Glossary
Dataset Shift
Also called: distribution shift, domain shift, covariate shift
When the data a model sees in deployment differs statistically from its training data, degrading performance without any visible error.
Dataset shift occurs when a model meets a different scanner, a different acquisition protocol, a different patient population, or a different clinical era than it learned from. It is the dominant cause of real-world performance loss in medical AI.
Where This Gets Misread
The defining feature is silence. There is no error message, no flag, and no visible artifact. The model keeps producing confident outputs that are quietly worse, and nobody finds out unless someone is monitoring performance against outcomes. Most deployments are not.