Drug and Dosage Recommendation Based on Explainable Generative AI Using Patient-Specific Modeling

Despite advances, personalized medicine involving patient-specific drug selection and dosing remains challenging. In this paper, we propose an interpretable, generative AI framework for drug recommendation and dosing. While clinical decisions often follow general guidelines, our approach uses patient-specific data to generate evidence-supported candidate drug-and-dose recommendations informed by up-to-date, provenance-linked sources. The Retrieval-Augmented Generation (RAG) framework combines data from electronic health records (EHRs) and employs machine learning and explainable AI methods. The digital twin approach supports this by simulating counterfactual treatment effects in a virtual patient model. Hence, the approach provides not only patient-tailored recommendations but also a clear, evidence-supported justification for the suggested drug and dose. We demonstrate feasibility via illustrations and simulations and outline record-based evaluations across multiple cohorts, reporting lower dose-prediction error, better calibration, and better discrimination of adverse drug reactions (AUC-ROC) and lower simulated ADR-risk indicators (RRI = 0.78; 95% CI 0.72–0.85) relative to guideline-based baselines, together with clinician-trust assessments. Retrospective results were consistent across heterogeneous cohorts, suggesting robustness to population and data shifts. The dataset comprises 75% real anonymized EHR records and 25% validated synthetic samples used only for training, with all test metrics computed on real-only data. We evaluate our approach on a de-identified, multi-cohort EHR-pharmacogenomics dataset of 20,030 patients spanning Hypertension, Diabetes Mellitus, Oncology, and Renal Impairment cohorts (hereafter, the ADR-20K multi-cohort dataset), constructed from anonymized EHRs, PharmGKB resources, and GAN-augmented synthetic samples with patient-level splits and temporal blocking to prevent leakage.