In 2019 my team at Vanderbilt University Medical Center and a group from GE HealthCare set out to predict who would benefit from immunotherapy and who would be harmed by it. Five years later we had a body of peer-reviewed work, filed international patents, and a model trained on routinely collected clinical data, not a data dump rotting on a shared drive. I want to explain why, because the difference was structural, not luck.
I have sat on both sides of these deals. I have watched well-funded academic-industry AI collaborations produce a press release, a single static data transfer, and then nothing: no deployable model, no usable IP, no publication anyone cites. The failure mode is so common that my colleagues and I eventually wrote it up and presented it at ASCO: Overcoming barriers in academic-industry partnerships to improve predictive modeling in immuno-oncology (ASCO Annual Meeting, 2022). The short version is that the standard contract structure is built to fail, and a few deliberate choices are what separate the collaborations that ship from the ones that stall.
The default structure is built to fail
The conventional academic-industry data deal looks tidy on paper. The academic medical center signs a contract, exports a de-identified dataset, and hands it to the company's data scientists. The two sides then work in isolation. This is the "data dump," and it fails in four predictable ways.
First, the data is static. Once it's transferred, it can't be refined, so the inevitable discrepancies (a lab value coded three different ways, a date that means admission in one feed and diagnosis in another) never get resolved. Second, the model picks up artifacts that nobody flags, because the people who know the clinic aren't in the room when the outputs are reviewed. Third, when someone finally wants to run the model on a real patient, they discover it depends on manually curated fields that don't exist in the live record, so it can't be deployed in the clinic it was built for. Fourth, and most corrosive over time: the clinicians aren't involved, so the academic partner gets nothing back except a check, and the relationship has no reason to continue.
Every one of those failures comes from the same root cause: the two sides act in isolation. Fix the isolation, and you fix most of the rest.
What we did instead
The GE HealthCare Digital Precision Oncology collaboration ran from 2019 to 2024, with me as principal investigator on the Vanderbilt Health side and Jan Wolber leading on GE's. The teams were genuinely distributed (clinicians and curators in the United States, modelers in Hungary and Germany), which made the discipline below non-optional rather than aspirational. Here is the recipe we actually used, the same one we laid out at ASCO in 2022:
- Iterate on the data, don't dump it. Our clinical and curation experts met with the industry modelers to dynamically refine the de-identified datasets, reconciling structured data with manually curated data and resolving discrepancies rather than training on top of them.
- Put clinicians in front of the model outputs. Practicing oncologists reviewed predictions specifically to catch artifacts and steer the final models. A radiologic or temporal artifact that looks like a signal to a modeler is often obvious to someone who has treated the disease.
- Automate the curation you'll need at deployment. Our curators worked with in-house machine-learning people to build algorithms that automatically extracted natural-language elements from the EHR, so the model didn't quietly depend on hand-curation that would vanish the moment it left the lab.
- Keep everyone in the same recurring meeting. Clinical and industry stakeholders met regularly with the modelers to keep the work pointed at something clinically and commercially useful, not just statistically tidy.
None of that is exotic. It is mostly a refusal to let the two organizations drift into separate workstreams that only meet at contract milestones.
The deployability constraint that made it real
The single most important decision we made was about the data the model was allowed to consume. We built against routinely collected EHR data, the stuff that already exists for every patient as a byproduct of care, rather than a bespoke research cohort with hand-annotated features. That choice is annoying. Routine data is messier, and you give up some apparent accuracy. But it is the only choice that produces a model you can actually run in clinic, because a model that needs a pristine research cohort can only ever be demonstrated, never deployed.
The flagship result, Prediction of Effectiveness and Toxicities of Immune Checkpoint Inhibitors Using Real-World Patient Data (Lippenszky et al., JCO Clinical Cancer Informatics, 2024), showed it works. Using routine EHR data from more than 2,200 patients treated with immune checkpoint inhibitors, we built models that predicted pneumonitis, hepatitis, colitis, and one-year overall survival, with AUCs in the mid-0.70s (roughly 0.73 to 0.76). Crucially, the models required no additional data collection or documentation in the clinic. That last clause is the whole point. The model fits the workflow that already exists.
It was also deliberately pan-cancer. The cohort spanned melanoma, lung, genitourinary, and other tumor types rather than chasing a single high-prevalence disease, because immunotoxicity is a cross-cutting problem and a model that only works in one cancer is a narrower asset for both partners.
Why the IP and the publications both happened
The structure produced two kinds of durable output at once, and that is not a coincidence. Because clinicians were embedded in the modeling loop, the work was publishable: multiple peer-reviewed manuscripts came out of it, not just an internal deck. And because the curation and model-generation methods were built deliberately rather than improvised, they were patentable. The resulting U.S. patent on the model-generation framework names the inventors from both institutions together. Academic credit and commercial IP are usually treated as a zero-sum trade. They aren't, if the work is structured so that the same activity generates both.
That joint inventorship is also the answer to the fourth failure mode above. The academic partner got real intellectual output, the industry partner got deployable methods and IP, and both had a reason to keep going. The collaboration ended on schedule in 2024, but the line of research did not stop with it. The methods and the questions it raised continue to shape how we bring predictive analytics into routine oncology decisions. Partnerships that ship tend to leave that kind of residue.
What to actually require if you're funding one of these
If you run strategy at a health system or at a company doing clinical AI, do not sign the data-transfer-only version of this deal. Write the structure into the agreement instead. Require recurring joint working sessions, not milestone hand-offs. Require that clinicians review model outputs as a deliverable, not a courtesy. Require that the model be built against data that exists in the live record, and treat any dependence on hand-curated features as a deployment defect to be fixed before launch. And settle the publication and IP terms up front so that academic credit and commercial ownership are designed to coexist rather than compete.
The full Digital Precision Oncology collaboration is documented as a case study, and I've written more elsewhere about why the data layer underneath these models matters more than the models themselves. But the partnership lesson stands on its own. The collaborations that produce deployable models and real IP are not the ones with the biggest budgets or the cleanest data. They are the ones where the two sides refuse to work in isolation, and build that refusal into the contract.