Study Finds AI Models to Predict Schizophrenia Treatment Outcomes Don't Generalize Well
-
AI algorithms used to predict patient outcomes for schizophrenia treatments lack generalizability - they only worked well for the specific clinical trial they were developed for.
-
The study evaluated an AI model's performance on its initial training data and on independent clinical trials data from over 1,500 patients.
-
Contributing factors to the lack of generalizability may include differences in patient populations, insufficient or restricted data, and the context-dependency of patient outcomes.
-
The researchers recommend skepticism when evaluating predictive model findings that lack independent sample validation.
-
The fragility of predictive models suggests excellent performance in one clinical context is not a strong indicator of future performance.