Can artificial intelligence improve precision dosing? a systematic review of narrow therapeutic index drugs

Background Precision dosing of narrow therapeutic index drugs remains challenging because of substantial interindividual variability and the narrow margin between therapeutic efficacy and toxicity. Artificial intelligence (AI)–based approaches may improve individualized dosing by integrating complex clinical and pharmacogenomic data. This systematic review evaluated the predictive performance and applications of AI-based approaches for precision dosing in narrow therapeutic index drugs. Methods A systematic review was conducted according to PRISMA guidelines. PubMed, Scopus, Cochrane Library, and IEEE Xplore were searched from inception to December 2025 for studies evaluating AI-based approaches for precision dosing of narrow therapeutic index drugs. Eligible studies included clinical investigations assessing predictive performance or clinical outcomes of AI-guided dosing models. Methodological quality and reporting completeness were evaluated using the PROBAST and TRIPOD tools, respectively. Results Sixteen studies published between 2018 and 2025 were included, with most focusing on warfarin ( n = 8) and tacrolimus ( n = 5). Ensemble learning approaches, including random forest, XGBoost, extra trees regressor, heuristic stacking, and deep forest models, frequently demonstrated favorable predictive performance compared with conventional methods. Pharmacogenomic variables, particularly CYP2C9 and VKORC1 for warfarin and CYP3A5 for tacrolimus, were consistently identified as important predictors. The identified AI applications were primarily related to individualized dose prediction and therapeutic drug monitoring across anticoagulant, immunosuppressant, and antiepileptic therapies. However, substantial heterogeneity was observed in AI methodologies, target outcomes, validation strategies, and reported performance metrics. Most studies were retrospective and single-centre, while external validation and assessment of clinically meaningful outcomes remained limited. Conclusions AI-based approaches show promising potential for improving precision dosing of narrow therapeutic index drugs, particularly when integrating pharmacogenomic and clinical variables. Nevertheless, current evidence is limited by methodological heterogeneity, inconsistent reporting, and insufficient external validation. Further prospective multicentre studies evaluating real-world clinical outcomes are needed before widespread clinical implementation can be recommended.