Lyle Palmer

Deep learning promises to transform medicine, but the field is currently recapitulating — at speed — many of the methodological growing pains that epidemiology and genomics endured over recent decades. This talk draws out those lessons. The candidate-gene era of genomics produced thousands of published associations that failed to replicate; the field recovered only through large collaborative consortia, stringent statistical thresholds, mandatory replication, and a culture of data sharing — a ready-made template for addressing the reproducibility crisis now confronting medical AI. Epidemiology’s central preoccupations — confounding, selection bias, spectrum effects, and generalisability across populations — map directly onto the characteristic failure modes of clinical deep learning models, which routinely learn spurious shortcuts, encode demographic information, and degrade under dataset shift when moved between institutions, scanners, or patient populations.

Both disciplines also teach that association is not translation: as with polygenic risk scores, impressive retrospective performance offers no guarantee of clinical utility, and rigorous pre-deployment evaluation — including prospective ‘silent trials’ run alongside routine care — is needed before models are permitted to influence clinical decisions. Drawing on examples from the presenter’s work in genetic epidemiology, large-scale population cohorts, linked administrative data, and medical imaging AI, the talk argues that the fastest route to safe and trustworthy medical AI is not to relearn these lessons independently, but to import the epidemiological toolkit — careful study design, bias assessment, external validation, and a culture of replication — into everyday machine learning practice.