Reinforcement learning alerts to Alzheimer's 20.9 months early

In clinical electronic health record cohorts, researchers developed a model-free reinforcement learning framework to optimize the timing of early alerts for chronic conditions. The framework addresses the trade-off between alert earliness and diagnostic specificity by formulating surveillance as a Partially Observable Markov Decision Process with asymmetric reward. To handle right-censored patient data, the method uses pseudo-label imputation, alongside entropy regularization to enable threshold adjustments without model retraining. In synthetic data, the advantage of look-ahead planning peaked when diagnostic information emerged in predictable bursts. Validated on real-world clinical records, the policy achieved an actionable lead time of 20.9 months prior to Alzheimer's disease diagnosis at 90% specificity. In a pediatric cohort, the framework also achieved an 8.7-month lead time prior to autism diagnosis at 90% specificity.