Every health system I talk to wants to know which model to buy. Almost none are asking the question that actually determines whether the model will work: is our data ready to feed it?
The current moment makes the model feel like the product. It isn't. In oncology, the model is the easy, commoditizing part. The decisive, under-resourced part is the structured data underneath it, and that's where leadership attention and budget should go first.
What I've watched fail
Over five years of building real-world prediction models for immune checkpoint inhibitor outcomes, the constraint was never the algorithm. It was whether the inputs the model needed, stage, biomarker status, line of therapy, toxicity, existed as structured, trustworthy fields rather than prose buried in notes. We deliberately built against routinely collected EHR data instead of a hand-curated research cohort, precisely because a model that needs a pristine cohort can't be deployed in the clinic it was built for.
The same lesson showed up when we ran one of the first peer-reviewed clinical evaluations of ChatGPT: a fluent answer is not a reliable one, and you cannot tell the difference without the structured ground truth to check it against.
Why standards are the unglamorous answer
This is the case for data standards, and it's why I chair the mCODE Executive Committee. mCODE, the minimal Common Oncology Data Elements, defines the core oncology data every patient's record should carry, in a form that travels between institutions. It's now implemented at more than 70 institutions across six countries and is the only method of submitting data to CMS's Enhancing Oncology Model. None of that is glamorous. All of it is what makes downstream AI possible.
The institutions that will get real value from clinical AI are the ones that invested early in the boring layer: structured capture at the point of care, genomics integrated into the record rather than stranded in PDFs, and a governance model for what "good data" means. At Vanderbilt Health, that investment is why we hold more structured genomic data in the EHR than any other institution in the United States, and why the AI conversation there starts from a different place.
What leaders should actually fund
If you run AI strategy for a health system, three priorities should precede your model selection:
- Structured capture at the source. Every field you want a model to use has to be captured as data, not narrative, by the people doing the work, which means workflow and incentives, not just schema.
- Standards adoption (mCODE or equivalent). Interoperable data is what lets you validate, benchmark, and eventually share, and what keeps you eligible for programs like the EOM.
- Governance before deployment. Decide how you'll measure a model's accuracy and its failure modes in your population before it touches a patient.
Buy the model last. By the time you do, the hard work, the part that determines whether it helps anyone, will already be done.