RELEAP: reinforcement-enhanced label-efficient active phenotyping for electronic health records

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Electronic health record (EHR) phenotyping often uses noisy proxy labels, impacting risk prediction reliability.
• Existing active learning methods do not optimize downstream prediction performance directly.

Purpose of the Study:

• To develop a framework that uses downstream prediction performance as feedback for phenotype correction and sample selection.
• To improve the reliability of EHR-based risk prediction under constrained labeling budgets.

Main Methods:

• Proposed reinforcement-enhanced label-efficient active phenotyping (RELEAP), a reinforcement learning-based active learning framework.
• RELEAP adaptively integrates querying strategies and updates its policy based on downstream model feedback.
• Evaluated RELEAP on a Duke University Health System cohort for incident lung cancer risk prediction using logistic regression and penalized Cox survival models.

Main Results:

• RELEAP improved over proxy-only baselines and approached oracle performance.
• Logistic AUC increased from 0.774 to 0.807; survival concordance index increased from 0.715 to 0.749.
• Performance gains were stable across iterations and consistent in sex-stratified analyses.

Conclusions:

• RELEAP links phenotype refinement to prediction outcomes, identifying samples that improve downstream discrimination and calibration.
• Offers a principled alternative to fixed active learning rules.
• Provides a scalable, label-efficient paradigm for EHR-based risk prediction, reducing manual review and enhancing reliability.