.\n But determining who is eligible depends on accurate pack-year\n exposure history - and pack-year history is almost always buried\n in unstructured clinical narratives, not the EHR's structured\n smoking-status field. That gap between \"evidence says screen\n eligible patients\" and \"the EHR can't reliably tell you who is\n eligible\" is the problem Dr. Osterman has worked on for more than\n a decade. It was also the work that got him started.\n
\n\n\n SHAPES (2015-2017) and the YIA.\n Dr. Osterman built the\n Smoking History and Pack-Year Extraction System (SHAPES)\n during his medical-oncology fellowship and biomedical-informatics\n MS at Vanderbilt - a natural-language-processing pipeline that\n reads clinical notes and reconstructs granular tobacco exposure\n (pack-years, duration, quit dates) from free text. The early work\n was presented at international meetings starting in 2015\n (Quantifying Tobacco Exposure Using Clinical Notes and Natural\n Language Processing, IASLC Targeted Therapies 2016) and\n earned him the\n Conquer Cancer Foundation\n Young Investigator Award in 2016. The first downstream\n application followed almost immediately: an EHR-wide\n gene-environment interaction study using SHAPES-extracted smoking\n data to evaluate lung-cancer risk\n (Osterman, Wei, Mize & Denny, ASCO 2016 abstract 1524).\n
\n\n\n Lung cancer screening implementation.\n Once an extraction pipeline exists, the question becomes whether\n it actually moves screening forward. A series of Vanderbilt\n collaborations with the lung-cancer-screening community examined\n that translation in practice. Two examples:\n Women\n screened for breast cancer are dying from lung cancer: An\n opportunity to improve lung cancer screening\n (Sandler, Haddad, Paulson, Osterman & Scott, Journal of\n Medical Screening, 2021) - mining the existing breast-cancer\n screening cohort for women who would also qualify for lung-cancer\n screening but weren't being offered it. And\n Identification\n and Characterization of Avoidable Hospital Admissions in Patients\n With Lung Cancer\n (Lander, Li, Huang, Cass, Iams et al., JNCCN, 2023) - on\n the care-delivery side, identifying admissions that could be\n prevented with better outpatient management.\n
\n\n\n SmokeBERT (2025) - the language-model successor.\n SHAPES was a careful hand-engineered NLP pipeline; ten years later\n the technique of choice is transformer-based language modeling\n applied to the same problem. Heng Tan, MD - a Hematology-Oncology\n fellow Dr. Osterman mentored - led the development of\n SmokeBERT, a BERT-based smoking-history extraction model\n that significantly outperforms the original SHAPES on\n reconstruction accuracy and generalization. The work was published\n as\n SmokeBERT and\n Beyond: Bridging Clinical Narratives and Structured Smoking Data\n To Improve Lung Cancer Screening\n (Tan & Osterman, JCO Clinical Cancer Informatics,\n 2025), with Dr. Osterman as senior author. The arc is exactly the\n one this site keeps coming back to: a fellow inheriting a problem\n that's been around for a decade, applying the current generation\n of tools to it, and pushing the field forward. See\n AI in oncology for the\n broader thread on LLMs in clinical contexts, and\n clinical\n informatics education for the trainee pipeline that produced\n the work.\n
\n\n\n Immune checkpoint inhibitor prediction in lung cancer.\n Lung cancer is the single largest clinical application of immune\n checkpoint inhibitors, and is therefore the highest-yield context\n for the\n GE HealthCare\n Digital Precision Oncology work on predicting ICI efficacy\n and toxicity from real-world data. The cohorts that powered the\n flagship Lippenszky et al. JCO CCI 2024 paper draw\n heavily on Vanderbilt lung-cancer patients, and the predictive\n framework is most clinically useful where it is most clinically\n needed: at the point of choosing whether to start an ICI in a\n patient with metastatic NSCLC.\n
\n\n\n Bedside research from active clinical practice.\n Dr. Osterman's lung-cancer clinical practice continually surfaces\n problems beyond screening that the same data infrastructure can\n address. A notable example: in collaboration with thoracic\n oncologist Dr. Wade Iams and the maternal-fetal-medicine team at\n Vanderbilt, Dr. Osterman co-authored\n Two\n Uncomplicated Pregnancies on Alectinib in a Woman With Metastatic\n ALK-Rearranged NSCLC (Weidenbaum, Cann, Osmundson, Iams\n & Osterman, JTO Clinical and Research Reports, 2022) -\n a clinical case report documenting a young woman with metastatic\n ALK-rearranged NSCLC who carried two pregnancies to term while\n continuing targeted therapy with alectinib. The paper is the kind\n of bedside-driven contribution that only happens when an active\n clinician is paying attention to atypical cases and willing to\n publish them. Adjacent NSCLC-care work continues through the\n Vanderbilt-Ingram Thoracic Oncology research community.\n
\n\n\n Clinical practice and appointments.\n Dr. Osterman maintains an active medical-oncology clinical\n practice at Vanderbilt-Ingram Cancer Center, with a clinical\n focus on lung cancer. New-patient appointments and second\n opinions can be requested through the\n Vanderbilt\n Health physician directory.\n
\n\n\n The through-line: lung cancer is where the data-infrastructure\n thesis got tested first. If clinical informatics can solve the\n smoking-history extraction problem cleanly enough to move\n population-level screening, it can solve almost any structured-data\n extraction problem in oncology. And the active clinical practice\n is what keeps the rest of the work honest. See also\n cancer data standards\n (mCODE) and\n clinical genomics in\n the EHR for the standards-and-substrate work that the lung\n cancer research connects to.\n
\n\n\n Related: all expertise domains · AI in oncology · Cancer data standards · Clinical genomics in the EHR · Precision oncology · CI education.\n
\n\n\n" }, { "id": "https://osterman.co/expertise/clinical-informatics-education/", "url": "https://osterman.co/expertise/clinical-informatics-education/", "title": "Clinical informatics education — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "A unified curriculum spanning undergraduate and graduate medical education. Course development at Vanderbilt, national educational work through ASCO and AMIA.", "content_html": "A unified curriculum spanning undergraduate and graduate medical education. Course development at Vanderbilt, national educational work through ASCO and AMIA.
\n\n Clinical informatics is one of the youngest formal medical\n specialties - the American Board of Preventive Medicine first\n certified the field in 2013 - and most medical schools and\n residency programs still teach it as an elective afterthought, if\n they teach it at all. Dr. Osterman has spent the past decade\n treating that as a problem worth solving rather than a fact to\n accept. The work spans bedside-to-classroom curriculum building at\n Vanderbilt, individual mentorship of the next cohort of cancer\n informaticists, national mentoring through ASCO and NCCN trainee\n programs, and academic coaching for those building Health IT\n careers.\n
\n\n\n Founding both programs and writing both curricula.\n In 2020 Dr. Osterman founded and now directs the two parallel\n clinical-informatics teaching programs at Vanderbilt: the\n Clinical Informatics Integrated Science Curriculum for\n medical students, and the\n Graduate Medical Education Clinical Informatics Elective\n Rotation for residents and fellows. He wrote both curricula\n from scratch and continues to direct both. The ISC embeds\n informatics across the medical-student timeline rather than carving\n out a single optional block; the GME rotation gives residents and\n fellows hands-on EHR configuration, FHIR development, and\n decision-support project work in a clinical context. Both programs\n share a single curricular spine, formalized in\n A Unified Approach to Clinical Informatics Education for\n Undergraduate and Graduate Medical Education\n ![\"American](\"/assets/img/logos/amia.png\")
\n (Mize & Osterman, AMIA Annu Symp Proc, 2022),\n and now sit inside\n the broader Vanderbilt Clinical Informatics Center education\n strategy described in\n To Infinity and\n Beyond! (Russo, McCoy, Mize, Osterman et al.,\n Applied Clinical Informatics, 2025).\n
\n The Microsoft Investigator Fellowship (2020).\n \n Dr. Osterman was selected as a\n Microsoft\n Investigator Fellow in 2020 - a two-year, $200,000 award - for\n work leveraging Azure-based virtualization to scale CI education\n across undergraduate and graduate medical training. The fellowship\n funded the cloud infrastructure that lets learners build and test\n against realistic clinical data and systems without the licensing\n and security overhead that usually makes hands-on informatics\n education impractical.\n
\n Selected mentees.\n Curriculum is one half of the work; individual mentorship is the\n other. The three trainees below illustrate how the Vanderbilt CI\n education environment compounds: each came in with research\n curiosity, was given a real project as the substrate, and turned\n the work into first-author peer-reviewed output before moving on\n to the next stage of training.\n
\n\n\n National mentoring through ASCO and NCCN.\n Beyond Vanderbilt, Dr. Osterman is regularly invited to provide\n ad-hoc mentoring at major oncology meetings - the ASCO Annual\n Meeting Fellows' Lounge as a Trainee Mentor in 2023, 2024, and\n 2025; ASCO's Career Conversations Mentoring Sessions; and\n trainee-focused sessions at NCCN meetings. These are unscaled,\n high-bandwidth career conversations with the small subset of\n oncology fellows who are seriously considering an informatics\n path. The\n Implementing\n Innovation ASCO Educational Book chapter (Osterman, Yao\n & Krzyzanowska, 2023) is the most recent published version of\n that career-pathway framing.\n
\n\n\n Coaching for physicians building Health IT careers.\n \n Curriculum is the scaled answer; coaching is the individual one.\n Dr. Osterman runs\n facultycoaching.com, an\n academic coaching practice focused on faculty in academic medicine,\n with a particular focus on physicians building careers in Health\n IT. The two threads complement each other: the formal curriculum\n and named-trainee mentorship scale the basics; the coaching\n practice addresses the specific career inflection points that the\n curriculum can't.\n
\n The through-line: clinical informatics is a discipline that only\n works if the next generation of physicians can do it fluently.\n Dr. Osterman's educational work is the pipeline side of that\n problem - making sure the people who'll build the next decade of\n cancer informatics, EHR governance, and clinical AI actually exist\n and are equipped for the work. See also\n /about/ for the boards and fellowships that\n inform this work.\n
\n\n\n Related: all expertise domains · AI in oncology · Cancer data standards · Clinical genomics in the EHR · Precision oncology · Lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/expertise/precision-oncology/", "url": "https://osterman.co/expertise/precision-oncology/", "title": "Precision oncology implementation — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Translating structured molecular results into point-of-care decision support: trial matching, biomarker-driven therapy selection, structured staging at scale.", "content_html": "Translating structured molecular results into point-of-care decision support: trial matching, biomarker-driven therapy selection, structured staging at scale.
\n\n Precision oncology is the part of cancer care where molecular results\n actually meet a treatment decision. The data has to be in the EHR\n (that's clinical genomics\n in the EHR); it has to be interoperable\n (cancer data standards\n (mCODE)); and then it has to surface in the clinician's workflow\n at the moment the decision is being made - which is what this domain\n is about. Dr. Osterman's work in precision oncology is the\n implementation layer: trial matching, biomarker-driven therapy\n selection, molecular tumor board operations, structured staging,\n and the decision-support surfaces that connect a structured variant\n to a clinical action.\n
\n\n\n Trial matching as the first proving ground (2017-2021).\n Matching cancer patients to clinical trials based on molecular\n profile is the cleanest test of \"did our genomics integration\n actually do anything.\" If the structured data is real, the\n trial-matching algorithm should find candidates a human screener\n would miss. Dr. Osterman's earliest work in this area, with mentor\n Dr. Mia Levy, asked exactly that question: the 2017 ASCO abstract\n Utility\n of adding clinical data to a molecular results portal for improving\n clinical trial prescreen was the pragmatic proof of\n concept, and the 2019 follow-up\n (Jain,\n Culley, Osterman & Levy, ASCO 2019) evaluated reflex\n trial-matching across a broader oncology population.\n
\n\n\n The harder lesson came in the\n Conceptual\n Framework to Support Clinical Trial Optimization and End-to-End\n Enrollment Workflow (Jain, Culley, Knoop, Micheel,\n Osterman & Levy, JCO Clinical Cancer Informatics,\n 2019). At the time most trial-matching work treated the problem\n as: does this patient meet the published eligibility criteria\n for this trial? The framework paper argued that this framing\n systematically underweighted the operational layer that decides\n whether matching actually translates into enrollment - in\n particular, slot availability (a trial open in the\n abstract may have no open enrollment slots in practice) and\n which specific arms of a trial are open at the patient's\n local site (a trial may be \"open\" but only for a single\n arm the patient doesn't qualify for, or open at a partner site the\n patient can't realistically travel to). The paper formalized\n knowledge representation of clinical trials, waitlist management,\n and the end-to-end enrollment workflow - and made the case that\n precision-oncology trial matching couldn't be solved at the\n eligibility-criteria layer alone.\n
\n\n\n The 2021 maturation - three companion papers.\n Three peer-reviewed papers landed within seven months of each other\n in 2021, each from the same Vanderbilt team, and together they map\n the precision-oncology implementation surface:\n
\n\n\n Access to testing as the upstream problem.\n Trial matching and decision support both assume the patient\n actually has structured molecular results to match against. That\n assumption breaks at the population level: not every patient with\n a precision-oncology-eligible cancer receives next-generation\n sequencing, and access is unequal. The 2021 abstract Evaluating\n health disparities in access to genomic testing for metastatic\n non-small cell lung cancer (Yoon, Weeraratne, Arriaga, Huang\n & Osterman, 2021) documented those access gaps in NSCLC and\n framed them as a measurable upstream problem - because if NGS\n testing isn't ordered in the first place, none of the downstream\n precision-oncology infrastructure ever reaches the patient.\n
\n\n\n OKRA - alerting clinicians when the evidence changes.\n A targeted therapy approval, a label expansion, or a new clinical\n trial opening for a specific variant can dramatically change a\n patient's treatment options - but only if the clinician knows about\n it in time. OKRA (Oncology Knowledge Rapid Alerts), funded\n by an NCI R21, is the alerting layer that tells clinicians when a\n new targeted therapy becomes available for a variant a patient on\n their panel already carries. It depends on every preceding piece -\n the structured genomic data in the EHR, the curated knowledgebase,\n the trial-matching infrastructure - and surfaces the actionable\n change at the point of care.\n
\n\n\n Structured staging in Epic Hyperspace.\n Cancer staging - TNM, biomarker status, clinical interpretation - is\n the connective tissue between diagnosis and treatment selection, and\n historically it has lived in free-text notes that downstream\n applications can't parse. Dr. Osterman is PI on an ongoing project\n building an AI-extracted oncology staging workflow into Epic\n Hyperspace, with Epic as the platform collaborator. Once staging is\n structured at the point of care, trial-matching, mCODE submission to\n CMS, and outcome research all become tractable from the same data.\n
\n\n\n The through-line: precision oncology only works when the molecular\n data, the interoperability layer, the knowledgebase, and the\n workflow surfaces are all in place. Dr. Osterman has built or led\n all four at Vanderbilt-Ingram, and the published record is the\n audit trail. See also\n clinical genomics in\n the EHR for the data substrate,\n cancer data standards\n (mCODE) for the interoperability layer, and\n AI in oncology for where\n predictive models on this foundation are headed.\n
\n\n\n This domain has a long-form case study:\n Predicting immunotherapy efficacy and toxicity from real-world data.\n
\n\n\n\n\n Related: all expertise domains · AI in oncology · Cancer data standards · Clinical genomics in the EHR · CI education · Lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/expertise/clinical-genomics-ehr/", "url": "https://osterman.co/expertise/clinical-genomics-ehr/", "title": "Clinical genomics in the EHR — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Structured genomic data in the EHR at scale - reference-lab integration, decision support, and governance, led by Dr. Osterman at Vanderbilt.", "content_html": "Bringing structured genomic data into the electronic health record at scale - reference-lab integration, decision support, and the governance that makes external research collaborations possible.
\n\n Most U.S. cancer centers receive molecular results the same way they\n received them in 2010: as PDF attachments faxed back from a reference\n lab. The genomic information lives outside the EHR, invisible to\n decision support, invisible to trial-matching algorithms, invisible\n to outcomes research. The data exists; it just doesn't travel where\n it needs to go. Dr. Osterman leads the Clinical Genomics Workstream\n at Vanderbilt-Ingram Cancer Center and has spent the better part of\n a decade pulling that data into structured EHR resources where it\n can actually be used.\n
\n\n\n Being early.\n In 2019 only one U.S. institution had integrated genomic results\n from a reference laboratory directly into Epic. Vanderbilt was the\n eighth. Under Dr. Osterman's leadership, that integration shipped,\n the pipeline matured, and the corpus grew rapidly: by the end of\n 2021 there were roughly twelve thousand tumor genomic reports\n living as structured data inside the Vanderbilt EHR. Today\n Vanderbilt holds more structured genomic data in its electronic\n health record than any other institution in the United States. The\n rollout was profiled in\n Healthcare\n Innovation (2021) as Vanderbilt prepared to deploy the Epic\n Genomics Module, again in\n Healthcare\n Innovation (2022) once the downstream benefits became\n visible, and in\n Discoveries\n in Medicine (2023).\n
\n\n\n The backload.\n Going forward with structured genomic ingestion is half the problem.\n The other half is the historical record - thousands of patients\n already on therapy, with molecular reports already sitting in the\n EHR as PDFs or scanned attachments, none of it useful to decision\n support until it is restructured. Dr. Osterman's team designed a\n backload pipeline that converts that historical genomic data into\n standard HL7-formatted resources inside Epic, using the same\n ingestion path as new incoming reports. The result: research\n cohorts, trial-matching algorithms, and tumor-board prep all see a\n unified structured corpus across the patient's entire history, not\n just from cut-over forward. The full upstream workflow is documented\n in the NCCN abstract\n BIO23-019:\n Precision Oncology - Integrating Structured Genomic Data Into the\n Electronic Health Record (Vento & Osterman, 2023).\n
\n\n\n Sharing the work with the broader community.\n Two keynote-scale talks pulled the implementation work into\n national view. At the\n American College of Medical Genetics annual meeting in\n March 2022, Dr. Osterman delivered Data to Knowledge to Wisdom:\n Improving Care Through Creative Use of Information Systems -\n framing what becomes possible when genomic results stop being\n attachments and start being structured clinical signal.\n ![\"American](\"/assets/img/logos/acmg.png\")
\n Five months later, at the August 2022 Epic User Group Meeting (UGM) -\n the largest annual gathering of Epic-using health systems - he\n presented Precision Medicine: Using Structured Genomic Data in\n Clinical Decision Support, walking other institutions through\n the build details so they could replicate it locally. The\n MyCancerGenome integration that powers the clinical interpretation\n layer was written up the year before\n (Holt et al.,\n JCO CCI, 2021).\n
\n Showing the standard on top - with students.\n Structured data inside one institution's EHR is necessary but not\n sufficient: research collaborators outside Vanderbilt still can't\n get to it without an interoperable surface. Dr. Osterman led a team\n of undergraduate Vanderbilt computer-science students through the\n first\n mCODE Genomics Pilot\n Project at Vanderbilt University Medical Center (Li et al.,\n JCO Clinical Cancer Informatics, 2024) - a working\n FHIR-on-Azure pipeline that lifts the structured genomic data out\n of the EHR and into\n mCODE-conformant\n resources downstream applications can consume. The pilot doubled\n as an early-career training experience: undergraduates learning\n cloud development, FHIR, and oncology data shaping while building\n something other implementers could reuse.\n
\n\n\n What it enables.\n Structured genomic data unlocks several adjacent decision surfaces,\n each of which has its own published track: trial-matching at scale\n (Jain et al., JCO\n CCI, 2021) showed what could be matched and what couldn't\n using NGS-result data alone; the\n molecular tumor board\n framework (Jain et al., The Oncologist, 2021) showed\n how to govern the human-review side around the structured data.\n Industry collaborations build on the same foundation -\n Tempus AI on\n reference-laboratory integration and multi-omic discovery,\n nference on research-data\n infrastructure (formalized in a\n 2023\n agreement), GE HealthCare on immunotherapy outcome prediction,\n and Microsoft on Azure-based FHIR infrastructure. None could have\n happened the same way without the integration work first.\n
\n\n\n The through-line: build the boring infrastructure first; the\n research, the AI, and the partnerships compound on top of it. See\n also cancer data\n standards (mCODE) for the standards layer that makes this data\n portable beyond Vanderbilt, and\n AI in oncology for what\n becomes possible once both the structure and the standards are in\n place.\n
\n\n\n This domain has a long-form case study:\n mCODE: from a data standard to regulatory infrastructure.\n
\n\n\n\n\n Related: all expertise domains · AI in oncology · Cancer data standards · Precision oncology · CI education · Lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/expertise/cancer-data-standards/", "url": "https://osterman.co/expertise/cancer-data-standards/", "title": "Cancer data standards (mCODE) — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Oncology data interoperability via HL7 FHIR mCODE - standards governance, regulatory uptake, and research infrastructure, chaired by Dr. Osterman.", "content_html": "International oncology data interoperability via HL7 FHIR. Standards governance, regulatory uptake, and the infrastructure that makes downstream research and quality reporting tractable.
\n\n Cancer data is generated everywhere - at the bedside, in pathology, in\n molecular labs, in radiology, in survivorship clinics - and historically\n almost none of it could travel from one institution to another without\n manual remapping. Dr. Osterman's work in cancer data standards is the\n patient infrastructure project that makes the rest of cancer informatics\n tractable: standardize the data model, get clinicians and vendors and\n regulators to agree on it, and the downstream applications - quality\n reporting, trial matching, real-world evidence, AI - all become possible.\n The arc below traces that work from early decision-support tooling to\n chairing the standard that now anchors a CMS value-based care program.\n
\n\n\n Early roots (2014-2017).\n Before mCODE existed, Dr. Osterman served as the non-small-cell-lung-\n cancer section editor for\n HemOnc.org, the open-access oncology\n regimen and drug-interaction reference under Dr. Jeremy L. Warner. The\n work was a public exercise in curating cancer information into a\n shared, machine-readable form. The collaboration produced a 2015 ASCO\n educational session and accompanying chapter\n (Rioth,\n Osterman & Warner, ASCO Educational Book, 2015) on\n website-based information resources for clinical practice. That work\n established Dr. Osterman early on as part of the small national\n community thinking carefully about structured oncology information -\n the community that, a few years later, would propose mCODE.\n
\n\n![\"mCODE](\"/assets/img/papers/mcode-diagram.svg\"\n)
\n \n 2018-2019: proposing the standard.\n Dr. Osterman joined the original 2018 mCODE working group convened to\n design a minimum interoperable cancer record on top of HL7 FHIR. The\n standard was released publicly at the 2019 ASCO Annual Meeting. Dr.\n Osterman was selected to present it in the meeting's educational\n session on big data and cancer care - the work was captured in the\n ed book chapter\n The Impact of Big\n Data Research on Practice, Policy, and Cancer Care\n (Chambers et al., ASCO Educational Book, 2019). The initial\n release coincided with then-ASCO-President Dr. Monica Bertagnolli's\n presidential address. The published version of the standard's value\n proposition followed the next year\n (Osterman, Terry\n & Miller, JCO Clinical Cancer Informatics, 2020 -\n \"The Promise of mCODE\").\n
\n\n\n 2021-present: governance and consensus.\n Dr. Osterman was appointed Chair of the mCODE Technology Review Group\n in January 2021 - the body that adjudicates additions and changes to\n the standard and ushers them through HL7's community ballot process.\n The TRG has shepherded more than two hundred community comments\n through HL7's open consensus mechanism, which is what makes mCODE\n durable rather than just one consortium's preferred shape. The\n community-facing side of that work has shown up in regular ecosystem\n updates: the 3-year Update talk at the AMIA Annual Symposium\n (Washington DC, November 2022) and Interoperability in Action:\n Progress in Implementing the mCODE Oncology Data Standard at the\n ASCO Annual Meeting (Chicago, June 2023). He now serves as Chair of\n the mCODE Executive Committee, with governance shared across the\n American Society of Clinical Oncology, the American Society for\n Radiation Oncology, the U.S. Food and Drug Administration, the\n National Cancer Institute, and the Alliance for Clinical Trials.\n
\n\n![\"HL7](\"/assets/img/logos/hl7.svg\")
\n ![\"American](\"/assets/img/logos/asco.png\")
\n ![\"U.S.](\"/assets/img/logos/fda.svg\")
\n ![\"National](\"/assets/img/logos/nci.svg\")
\n![\"Flow](\"/assets/img/papers/mcode-publish-process.svg\"\n)
\n \n Regulatory uptake and international reach.\n The executive-committee chairship's stated goal is to embed mCODE more\n deeply into the broader standards community. The clearest concrete\n outcome is CMS' Enhancing Oncology Model: every practice participating\n in EOM must submit data via mCODE, and only via mCODE. That regulatory\n hook converted mCODE from a promising consortium standard into\n infrastructure that anchors a federal value-based care program.\n Internationally, Dr. Osterman presented at the 2026 HL7 International\n Working Group Meeting in Rotterdam - opening discussions about mCODE\n as a true international cancer data standard, building on existing\n implementations in Taiwan, Brazil, and Canada. The fuller history is\n in the mCODE case study.\n
\n\n\n Implementation demonstrations.\n Standards are easier to defend in committee meetings than to deploy in\n production. Dr. Osterman led a team of undergraduate Vanderbilt\n computer-science students through the first\n mCODE Genomics Pilot\n Project at Vanderbilt University Medical Center (Li et al.,\n JCO Clinical Cancer Informatics, 2024) - a working\n FHIR-on-Azure pipeline that lifts structured genomic data out of the\n EHR and into mCODE-conformant resources downstream applications can\n consume. The project doubled as a teaching exercise: undergraduates\n learning cloud development, FHIR, and oncology data shaping while\n building something the standard's other implementers could reuse. The\n mCODEGPT\n zero-shot extraction work (Zhang et al., Communications Medicine,\n 2025) extends the same idea to unstructured text: when an LLM has a\n real schema to aim at, the outputs become trustworthy. Both papers\n exist because the standard exists first.\n
\n\n\n Dr. Osterman's standards work also runs through several related\n communities - the NCCN Digital Oncology Forum, the Epic Adult Oncology\n Steering Committee, and the Epic Beacon Community Operations Group\n (which he founded as a multi-institution governance body for the\n dominant U.S. oncology EHR module). The framing he returns to in\n talks: a standard is a force multiplier. One good cancer-data\n standard, adopted broadly, replaces hundreds of one-off institutional\n mappings, lowers the cost of every downstream AI/research/quality\n project, and is the only sustainable answer to cancer's data\n fragmentation problem. See also\n clinical genomics in the\n EHR for the implementation side and\n AI in oncology for what\n becomes possible once the standard is in place.\n
\n\n\n This domain has a long-form case study:\n mCODE: from a data standard to regulatory infrastructure.\n
\n\n\n\n\n Related: all expertise domains · AI in oncology · Clinical genomics in the EHR · Precision oncology · CI education · Lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/expertise/ai-in-oncology/", "url": "https://osterman.co/expertise/ai-in-oncology/", "title": "AI in oncology — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Machine learning and large language models for clinical decision support, efficacy and toxicity prediction, and structured data extraction at scale.", "content_html": "Machine learning and large language models for clinical decision support, efficacy and toxicity prediction, and structured data extraction at scale.
\n\n Dr. Osterman's AI work in oncology is grounded in clinical practice. He\n treats patients, sees what data is actually available, and builds models\n around that constraint rather than around what would be ideal in a\n research setting. The work breaks into four threads: a multi-year\n pre-LLM bet on real-world immunotherapy prediction, an early structured\n evaluation of clinical large language models, a recent invited framework\n piece for the field, and ongoing national thought leadership at the\n policy table.\n
\n\n\n The first thread is the five-year GE HealthCare Digital Precision\n Oncology collaboration (2019-2024), which predated the current AI\n mainstream and was built against two prevailing assumptions in the\n immune checkpoint inhibitor literature. Most prior ICI-prediction\n work used carefully curated cohorts; the DPO work used routinely\n collected EHR data so resulting models could be deployed without\n first reorganizing the underlying data pipeline. Most prior work\n focused on a single cancer type; the DPO work was deliberately\n pan-cancer, accepting the sample-size challenge in exchange\n for models that generalized. The program produced three peer-reviewed\n manuscripts, multiple abstracts, twelve conference presentations, and\n a U.S. patent (Wolber, Csernai, Kiss, Lippenszky, Horváth, Osterman et\n al.). Dr. Osterman led teams across the U.S., Hungary, and Germany,\n and presented the work internationally, including the\n ESMO Immuno-Oncology Annual Congress in Geneva. The full arc is\n covered in the Digital\n Precision Oncology case study.\n
\n\n\n The second thread began in early 2023 when ChatGPT entered the clinic.\n Dr. Osterman saw the need for an actual evaluation rather than the\n speculative discourse that dominated the conversation. He designed a\n multidisciplinary prospective trial structure for evaluating LLM\n responses to physician-authored questions and supervised\n Dr. Rachel Goodman (then a medical student) and Dr. Douglas Johnson\n in executing it. The result, Accuracy and Reliability of Chatbot\n Responses to Physician Questions\n (Goodman\n et al., JAMA Network Open, 2023), was one of the earliest\n peer-reviewed clinical LLM evaluations and has been widely cited in\n the policy debates that followed.\n
\n\n\n In August 2025, Cancer invited Dr. Osterman and colleagues to\n write Artificial intelligence across the cancer care continuum\n (Riaz, Khan & Osterman,\n 2025) - a framework piece mapping AI applications across screening,\n diagnosis, treatment selection, toxicity prediction, survivorship, and\n quality of life. The framing is intentional: AI is not one thing in\n oncology; it is many overlapping technologies at different stages of\n validation. The review now functions as one of the reference frames\n the field uses to scope its own work.\n
\n\n![\"The](\"/assets/img/papers/cancer-2025-continuum.png\"\n)
\n \n \n The work has reach well beyond academic medicine. Dr. Osterman\n delivered the keynote at the 2025 NCCN AI policy meeting\n (AI and the Cancer Journey: Navigating New Frontiers in Policy\n and Technology, September 2025) and was the AI keynote speaker at\n the 2026 NCCN Annual Conference (Harnessing Artificial Intelligence\n to Improve Oncology Care, March 2026). NCCN's audience extends\n beyond physicians to oncology nursing, industry leadership, regulators,\n and payers - the constituencies that have to act on AI-in-oncology\n conclusions, not just produce them.\n
\n The through-line across these four threads is structure. ICI prediction\n works because the EHR data underneath is curated and consistent. LLM\n evaluation matters because that's where unstructured text becomes\n downstream clinical action. The Cancer review and the NCCN\n keynotes argue the same point at policy scale: AI's clinical value\n scales with the quality of the data and standards underneath it. See\n also cancer data standards\n (mCODE) and\n clinical genomics in the\n EHR for the foundation that makes this work tractable.\n
\n\n\n This domain has a long-form case study:\n AI in oncology.\n
\n\n\n\n\n Related: all expertise domains · Cancer data standards · Clinical genomics in the EHR · Precision oncology · CI education · Lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/nccn/", "url": "https://osterman.co/collaborations/nccn/", "title": "NCCN Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "NCCN Digital Oncology Forum member and AI keynote speaker at the 2026 NCCN Annual Conference and the NCCN AI policy summit.", "content_html": "\n \n
Member of the NCCN Digital Oncology Forum and an active voice across NCCN policy and clinical channels. Featured AI keynote speaker at the 2026 NCCN Annual Conference and the NCCN AI policy summit.
\n\n Dr. Osterman is a member of the NCCN Digital Oncology Forum and an\n active voice across NCCN policy and clinical channels. The most\n visible recent work has been NCCN's AI-in-oncology agenda. In\n September 2025, Dr. Osterman delivered the keynote at the NCCN AI\n policy meeting (AI and the Cancer Journey: Navigating New\n Frontiers in Policy and Technology). In March 2026 he was the\n AI keynote speaker at the\n NCCN\n Annual Conference (Harnessing Artificial Intelligence to\n Improve Oncology Care) - reaching an audience that includes\n not only physicians but oncology nursing, industry leadership,\n regulators, and payers. The framing of regulation, integration, and\n clinical impact was published in\n JNCCN\n and featured in\n Oncodaily.\n The full arc lives in the\n AI in oncology case study.\n
\n\n\n Long-form narrative:\n AI in oncology: from ChatGPT in the clinic to mapping the care continuum.\n
\n\n\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · GE HealthCare · Epic · Microsoft · Tempus AI · nference · ai in oncology.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/nference/", "url": "https://osterman.co/collaborations/nference/", "title": "nference Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Vanderbilt Health's nference partnership (2023) - privacy-preserving research access to clinical data for real-world evidence.", "content_html": "\n \n
Vanderbilt Health's partnership with nference reduces the friction of accessing clinical data for research while preserving privacy and regulatory compliance. Formalized in a 2023 agreement.
\n\n Dr. Osterman leads Vanderbilt Health's partnership with nference,\n formalized in a 2023 agreement\n (VUMC\n News, July 2023) to advance real-world evidence generation in\n complex disease populations. The collaboration reduces the friction\n of accessing clinical data for research while preserving privacy\n and regulatory compliance - extending the broader structured-data\n agenda described in\n clinical genomics in\n the EHR.\n
\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · GE HealthCare · Epic · Microsoft · Tempus AI · NCCN · clinical genomics ehr.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/tempus-ai/", "url": "https://osterman.co/collaborations/tempus-ai/", "title": "Tempus AI Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Tempus AI partnership - structured genomic-results integration into clinical workflows, digital pathology, and multi-omic discovery.", "content_html": "\n \n
Multi-year partnership with Tempus AI spanning three workstreams: structured electronic integration of genomic results into clinical workflows, digital pathology, and multi-omic discovery infrastructure.
\n\n Vanderbilt's research collaboration with Tempus AI spans three\n workstreams: structured electronic integration of genomic results\n into Vanderbilt's clinical workflows, digital pathology, and\n multi-omic discovery infrastructure. The reference-laboratory\n integration was\n profiled\n by VUMC News in November 2024 as part of a broader effort to\n make reference-laboratory data first-class within the electronic\n health record - the same body of work that put Vanderbilt's\n structured-genomic-data EHR ahead of any other U.S. institution\n (see clinical genomics\n in the EHR).\n
\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · GE HealthCare · Epic · Microsoft · nference · NCCN · clinical genomics ehr · precision oncology.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/microsoft/", "url": "https://osterman.co/collaborations/microsoft/", "title": "Microsoft Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Microsoft Investigator Fellow (2020) - a $200K award for scaling clinical informatics education via Azure, plus Azure Research Community Advisory Board service.", "content_html": "\n \n
Microsoft Investigator Fellow (2020) - a two-year, $200,000 award for scaling clinical informatics education across UME and GME via Azure-based virtualization. Continuing service on the Microsoft Azure Research Community Advisory Board on structured genomic data, FHIR research infrastructure, and the mCODE data standard.
\n\n Dr. Osterman was selected as a\n Microsoft\n Investigator Fellow in 2020 - a two-year, $200,000 award - for\n work leveraging Azure-based virtualization to scale clinical\n informatics education across undergraduate and graduate medical\n training. The fellowship funded the cloud infrastructure that lets\n learners build and test against realistic clinical data and systems\n without the licensing and security overhead that typically makes\n hands-on informatics education impractical. The infrastructure is\n now part of the Vanderbilt CI Integrated Science Curriculum and the\n GME Clinical Informatics Elective Rotation\n (see CI\n education).\n
\n\n\n Beyond the fellowship, Dr. Osterman serves on the\n Microsoft Azure Research Community Advisory Board (AMC CAB),\n contributing on structured genomic data, FHIR-based research\n infrastructure, and the\n minimal Common\n Oncology Data Elements (mCODE™) standard. The mCODE Genomics\n Pilot at Vanderbilt\n (Li et al.,\n JCO CCI, 2024) was built on Azure, demonstrating an\n end-to-end FHIR-on-Azure pipeline lifting structured genomic data\n out of Epic into mCODE-conformant resources.\n
\n\n\n Long-form narrative:\n mCODE: from a data standard to regulatory infrastructure.\n
\n\n\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · GE HealthCare · Epic · Tempus AI · nference · NCCN · clinical informatics education · cancer data standards · clinical genomics ehr.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/epic/", "url": "https://osterman.co/collaborations/epic/", "title": "Epic Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Dr. Osterman's multi-level Epic collaboration - Adult Oncology Steering Committee, Beacon Community Operations Group founder, Genomics Module and AURA.", "content_html": "Epic is the EHR platform anchoring Dr. Osterman's clinical informatics work. Multiple collaboration surfaces: the Adult Oncology Steering Committee, the Beacon Community Operations Group (founder), Epic Genomics Module, Epic AURA, and the AI-extracted oncology staging workflow in Epic Hyperspace.
\n\n Epic is the EHR platform that anchors most of Dr. Osterman's clinical\n informatics work, and the collaboration with the company takes\n several distinct shapes: governance, build, training, and research.\n
\n\n\n Governance.\n Dr. Osterman serves on Epic's Adult Oncology Steering Committee, the\n body that influences the direction of Epic's oncology-specific\n features. He also founded the\n Epic Beacon Community Operations Group (BCOG), a\n multi-institution governance body for the dominant U.S. oncology EHR\n module - the group of academic cancer centers that coordinate on\n shared challenges and shared builds across the Beacon platform.\n
\n\n\n Build and integration.\n Three implementation pieces have shipped at Vanderbilt under\n Dr. Osterman's leadership: the\n Epic Genomics Module as the structured-variant home inside\n the EHR; Epic AURA as the bridge that pulls\n reference-laboratory genomic reports in over standardized interfaces;\n and the AI-extracted oncology staging workflow being built into Epic\n Hyperspace (Dr. Osterman is PI). The implementation work is\n documented in\n Li et al., JCO\n CCI, 2024 (mCODE Genomics Pilot) and the\n Vento &\n Osterman NCCN abstract (2023) on the upstream workflow.\n
\n\n\n Speaking and community.\n Featured speaker at Epic's annual User Group Meeting (UGM) and\n Expert Group Meeting (XGM) on community governance, structured\n genomic data backloading, in-basket workflow, and other topics\n where the BCOG cross-institution conversations have generated\n reusable patterns. The 2022 UGM talk\n Precision Medicine: Using Structured Genomic Data in Clinical\n Decision Support is the canonical public version of the\n Vanderbilt build story.\n
\n\n\n Long-form narrative:\n mCODE: from a data standard to regulatory infrastructure.\n
\n\n\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · GE HealthCare · Microsoft · Tempus AI · nference · NCCN · clinical genomics ehr · cancer data standards · precision oncology.\n
\n\n\n" }, { "id": "https://osterman.co/collaborations/ge-healthcare/", "url": "https://osterman.co/collaborations/ge-healthcare/", "title": "GE HealthCare Collaboration — Travis Osterman", "date_published": "2026-05-28T19:49:00.000Z", "summary": "Five-year GE HealthCare research collaboration - predicting immunotherapy efficacy and toxicity from real-world EHR data; Dr. Osterman was PI.", "content_html": "\n \n
Five-year strategic research collaboration on predicting immune checkpoint inhibitor efficacy and toxicity from real-world EHR data. Dr. Osterman served as principal investigator on the flagship Digital Precision Oncology study.
\n\n Vanderbilt and GE HealthCare announced a five-year strategic research\n collaboration in January 2019\n (dotmed.com)\n centered on applying machine learning to predict immune checkpoint\n inhibitor (ICI) effectiveness and toxicity from real-world EHR data.\n Dr. Osterman served as principal investigator from Vanderbilt; Dr. Jan\n Wolber led on the GE HealthCare side. Multi-disciplinary teams on\n both sides spanned medical oncology, radiochemistry, biostatistics,\n bioinformatics, and machine learning. The collaboration ran through\n 2024 and produced three peer-reviewed manuscripts, twelve conference\n presentations, multiple abstracts, and a U.S. patent on the\n underlying model-generation framework.\n
\n\n\n The flagship paper -\n Prediction of\n Effectiveness and Toxicities of Immune Checkpoint Inhibitors Using\n Real-World Patient Data\n (Lippenszky et al., JCO Clinical Cancer Informatics, 2024) -\n demonstrated that routinely-collected EHR data is enough to build\n clinically useful prediction models across both efficacy and\n toxicity, without bespoke biomarker panels or unstructured-data\n extraction pipelines. The imaging companion paper extended the\n framework to pre-treatment thorax CT\n (Smith et al.,\n JCO CCI, 2025) using radiomic features and deep learning\n to predict ICI-induced pneumonitis from 2,700 CT volumes.\n
\n\n\n Beyond Vanderbilt the work was externally validated on multi-center\n German pan-cancer cohorts in collaboration with the Essen-Bochum\n team (Kiss, Lippenszky et al., SITC 2023). The full narrative arc -\n what the team bet against the prevailing assumptions of the field\n (real-world data over curated cohorts, pan-cancer over single-cancer),\n what shipped, and what's next - lives in the\n Digital Precision\n Oncology case study.\n
\n\n\n Long-form narrative:\n Digital Precision Oncology with GE HealthCare.\n
\n\n\n\n\n This collaboration is publicly disclosed through two independent records:\n ASCO COI and\n CMS Open Payments.\n
\n\n\n Related: all collaborations · Epic · Microsoft · Tempus AI · nference · NCCN · ai in oncology · lung cancer.\n
\n\n\n" }, { "id": "https://osterman.co/christmas/", "url": "https://osterman.co/christmas/", "title": "Christmas archive", "date_published": "2026-05-26T20:48:05.000Z", "summary": "Annual Osterman family Christmas pages and newsletters. Older years are gated; the current year is public.", "content_html": "\n\n Annual Osterman family Christmas pages and newsletters. The current year\n is open; older years are gated.\n
\n![\"\"](\"/assets/img/wp/uploads/2025/11/reface-2025-11-06-05-01-26.png\")
Welcome to our Christmas 2025 update. It has been another exciting year. Be sure to check out our 2025 newsletter for details.
\n![\"\"](\"/assets/img/wp/uploads/2025/11/PSX_20250101_225325-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250220_183656567-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250111_224527376.PORTRAIT-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/edfd3522-bda2-47f8-95b4-1a0be21df1a2.jpeg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250220_172933981-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250304_221011658-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250329_204246882-720x540.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20250903_220524080.MP_-405x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20251008_151706886-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20251008_152239490-540x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20251031_221242535-720x540.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20251018_172859141-720x540.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/Screenshot_20251021-175806-707x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/629a49872-480x720.jpg\")
![\"\"](\"/assets/img/wp/uploads/2025/11/PXL_20251130_213718285.png\")
Willow (7)
\nWillow is a busy second-grader. In addition to school, she has enjoyed dance this year with lots of new friends. She continues to be the artist and crafter of the family and can often be found surrounded by markers, glue, tape, glitter, pipe cleaners, and piles of scrap paper.
\nGavin (9)
\nGavin is now in third grade. His highlights include becoming an expert modded Minecraft player. He loves exploring new worlds with his family, learning all the winding questlines, and gathering chests filled with the best loot. He continues to be active in scouts and swimming.
\nOwen (13)
\nOwen is loving seventh grade with highlights of joining the Brentwood Middle School Student Government Association and making a Southeastern cut time in swimming. He has enjoyed bringing his unique brand of junior high humor to the dinner table.
\nLaura
\nLaura has enjoyed immersing herself in the new world of 3d printing. She also coordinated getting all the McEwens to a Blue Jays game during their historic 2025 season. She continues to be an active volunteer at the schools and in extracurricular activities. Finally, she led the campaign to add a new member to the family (pictured above).
\nTravis
\nThis has been a hard and reflective year for Travis. Brenda Osterman, his mother, died. She was an amazing, kind, and giving mother, sister, wife, and grandmother who fought Parkinson's Disease for twenty-five years. She was a testament to seizing each day and an inspiration to those around her, as reflected in the outpouring of family, friends, and community at her memorial services. Life continues but differently.
\nChristmas Present
\nDr. William Weaver is the Chief of Surgery at the Fayetteville VA Medical Center in North Carolina. He has had two recent stories covered by NPR. In the first, he recounts being one of 14 black students integrating the all-white West High School in Knoxville, TN. I heard the second story today during my commute and was moved by his Story Corps recording recounting the theft of his brother's bike during Christmas when he was a child. This story of family, giving, and forgiveness comes at a great time. Thank you NPR and thank you Dr. Weaver for sharing.
\n\n" }, { "id": "https://osterman.co/good-enough-dad/", "url": "https://osterman.co/good-enough-dad/", "title": "Good Enough Dad", "date_published": "2013-03-25T00:00:00.000Z", "summary": "A 2013 post about the NPR StoryCorps interview between Tim Harris, a restaurant owner with Down syndrome, and his father.", "content_html": "![\"Tim](\"/assets/img/wp/uploads/2013/03/tim-harris-150x150.jpg\")
National Public Radio (NPR) runs a series called Story Corps which is simply recordings of people interviewing people. The link at the top is probably my favourite interview I've heard all year, in any context. Tim is a college graduate with Down's Syndrome and is owner of Tim's Place in Albuquerque, New Mexico. The interview is mostly a father interviewing his son. If you're a father or son or both, the interview is three minutes and thirteen seconds long and you'll probably get less than sixty seconds in before tearing up.
\n\n" }, { "id": "https://osterman.co/hydroponics-ebb-and-flow-cost/", "url": "https://osterman.co/hydroponics-ebb-and-flow-cost/", "title": "Hydroponics - Cost To Get Started with Ebb and Flow Hydroponics", "date_published": "2012-10-29T00:00:00.000Z", "summary": "A 2012 cost breakdown for a home ebb and flow hydroponics system — lights, flood table, pump, nutrients, and aeration — totaling about $600.", "content_html": "
To get started with hydroponics at home, you should take an inventory of your goals and budget to begin. For a 48x24 inch ebb and flow (flood and drain) system, Your total cost will be around $600. If you grow potatoes, it will take quire a while to recoup this cost. Basil, however, at $3 for a young, fresh plant for pesto, will probably return your investment in as few as 2 years (6 inch diameter per plant = 8x4 plants in table = 32 plants x $3/plant = $96/harvest / 3 months maturation = $32/month - electricity - more nutrient - seeds = ~$25/month). If you choose to use all organic and can find a buyer, expect a much more lucrative return.
\n![\"hydroponics](\"/assets/img/wp/uploads/2012/08/ebb-flow.png\")
\nIf you don't have $600 laying around, you can step into hydroponics more slowly by starting with a deep water culture system or smaller ebb and flow setup. I'll argue though that planning a setup this size and buying it incrementally will make the most sense. We'll breakdown the various components here.
\nThe Amazon product widgets originally embedded here no longer render. Product links: T5 grow light fixtures (1, 2).
\nStarting with the light source makes the most sense. Begin by purchasing the most lumens per square foot that you can afford. Note that will not necessary be the largest fixture. A 2x2 ft fixture that packs in many bulbs is just as good. Direct sunlight is around 9000-10,000 lumens per square foot (or 100,000 lumens/sq meter) so that is our target.
\nThe fixture show above contains 8 T5 bulbs that will produce 40k lumens each or 320K lumens total over an area of approximately 4x2 feet (8 sqft). That makes 40,000 lumens per square foot. After accounting for inefficiency in reflection for light that is initially directed up, light that is lost horizontally, and decreased efficacy of the bulb as it ages, we should still be around the 10k/sqft sunlight equivalent. We'll ignore wavelengths for now and say that T5 is a good bulb choice because they are cool enough to be very close to the plant canopy.
\nYou can germinate plants with much less light. This amount of light will allow your plants to thrive, bloom, and fruit however. Once you have this set up, try it for a few months on a traditional potted plant and you'll be sold.
\nTotal: $200
\nThe Amazon product widgets originally embedded here no longer render. Product links: flood table and grow media (1, 2, 3).
\nThe next large investment in your ebb and flow hydroponics setup will be a flood table, reservoir and grow medium. From the diagram above, hopefully you can get the idea that your reservoir can be just about anything that holds water. I scoped out craigslist for a few months as was able to find a fifty-five gallon aquarium for $50 that holds water just fine.
\nThe flood table will either need to be supported by the reservoir or by some sort of stand. Since the flood table will contain grow media which will displace some of the water, the volume to flood the the flood table will be less (maybe 50% less) that calculated volume. Assuming that the reservoir is covered to minimize evaporation loss, You can begin to estimate the maximal amount of time between filling the reservoir. This is also known as how long you can go on vacation.
\nA 48 x 24 x 6 flood table is 6912 cubic inches or just under 30 gallons. If we assume 50% displacement but grow media, we'll need 15 gallons to flood the table. Let us add in 2 gallons of water to keep the pump submerged and the minimum water requirement would be 17 gallons. Since our reservoir is 55 gallons, if we fill it completely, will will have 38 gallons of "extra" water that we will either loose to our plants or evaporation.
\nDifferent plants have different needs for water. Lettuce is notoriously good for large scale hydroponics as it sucks a large amount of water. This isn't to say that you can't raise hydroponic cactus though. Hydroponics Made Easy claims a young tomato plant will use 300ml-1L of water per day. We would have 96 plants in a full bed or 28.8-96L/day (7.6-25gal/day). So this system probably has between a 1-7 refill period depending on plant type, maturity, and density.
\nThe hydroponic grow media simply serves a means to keep roots moist but not wet (like a sponge. Hydroton is a commercially available clay pellet that is commonly used.
\nTotal: $200
\nThe Amazon product widgets originally embedded here no longer render. Product links: water pump and timer (1, 2).
\nThe water pump is the key component of the ebb and flow (flood and drain) hydroponic system. The goal is to have a pump capable of raising an appropriate volume per minute to flood your table relatively quickly but not so fast as to cause erosion of your grow media. The higher your flood table is from the base of your reservoir and the larger your flood table, the larger the pump your need. The larger the pump, the (generally) louder it will be and more electricity it will consume; although nowhere near the consumption of your grow lights.
\nThe water pump will be on a timer to flood your plants several times daily while your grow lights are on.
\nTotal: $50
\nThe Amazon product widgets originally embedded here no longer render. Product link: hydroponic nutrients (a second product is no longer listed).
\nPlants cannot live on water alone. The nutrients, associated pH solutions, water, and electricity will be the main recurring expenses.
\nTotal: $50
\nThe Amazon product widgets originally embedded here no longer render. Product links: air pump, air stones, and check valve (1, 2, 3, 4).
\nYou can put off aeration to the end but you will not want to put this off too long. like your plants, algae will have all the ingredients to grow. Changing your water at regular intervals and keeping your water aerated will help combat this thread.
\nRemember that air stones will help get more of the gas dissolved into the water so plan on picking up a few air stones. Also, a one-way check valve is a good idea if your pump is not submersible as there is always a chance of backflow of water when the system is turned off.
\nAeration, unlike lighting and the water pump can continue 24 hours a day. Make sure that your air outlet are not is the same area as your water pump intake as this will decrease efficiency and life of your water pump.
\nTotal: $30
\n![\"enders](\"/assets/img/wp/uploads/2012/10/enders_game-150x150.jpg\")
September and October have now become the months that one interviews for the internal medicine sub-specialties. For years I have been a fan of using my daily commute and extended travel times to "read" audiobooks or listen to medical lectures. I have been on the road over the past two months and decided with my last trip out of state to break up my MKSAP study with a leisure novel - Ender's Game. This will not be a book review. I am neither well-read enough to make insightful comparisons nor well written enough to write a critique to far outside my comfort zone of technical writing. One idea did come to me as I sat in the parking lot after arriving at work, listening to the final ten minutes of Orson Scott Card's commentary. He emphasizes that this (the audiobook) is the medium that he would most like his stories to be shared.
The Amazon product widget originally embedded here no longer renders. The product: Ender’s Game audiobook on Amazon.
\n![\"josh](\"/assets/img/wp/uploads/2012/10/josh-stuart-150x150.jpg\")
Stand Up 2 Cancer has several dream teams organized to help oncology research. Today, I came across an article from UCSC discussing how one of their associate professors in biomolecular engineering, Joshua Stuart (right), is joining as a member of the prostate cancer dream team. It looks like Stuart is already active in the breast cancer dream team.
From the article:
\nA second component of the UCSC team's work will be to develop a novel application called MedBook to connect patients, doctors, and researchers. Inspired by social media, MedBook is the brainchild of UCSC graduate student, Foundation trustee, and former Apple executive Ted Goldstein, who has described it as a \"Facebook for cancer.\" The idea is to bring together the findings from patient tumor samples and connect researchers in a way that will help them work collaboratively to discover new models of disease and match patients with targeted treatments.
\n\"The precision prostate oncology MedBook will act like an iPad app, linking patients, doctors, and researchers in a participatory information platform that will speed the matching of new medicines targeted at new druggable pathways to the patients who will benefit most,\" said Jonathan Simons, M.D., president and CEO of the Prostate Cancer Foundation.
\nA few other informaticists are already participating in the dream team projects:\nArul M. Chinnaiyan, M.D., Ph.D.\nDirector of Pathology Research Informatics\nDirector of Cancer Bioinformatics\nThe Regents of the University of Michigan
\nDavid Cherba, Ph.D.\nVan Andel Research Institute (VARI)
\nJason Corneveaux\nTranslational Genomic Research Institute (TGen)
\nThis is probably an under-representation of the importance that informatics plays in oncology, but we are getting there. Kudos to everyone providing support as we create tools to help us ... Stand Up 2 Cancer.
\n" }, { "id": "https://osterman.co/stanford-oncology-informatics/", "url": "https://osterman.co/stanford-oncology-informatics/", "title": "Stanford Oncology Informatics", "date_published": "2012-10-01T00:00:00.000Z", "summary": "A 2012 note on two oncology informatics leaders: Robert Carlson named NCCN CEO and George Sledge moving from Indiana University to Stanford.", "content_html": "![\"stanford](\"/assets/img/wp/uploads/2012/10/stanfordlogo-150x150.jpg\")
Last week was a busy week for The Cardinal. Two days after Dr. Robert Carlson was announced as the new CEO of the National Comprehensive Cancer Network (NCCN), the Indianapolis Star announced that Dr. George Sledge will be leaving Indiana University to become the division chief of hematology and oncology at Stanford.
![\"dr](\"/assets/img/wp/uploads/2012/10/carlson.png\")
Both distinguished clinician researchers have a history of breast cancer focus with academic interests at the interface of medical oncology and informatics (oncology informatics). Dr. Carlson has interests in:
Computer science research efforts focus on computer-based systems for patient care. These include: the development of expert systems to assist in the care of patients with cancer (ONCOCIN) or AIDS (T-Helper); the development of computer-based systems using speech and pen-based input for medical records to access sources of medical knowledge; the design of three-dimensional methods of performing and representing literature searches; and the development of standards for implementing of medical computer-based systems; the development of smart agents and medical ontologies.
\n![\"dr](\"/assets/img/wp/uploads/2012/10/sledge.png\")
Dr. Sledge was the ASCO 2011 president and used the opportunity of his presidential address to discuss The Challenge and Promise of the Genomic Era. I had the privilege of working with Dr. Sledge briefly in his breast cancer clinic and it helped solidify oncology as a career path for me. He loves and is loved by his patients, is an eloquent teacher, and has a fantastic sense of humor. He will be missed at Indiana University. I wish Dr. Sledge the best as he embarks on this next phase in his decorated career and thank both for their work championing oncology informatics.\n"
},
{
"id": "https://osterman.co/can-i-break-even-selling-ads-on-my-website/",
"url": "https://osterman.co/can-i-break-even-selling-ads-on-my-website/",
"title": "Can I Break Even Selling Ads On My Website",
"date_published": "2012-09-06T00:00:00.000Z",
"summary": "A 2012 look at one year of Google AdSense data from a small Linux tutorial site, with the math on how many posts it takes to break even.",
"content_html": "![\"ad](\"/assets/img/wp/uploads/2012/08/money-bag-with-dollar-sign-150x150.jpg\")
I have been using Google AdSense on one of my websites, Gentoovps.net, for one year now and wanted to answer the question, “Can I Break Even Selling Ads On My Website?” Just to be clear, I did not create the site for monetary purpose. I did not troll for the highest cpc keywords, try to trick users into clicking on ads, or anything else shady. The entire purpose of integrating ads into the site was to offset a small amount of the cost of hosting and domain registration.
![\"dollars](\"/assets/img/wp/uploads/2012/08/dollars-per-post-per-year.jpg\")
\nThis assumes though that older articles are still being viewed (whether being found via search, linked to, or navigated to once found by a user). To test this, I looked at a simple correlation between monthly pageviews and time (as a surrogate for total posts). Here is the scatter plot with a horrible correlation.
\n![\"can](\"/assets/img/wp/uploads/2012/08/monthly-page-views-total.jpg\")
\n![\"can](\"/assets/img/wp/uploads/2012/08/page-views.jpg\")
\nSo the more content you produce, the more page views you receive (as expected).
\n![\"time](\"/assets/img/wp/uploads/2012/08/time-warp-300x294.jpg\")
If domain registration plus annual hosting costs $150 and each post generates $0.15/year based on the above formula, then $150/$0.15 = 1000 posts.\nThere are a few caveats however. Scaling a website from 40 posts to 1000 posts (or 50 page views per day to 1250 page views per day) may require more resources (which cost more money). Remember also that at no point have you begun to compensate yourself for the time of writing the articles. If each post takes 3 hours, you are looking at a 3000 hour time investment to save $150/year. Right now, at the current US minimum wage ($7.25), that is an investment of $21,750 which would require 41 million page views to offset.
\nA pubmed search for electronic sign out (or hand off) yields 42 results. I wrote recently about secure patient health information and have been thinking more about what it takes to have a secure, physician-to-physician electronic sign out system. To begin, we'll assume two things: 1) no integration with existing systems. This is not a small assumption but I would like to work on a high level and ideally a hand off could be generated by proprietary EMR X and submitted in a secure way via the following system. 2) The system and interface should be as simple as possible to achieve the minimum required security. The definition of minimum security when it comes to signing out patient information is murky at best. Here is my solution in infographic form.
\n![\"secure](\"/assets/img/wp/uploads/2012/08/secure-electronic-hand-off.png\")
Immediately, you probably assume I have broken assumption #2 from above. The above diagram describes a system of accessing (and sharing) encrypted data through 3 layers of encryption. Three layers are not more beneficial from a data security standpoint but the process of changing over patients assumes multiple providers accessing the data, changing passwords, etc. Also, we should address problems like session hijacking, database hacking, and password compromise.\nPrimary data key - this is a random key that is generated for each unique data field to be encrypted. This key is used to encrypt the data itself. If the key is compromised only one unit of data are revealed. An encrypted version is stored in the database.
\nSecondary data key - this is a random key that is generated for each method of accessing each unit of encrypted data. Since each primary data key will be accessed by two ways per user, the total number of database entries is at least data units x 2. If the key is compromised only one unit of data are revealed. An encrypted version is stored in the database.
\nMaster user key - this is a random key that is generated once on creation of a user account and serves to be the initial key by which a secondary data key is encrypted. If the key is compromised all the user's data are revealed. An encrypted version is stored in the database.
\nSalted password hash - a salted password hash is used as the key to encrypted the master user key. In the event that a user needs to change his or her password, the only object that requires updating is the encrypted signature of the master key. This eliminates any lag as a result of changing passwords.
\nSession key - this is not mandatory, but may reduce exposure to session hijacking. Instead of placing the master key in the session to decrypt data, a session-specific key is created (along with a new set of secondary data keys). At conclusion of the session or after an expiratory period, the session key and secondary data keys can be removed.
\nTemp data key - in order to share information, a secondary data key can be cloned and encrypted using a temporary data key. This key can then be transmitted to the second user. Upon accessing the data, the second user will automatically create a secondary data key encrypted with their own mater key. At this point, two users can access hand off (or sign out) data which is encrypted with a single, unknown, primary data key.
\nIt is probably not ideal yet, but this process is superior to sending changeover via email or simply using whole disk encryption.
\n" }, { "id": "https://osterman.co/why-you-cant-be-anything-you-want-to-be/", "url": "https://osterman.co/why-you-cant-be-anything-you-want-to-be/", "title": "Why You Can't Be Anything You Want to Be", "date_published": "2012-08-23T00:00:00.000Z", "summary": "A 2012 reflection on Sean Johnson's essay arguing the market doesn't care what you love, and on finding work you wake up wanting to do.", "content_html": "My grandfather worked as a carpenter from the time he was a teenage apprentice until a couple years before his death. I loved making projects with him, and he taught me a lot about woodworking. One of the most striking things about him was that he loved going to work. Whether it was remodeling a kitchen, making a custom bench, or repairing a deck, he loved it. He told me once that he never woke up and told himself that he didn't want to go into work.
\n![\"doing](\"/assets/img/wp/uploads/2012/08/132544265116-300x262.jpg\")
Sean Johnson wrote an intriguing article, Why You Can't Be Anything You Want to Be, at his blog this month. It does fly in the face what my generation has been told about pursuing career happiness.
Reading the article, I feel lucky to be able to purse a career that makes me happy, seems to be in demand, and that I'll want to wake up and go do. Sean's five themes are as follows:
\n[Peter, Michael, and Samir are chatting as they hang around the printer]\nPeter Gibbons: Our high school guidance counselor used to ask us what you'd do if you had a million dollars and you didn't have to work. And invariably what you'd say was supposed to be your career. So, if you wanted to fix old cars then you're supposed to be an auto mechanic.\nSamir: So what did you say?\nPeter Gibbons: I never had an answer. I guess that's why I'm working at Initech.\nMichael Bolton: No, you're working at Initech because that question is bullshit to begin with. If everyone listened to her, there'd be no janitors, because no one would clean shit up if they had a million dollars.\nSamir: You know what I would do if I had a million dollars? I would invest half of it in low risk mutual funds and then take the other half over to my friend Asadulah who works in securities...\nMichael Bolton: Samir, you're missing the point. The point of the exercise is that you're supposed to figure out what you would want to do if...\n[printer starts beeping]\nMichael Bolton: \"PC Load Letter\"? ...\n
Who knows, maybe Rich Dad Poor Dad will soon be required reading before entering college. Figuring out the best way to find happiness at the bazaar where we all swap time, money, and ideas has never been easy for any generation. I'd be happy to have a career like my grandfather and wake up ready to go to work each morning. Medicine has been great at that so far.
\n" }, { "id": "https://osterman.co/diuretics-and-heart-failure/", "url": "https://osterman.co/diuretics-and-heart-failure/", "title": "Diuretics and Heart Failure", "date_published": "2012-08-20T00:00:00.000Z", "summary": "A 2012 archival medical self-quiz on diuretics in heart failure, originally an interactive question on the old WordPress site.", "content_html": "![\"medical](\"/assets/img/wp/uploads/2012/07/Standardized-Testing.jpg\")
\n[slickquiz id=3]\n![\"mythTV-logo\"](\"/assets/img/wp/uploads/2012/08/mythTV-logo.jpg\"){kind=logo}
Recently, I was in needed of an Nvidia pci video card for a new installation of Mythtv I put together for watching the olympics. Last Christmas my sister and I gave our parents a new computer so I inherited the old one for re-purposing. I didn't want to dump a lot of money into the box so I needed a hardware pci capture card to record over the air (OTA) hdtv and an nvidia pci video card with vdpau (or purevideo) support.
![\"Zotac-ZT-40605-10L\"](\"/assets/img/wp/uploads/2012/08/zotac-gt-430-pci-vdpau-video-card.jpg\")
After checking wikipedia's purevideo page, I started to look up chipset and see if any manufacturer had put a gpu with complete acceleration for MPEG-1, MPEG-2, MPEG-4 Part 2 (a.k.a. MPEG-4 ASP), VC-1/WMV9 and H.264 onto a pci card, made it fanless, and sold it for cheap. There was really only one card I could find. The Zotac ZT-40605-10L fit the bill. Total cost was around $60 after rebate and decodes full 1080p with minimal use of system resources. It supports 4th generation vdpau and while I may have the option to use pci-express in a future build, I will almost certainly have an open pci slot. There were some concerns that 512mb of video ram might be insufficient, but I haven't run into this problem.\n![\"dvico](\"/assets/img/wp/uploads/2012/08/pci-capture-card-150x150.jpg\")
Before the nationwide conversion to OTA HD, I loved hauppage capture cards for how easy they were to install so I was hesitant to stray away at first. The PCI requirement was, again, a difficult one. I ended up finding a dvico fusion hdtv5 RT Lite on ebay for less than $20 (including shipping). It works great and I couldn't be happier. I bought a thin, $40, HD antenna and the combination is perfect.\nInstallation was a bit harder than I remember the pvr 250 being from many years ago but it works and I don't plan on changing to anything else for now. The other benefit of this card is that if you wanted to put two or three of them into a single system, the motherboard wouldn't cost you an arm and a leg.
\n\n
![\"fda](\"/assets/img/wp/uploads/2012/07/pharmacy-280x300.jpg\")
Cefepime Seizure Risk56% of these cases involved patients >65 years of age (range: 7-95 years) and 69% of the 59 cases involved female patients. Renal dysfunction was present in 58/59 patients (renal status was unknown in one patient). In 56/59 patients, cefepime dosing was not appropriately adjusted for renal impairment as recommended in the cefepime label.\n\nYou've probably been dose-adjusting cefepime already, this just adds even more data for your support.\n\n\n
- FDA Drug Safety Communication
\n" }, { "id": "https://osterman.co/coffee-and-medicine/", "url": "https://osterman.co/coffee-and-medicine/", "title": "Coffee and Medicine", "date_published": "2012-08-09T00:00:00.000Z", "summary": "A 2012 archival note rounding up studies on coffee and health, including the NEJM analysis linking coffee drinking to lower all-cause mortality.", "content_html": "
Coffee and medicine go together like ... well ... coffee and doctors. ![\"travis-coffee-company\"](\"/assets/img/wp/uploads/2012/08/travis-coffee-company.jpg\")
I have loved coffee at least since medical school - most of us succumb to the peer pressure of drinking coffee while studying medicine. My earliest memory of coffee was sitting at my grandparents home where my grandmother would let me have the last few sips of her coffee as a child. Today, almost randomly, I came across the Travis Coffee Company. You can imagine my excitement.
This is the coup de gras paper and should be kept in tablet or smart phone of any avid coffee drinker to defend the, "don't you think that's too much coffee?" Freedman, et al showed in the article Association of Coffee Drinking with Total and Cause-Specific Mortality, published New England Journal of Medicine May, 2012, showed the following:
\nAdjusted hazard ratios for death among men who drank coffee as compared with those who did not were as follows: 0.99 (95% confidence interval [CI], 0.95 to 1.04) for drinking less than 1 cup per day, 0.94 (95% CI, 0.90 to 0.99) for 1 cup, 0.90 (95% CI, 0.86 to 0.93) for 2 or 3 cups, 0.88 (95% CI, 0.84 to 0.93) for 4 or 5 cups, and 0.90 (95% CI, 0.85 to 0.96) for 6 or more cups of coffee per day (P<0.001 for trend); the respective hazard ratios among women were 1.01 (95% CI, 0.96 to 1.07), 0.95 (95% CI, 0.90 to 1.01), 0.87 (95% CI, 0.83 to 0.92), 0.84 (95% CI, 0.79 to 0.90), and 0.85 (95% CI, 0.78 to 0.93) (P<0.001 for trend).\n
The best part (for me) in regards to the above paper is that the more coffee you drank (at least to six cups a day), the larger the benefit. I think it's about time for a refill. Coffee is not a medicine and is not FDA approved for primary prevention of any disease. Almost every other addictive substance has a long list of warnings and adverse effects. It's nice to have data supporting moderate to excessively indulgence. Cheers.
\n![\"coffee-stack\"](\"/assets/img/wp/uploads/2012/08/coffee-stack.jpg\")
\n"
},
{
"id": "https://osterman.co/hostigation-review/",
"url": "https://osterman.co/hostigation-review/",
"title": "Hostigation Review",
"date_published": "2012-08-06T00:00:00.000Z",
"summary": "A 2012 review of Hostigation VPS hosting after one year, covering uptime, support response time, and why the service earned a renewal.",
"content_html": "![\"hostigation\"](\"/assets/img/wp/uploads/2012/07/vps-hosting-150x150.jpg\")
\nMarch 31, 2012 the network was down for aprpoximately 5 hours. Quadranet, who provides the physical servers, were sorting through a DDoS attack. Since I could not access their main site, I sent a tweet to @hostigation and got a twitter reply 53 minutes later explainging the situation. I thought that was great response time. The DDoS attack was clearly out of their control and while, I suppose, they could have tried to update dns records to an alternative server to explain the situation, I’m not paying for enterprise-level support.\n
One month is not a fair amount of time to really judge a hosting providing. Over the past year, everything has just worked. Being able to boot right into the gentoo live cd instead of booting ubuntu into recovery mode and installing gentoo is great. There are other cheaper or faster vps providers out there. Tim appears to be a very competent sysadmin and with 99.9% uptime, speedy service when I need it, and good communication, my money will stay here.\n\nI’ll update if anything changes.
![\"double](\"http://www.allergyclinic.co.uk/images/double-blind.jpg\")
The short answer is probably not - the randomized control trial is safe for now. The longer answer when it comes to exploring molecular targets in oncology we may need to rethink the double-blind randomized control trial.
Since current methods to assess potential cancer treatments are cumbersome, expensive, and often inaccurate, can we develop other methods to rapidly test interventions for cancer treatment or prevention?\n\n*NCI provacative questions\n\nAdvances in 3-dimensional cell culture suggest that multiple cell types can be assembled in vitro and that engineered tissues often mimic many of the features of human organs. If systems can be developed that mimic the natural environment of tumors, perhaps these models will recapitulate drug action. It also seems possible that complex cell-free systems could be developed that would recapitulate at least some features of drug responses. Since it seems unlikely that any one new system will serve as an accurate model for all tumors, each may need to be tuned to the particular features of a particular tumor type or subtype.
\n\n
- NCI PQ 17*
![\"lung-cancer-genome\"](\"/assets/img/wp/uploads/2012/07/lung-cancer-genome.jpg\")
![\"genome](\"http://images.forbes.com/media/2010/06/03/0603_chart-cost-human-genome_398x371.jpg\")
\nOnce a driver mutation is identified in animal models, it will be increasingly difficult develop targeted therapies if the rate of mutation in the tumor is small. In the past, it was cost prohibitive to sequence many different tumors looking for these less common mutations. Times are changing and our technology is likely outpacing our investigative model.
\n\n" }, { "id": "https://osterman.co/windows-portable-sync-comparison/", "url": "https://osterman.co/windows-portable-sync-comparison/", "title": "Windows Portable Sync Comparison", "date_published": "2012-07-30T00:00:00.000Z", "summary": "A 2012 comparison of three portable Windows file-sync tools (DSynchronize, Toucan, Synkron) testing true bidirectional sync with deletes.", "content_html": "
I looked at three different portable solutions that do not require installation or administrator privileges.
\n![\"dsynchronize\"](\"/assets/img/wp/uploads/2012/07/dsynchronize_icon.gif\"){kind=icon}
Dsynchronize was the first application I tried. The interface is included here. I was optimistic based on the option for \"bidirectional synchronization\" and the box at the bottom the allowed you to be prompted prior to file deletion. The synchronization process was fast and modification and new files were propagated easily. When I deleted a file on the \"destination\" folder, the file was recreated during the next sync. Deal breaker.![\"dsynchronize\"](\"/assets/img/wp/uploads/2012/07/dsync2.png\")
\n\n
![\"Toucan\"](\"/assets/img/wp/uploads/2012/07/Toucan_128.png\")
Toucan by far had the most features and well-polished interface of the three I reviewed. At first I could not understand why it did not have a sync function but after reading the documentation online, \"equalise\" is essentially the same. Toucan sports some very nice backup options including encryption and file renaming based on variables like date and timestamp. As expected, modification and new file creation were no problem, but the \"source\" and \"destination\" model tripped up design (from my standpoint) by recreating deleted files on the destination side. So, while beautiful, a deal breaker.![\"Toucan\"](\"/assets/img/wp/uploads/2012/07/Toucan2.png\")
\n![\"Synkron_header2\"](\"/assets/img/wp/uploads/2012/07/Synkron_header2.png\")
As I'm starting the think the only way to fix this problem is to run a set of Windows batch scripts, I come across synkron. The interface is fine and I can declare multiple \"master\" folders through the main interface. I stopped exploring, ran a sync, then deleted a file on the network drive. Upon resync, proof, appropriate deletion. We instantly had a winner. Since then, it raised in my book even higher when I clicked the restore tab and have the ability to restore recent versions of the deleted files. Fantastic. I can set it to sync and start and it will synchronize by work computer with my network drive at work and my usb stick. Just what I wanted.![\"Synkron2\"](\"/assets/img/wp/uploads/2012/07/Synkron2.png\")
\nOndansetron (Zofran) Prolonged QTThe most common scenario for using the higher dose is as an anti-emetic administered during chemotherapy. The updated label will state that no single intravenous dose should exceed 16mg.
\nThe new information on QT prolongation does not change any of the recommended oral dosing regimens for ondansetron. It also does not change the recommended lower dose intravenous dosing of ondansetron to prevent post-operative nausea and vomiting.
\nThe use of a single 32 mg intravenous dose of ondansetron should be avoided. New information indicates that QT prolongation occurs in a dose-dependent manner, and specifically at a single intravenous dose of 32 mg.\n\n\nPatients who may be at particular risk for QT prolongation with ondansetron are those with congenital long QT syndrome, congestive heart failure, bradyarrhythmias, or patients taking concomitant medications that prolong the QT interval
\nElectrolyte abnormalities (e.g., hypokalemia or hypomagnesemia) should be corrected prior to the infusion of ondansetron.
\nThe lower dose intravenous regimen of 0.15 mg/kg every 4 hours for three doses may be used in adults with chemotherapy-induced nausea and vomiting. However, no single intravenous dose of ondansetron should exceed 16 mg due to the risk of QT prolongation.
\nThe new information does not change any of the recommended oral dosing regimens for ondansetron, including the single oral dose of 24 mg for chemotherapy induced nausea and vomiting.
\n\n
- FDA Drug Safety Communication
\n[slickquiz id=2]
![\"raspberry](\"/assets/img/wp/uploads/2012/07/raspberry-pie-300x232.jpg\" "\\"raspberry")
My first Raspberry Pi post was quite popular so I thought I would do a second. You might have caught that production has ramped up and as more people get their hands on a Raspberry Pi, we are seeing more projects and videos pop up. Here are some of my favorite videos showing the core features of the Rapberry Pi in action.
The most popular video type is probably the unboxing. This demonstrates the boot time and bit of lag that you'll expect when running X windows on the Raspberry Pi.
\n\nThe Raspberry Pi has headers to support daughterboards and should integrate well with arduino-based projects. Here are a few videos showing off the Pi interacting with other devices.
\nTouchscreen Anyone? Here is a 3.5 inch LCD touchscreen driven by the Raspberry Pi.
\n\nHere is a proof of concept that controls LEDS through a web server on the Raspberry Pi running apache and php.
\n\nAndroid transporter with Raspberry Pi ("beaming content")
\n\nDemonstration of consumer electronics control (CEC) on Raspbery Pi via HDMI.
\n\nHighly touted, the GPU is said to be capable of 1080p output via hdmi. These video some the Raspberry Pi in action playing HD video.
\nAn example of 1080p playback.
\n\nXBMC playing mp4 followed by H.264 with Android remote control
\n\n\nA small computer? Cool, I'll make it into a gaming rig. While this isn't my first idea for a great Raspberry Pi project, you can apparently run Quake 3 and Mame as shown below. I'm much more excited to see Java running on the Raspberry Pi.
\nMame on Raspberry Pi. You'll want to jump to 3:50 and you'll see some examples.
\n\nQuake 3 on Raspberry pi.
\n\nJavaFx running on the RPi
\n\nWhat are you doing with your RPi? Post a comment or use the contact form above. I'll try to update the post periodically.
\n" }, { "id": "https://osterman.co/glycosuria-without-hyperglycemia/", "url": "https://osterman.co/glycosuria-without-hyperglycemia/", "title": "glycosuria without hyperglycemia", "date_published": "2012-07-16T00:00:00.000Z", "summary": "A 2012 archival medical self-quiz on causes of glycosuria without hyperglycemia, originally an interactive question on the old WordPress site.", "content_html": "\n[slickquiz id=1]
![\"raspberry](\"/assets/img/wp/uploads/2012/07/raspberry-pie.jpg\")
| Target price: | \nUS$ 35 | \n
| SoC: | \nBroadcom BCM2835 (CPU, GPU, DSP, and SDRAM) | \n
| CPU: | \n700 MHz ARM1176JZF-S core (ARM11 family) | \n
| GPU: | \nBroadcom VideoCore IV,[57] OpenGL ES 2.0, 1080p30 h.264/MPEG-4 AVC high-profile decoder | \n
| Memory (SDRAM): | \n256 MB (shared with GPU) | \n
| USB 2.0 ports: | \n2 (via integrated USB hub) | \n
| Video outputs: | \nComposite RCA (PAL & NTSC), HDMI (rev 1.3 & 1.4),[58] raw LCD Panels via DSI | \n
| \n | 14 HDMI resolutions from 640×350 to 1920×1200 plus various PAL and NTSC standards. | \n
| Audio outputs: | \n3.5 mm jack, HDMI | \n
| Onboard storage: | \nSD / MMC / SDIO card slot | \n
| Onboard network: | \n10/100 Ethernet (RJ45) | \n
| Low-level peripherals: | \n8 × GPIO, UART, I²C bus, SPI bus with two chip selects, +3.3 V, +5 V, ground | \n
| Power ratings: | \n700 mA (3.5 W) | \n
| Power source: | \n5 volt via MicroUSB or GPIO header | \n
| Size: | \n85.60 × 53.98 mm (3.370 × 2.125 in) | \n
| Weight: | \n45 g (1.6 oz) | \n
The video originally embedded here — Hak5’s Maker Faire 2012 interview with the Raspberry Pi’s creator — is no longer available (the Revision3 player has shut down).
\npasswd\nwhen you first login. Also, you may need to\n
rm -i /etc/udev/rules.d/70-persistent-net.rules\nand reboot to get networking running. Once that's done, open the\n
README\nfile on the desktop and follow the directions. The only thing I couldn't get to essentially work out of the box was the networking. Here is a screenshot from my quick installation.\n
![\"Raspberry](\"/assets/img/wp/uploads/2012/07/RaspberryPi-rpi-screenshot-Running-Oracle-VM-VirtualBox_003.png\")
The video originally embedded here — Ben Heck’s cartridge-based Raspberry Pi computer build — is no longer available (the Revision3 player has shut down).
\n![\"gentoo](\"/assets/img/wp/uploads/2012/07/gentoolinux1.png\")
My internet \"wasn't working\" and after figuring out that I could ping my router but not outside, it was clear that I needed to reset my router so that it would be assigned another IP address via my dsl modem. This scenario is common. The only thing that may be a little different for me is that I don't have one of those iconic blue on gray boxes with two antennas sticking up. I was able to log into my router via ssh to r![\"dell](\"/assets/img/wp/uploads/2012/07/dellgx1a.jpg\")
estart my network. My router is a full computer. It is a dell gx1 optiplex 400mHz system running Gentoo Linux that I installed installed initially in 2005.\nI've updated my installation a few times over the years, but I've not had to add or remove any of the hardware. More impressively, I've never had to reinstall the operating system. Why is that odd? This is a computer that runs 24 hours a day, seven days a week, every week, without a reboot. The most common reason for turning off or rebooting the computer.
\n![\"mycancergenome.org\"](\"/assets/img/wp/uploads/2012/06/mycancergenome-logo.jpg\"){kind=logo}
\n\nAs we continue to move toward personalized and targeted therapies, tools like this will become the norm. The challenge is, of course, integrating these tools into our standard work flow.
\nThe next problem is interpreting the amount of molecular and genomic data to make those data applicable to patients. These, however, are good problems to have.
\n" }, { "id": "https://osterman.co/secure-personal-health-information/", "url": "https://osterman.co/secure-personal-health-information/", "title": "Secure Personal Health Information in 10 Steps", "date_published": "2012-06-14T00:00:00.000Z", "summary": "A 2012 checklist of ten security measures for systems storing personal health information, from SSL to two-factor auth and encryption.", "content_html": "![\"secure](\"/assets/img/wp/uploads/2012/06/secure-patient-health-informationt1.png\")
SSL AccessIf you purchase a certificate from a signing authority and a third party attempts a man in the middle attack on one of your clients, the client's web browser will warn the client that the secure certificate is not signed by a valid authority.
\nOften when someone says their system is "secure," this is the extent of their claim.
\n<meta http-equiv=\"cache-control\" content=\"max-age=0\" />\n<meta http-equiv=\"cache-control\" content=\"no-cache\" />\n<meta http-equiv=\"expires\" content=\"0\" />\n<meta http-equiv=\"expires\" content=\"Tue, 01 Jan 1980 1:00:00 GMT\" />\n<meta http-equiv=\"pragma\" content=\"no-cache\" />
\n![\"sql](\"/assets/img/wp/uploads/2012/06/sql-injection.png\")
An sql injection attack is an attack on a database where a third party submits malformed input into a form that is submitted to the database. The malformed input then either allowed unauthenticated users to gain access to the system or worse, to allow remote code execution and compromise access to the whole system.\nThere are three major downsides however.
\nThe video originally embedded here — George Sledge’s ASCO 2011 presidential address (a Brightcove Flash player) — is no longer available.
\nwe pay too much attention to what successful people are like, and too little attention to where they are from: that is, their culture, their family, their generation, and the idiosyncratic experiences of their upbringing.\n
![\"oncology](\"/assets/img/wp/uploads/2012/06/outliers.jpg\" "\\"oncology")
\nI am far from unique but I do have passion for both clinical oncology and information technology. My head spins with possibilities of to integrate these fields in the next decade. Whether I will be able to contribute to this exciting field is yet to be determined. Of one thing I am certain - what a great time to be alive.
\n" }, { "id": "https://osterman.co/staph-infection/", "url": "https://osterman.co/staph-infection/", "title": "Staph Infection?", "date_published": "2012-06-06T00:00:00.000Z", "summary": "A 2012 clinical note on adding clindamycin to suppress staph TSST-1 toxin when infections worsen despite broad-spectrum antibiotics.", "content_html": "In a brief literature review there were some reports of linezolid having similar effects. There are not good data to support its routine use to bind toxin however.
\n![\"Just_staphylococcus_aureus\"](\"http://microbewiki.kenyon.edu/images/3/3d/Just_staphylococcus_aureus.jpg\")
I've often asked by medical students what additional coverage is gained with using piperacillin-tazobactam (Zosyn) over ampicillin-sulbactam (Unasyn). The question comes up because so many people reflexively reach for "vanc-zosyn" for broad coverage. The answer is Pseudomonas.
\nThe more difficult question is what additional coverage is gained with ampicillin-sulbactam over piperacillin-tazobactam. The answer is acinetobacter.
\nThe additional coverage comes from the the sulbactam portion. This isn't to say that piperacillin-tazobactam should be a first line treatment, but it's interesting nonetheless.
\n![\"acinetobacter\"](\"http://phil.cdc.gov/PHIL_Images/10095/10095_lores.jpg\")
I have an office for the first time in many years with my new role as chief resident. Unfortunately, I am more familiar with Google and Mozilla's office-like tools than I am with those from Microsoft. One of the first frustrations I found was having a browser and email client window open to be able to read email requests for meetings and then put them in my Google calendar. Two days ago, I stumbled upon the Google calendar add-on for thunderbird. It is perfect. I'll include a screenshot but it does exactly what I would expect. Thanks go to Bryan Clark.
\n![\"thunderbird](\"/assets/img/wp/uploads/2012/06/travis-osterman-thunderbird-google-calendar.png\" "\\"thunderbird")
![Download (400px)](\"/assets/img/osterman-headshot-400.jpg\"){kind=link}