Answers are grounded in this paper and cite their sources.
Background: Therapeutic drug monitoring (TDM) enables individualized pharmacotherapy for drugs with high interpatient variability and concentration-dependent effects. Conventional Bayesian approaches, although clinically established, may not fully capture pharmacokinetic complexity in dynamic or heterogeneous patient populations. Artificial intelligence (AI) and machine-learning methods offer complementary capabilities, including modeling nonlinear relationships and integrating high-dimensional clinical data. Methods: Peer-reviewed studies that applied AI or machine-learning methods to TDM between January 2017 and April 2026 were reviewed in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 guidelines. Studies involving adult and pediatric populations were included, whereas reviews, editorials, and animal studies were excluded. Results: Forty-seven studies covering a range of drugs, including vancomycin, tacrolimus, antiepileptics, mycophenolic acid, methotrexate, and voriconazole, were included. Tree-based ensemble methods, particularly eXtreme Gradient Boosting, random forest, and gradient boosting, were the most frequently applied algorithms. In general, these methods demonstrated strong predictive performance, with area under the receiver operating characteristic curve values up to 0.90 for classification tasks and coefficient of determination values up to 0.85 for regression tasks. In 13 studies, AI models demonstrated predictive performance comparable with or exceeding that of traditional population pharmacokinetic and Bayesian approaches for selected applications. However, these comparisons were heterogeneous and often based on retrospective data sets with limited external validation. Conclusions: AI-based methods have considerable potential for enhancing TDM, particularly for dose optimization, drug-exposure estimation, and toxicity prediction. However, evidence linking improved predictive performance to meaningful clinical outcomes remains limited.
No takes yet. Share an insight, caveat, or question.
Goncalves et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: