Reinforcement Learning for Drug Dosage Adjustment

Abstract

Drug dosage adjustment is a sequential clinical decision problem requiring repeated evaluation of therapeutic response, toxicity, and changing patient physiology. Conventional dosing protocols and supervised prediction models often recommend isolated doses without  adequately representing the delayed and cumulative effects of previous administrations.  This study proposes a pharmacokinetically informed, uncertainty-aware offline reinforcement-learning framework for individualized dosage adjustment using longitudinal electronic health records.  The identified research gap concerns the limited  integration of prolonged drug-action memory,  conservative policy learning, uncertainty estimation, and explicit safety constraints within a unified dosing architecture. The proposed framework represents residual drug exposure through an action-history state encoder and restricts recommendations to clinically plausible dose ranges supported by observational data.  A distributional critic estimates the range of possible treatment outcomes, while a safety layer penalizes  toxicity, abrupt dose changes, and departures from established therapeutic boundaries.  The intended experimental design evaluates therapeutic target attainment, toxicity exposure, policy reliability, dose stability, and performance across clinically relevant patient subgroups.  The framework is designed as an interpretable decision support system rather than an autonomous prescribing mechanism, enabling safer translation of reinforcement learning into precision pharmacotherapy