Machine learning models for vancomycin dose estimation in hospitalized adults and adolescents

Background Machine learning (ML) methods may improve individualized vancomycin dosing by leveraging high dimensional electronic health records data. This review aimed to identify studies that developed, validated, or implemented ML models to recommend or predict intravenous vancomycin doses in adolescents and adults, and to appraise their methodological quality and clinical readiness. Methods We searched electronic sources for original human studies using ML for vancomycin dosing or pharmaco-kinetic prediction. Two reviewers independently screened records, extracted data, and assessed risk of bias using PROBAST tool. Due to heterogeneity in targets, models and metrics, narrative synthesis was carried out. Results Twelve retrospective studies met inclusion criteria. Algorithms included tree-based ensembles, deep neural nets and hybrid population pharmacokinetics plus ML. Across studies ML methods consistently reduced prediction error and increased dosing accuracy compared to conventional population pharmacokinetic/Bayesian tools. External validation was present in several but not all studies. No study evaluated clinical outcomes. Risk of bias was variable, with concerns about participant selection, calibration reporting and generalizability. Conclusions ML approaches improve vancomycin exposure and dose prediction compared to traditional methods. However, prospective, multicenter evaluation targeting guideline-recommended AUC outcomes and trials linking ML guided dosing to clinical outcomes are needed before routine clinical deployment.