Domain shift

When the data a model meets in deployment differs systematically from the data it was trained on: a different scanner, a different acquisition protocol, a different patient mix, a different hospital. Performance drops, often sharply, and usually with nothing in the output to warn you. Domain shift is the reason external validation exists and the reason a model that works at one site cannot be assumed to work at the next. It is not the same as model drift, which is one site changing over time; domain shift is the gap between two sites or two populations at the same moment. Before buying, ask which scanners, protocols, and populations the model was validated on, and be honest about how far yours sit from them.

A nodule detection model validated on one vendor's CT gets installed in a department running two other scanners and a lower-dose protocol. Nothing errors. Detections just become less reliable, and the first person to notice is whoever audits, if anyone does. Ask which scanners, protocols and populations sit behind the published numbers, compare that list to your own estate honestly, and where they diverge, budget for a local validation before go-live rather than after the first miss.

Terms like this come up in real clinical scenarios across the HelloAI courses: bite-sized modules with verifiable certificates. An account takes one minute, no password needed.