Clinical AI prediction claims often confuse early detection with recognition signals
Prediction Is Not Detection: Evaluating Pre-Recognition Claims in Longitudinal Clinical AI
Artificial Intelligence
Summary
Early detection of medical problems is hard and important, but some AI tools seem better at predicting problems because they rely on recognizing signals after doctors start to notice issues. The authors explain that using these signals can make AI seem more accurate and early than it really is. They propose ways to test AI predictions more fairly by separating real early signs from those that appear after recognition.
What this means in practice
- •For hospital data teams: Improve evaluation of AI tools by separating true early detection from recognition-dependent signals in patient records.
- •For healthcare ai developers: Develop clinical AI systems that better validate early prediction claims using independent standards before clinical recognition.
Authors
Jing Yang, Long R. Jiao, Xiujun Cai, Zongjiu Zhang
Abstract
Clinically useful early detection requires validated pre-recognition lead time. Yet event-based evaluations of longitudinal clinical AI can treat recognition-mediated care-process signals as shortcuts and recognition-dependent endpoints as reference standards, inflating apparent performance and lead time while undermining cross-center transport. Such results may serve prognosis without establishing detection before recognition. We define an interval-censored pre-recognition transition, an independent as-of reference standard, and a prespecified recognition proxy to make the claim testable.