External Validation & Domain Shift

Random train/test splits from the same hospital are optimistic by construction: same scanners, same coding habits, same patient mix — and subtle leakage survives (near-duplicate images across splits, the same patient on both sides of the split).

External validation — a site your model never saw — is where the honest number lives. Temporal validation (train on 2019, test on 2021) is the cheaper intermediate: it at least measures drift over time, which no random split does.