Machine Learning in Personalized Drug-Dose Optimization: A Review

Dosing decisions for narrow-therapeutic-index and highly variable drugs have historically relied on population-based nomograms, therapeutic drug monitoring, and clinician heuristics, which frequently under- or over-dose individual patients. Machine learning (ML) offers a data-driven alternative capable of integrating genomic, physiological, and longitudinal clinical information to individualize dosing. This review synthesizes recent literature on ML-based precision dosing across five therapeutic domains—anticoagulation, antibiotics, immunosuppression, oncology, and diabetes—and examines the major methodological families in use: classical supervised learning, deep learning, neural ordinary differential equations, and reinforcement learning. Across domains, ML models consistently outperform linear pharmacogenetic and population pharmacokinetic algorithms in retrospective accuracy, and reinforcement learning shows promise for sequential, adaptive dosing in closed-loop systems such as the artificial pancreas. However, clinical translation remains limited by small and non-diverse training cohorts, weak external validation, limited interpretability of complex models, and an evolving regulatory landscape for adaptive software a s a medical device. Emerging approaches—hybrid mechanistic-ML models, neural ODEs, explainable AI, and patient-specific digital twins—are discussed as pathways toward safer, more transparent, and clinically actionable dose-individualization tools. The review concludes that realizing the potential of ML in precision dosing will require prospective validation, standardized reporting, and close integration with pharmacometric principles rather than treating ML as a replacement for mechanistic understanding.