The authors built a transformer to estimate individualised treatment effects from records of about 350,000 All of Us participants. It learned diagnosis and medication histories without outcome labels, then fed doubly robust estimation of treatment choice and outcome risk. Tests compared GLP-1 receptor agonists with DPP-4 inhibitors on cardiovascular events in diabetes, and immunotherapy with chemotherapy on cancer mortality. Treatment-choice and outcome models discriminated and calibrated better than tabular methods and the same transformer without pre-training. Predicted effects varied substantially between patients, with patterns linked to diagnoses, medication histories and comorbidities. The estimates rest on observational records, so unmeasured confounding remains possible, and no external cohort was used for validation.
Why it is interesting: Most health-record foundation models predict risk rather than estimate treatment effects. This one targets which patients may gain more from one drug than another, using observational data.
Cite this item
BibTeX
Free sign-up. A digest of the new papers twice a week, grouped by domain.