AI-Based Prediction of Drug–Patient Response for Personalized Diabetes Pharmacotherapy

Abstract

Type 2 diabetes mellitus is a heterogeneous metabolic disorder in which patients exhibit substantial inter-individual variation in therapeutic response. Although multiple pharmacological classes are available, treatment selection is frequently guided by population-level evidence and sequential therapeutic adjustment, which may result in delayed optimization of glycemic control. Artificial intelligence (AI) and machine-learning techniques provide an opportunity to integrate demographic, anthropometric, clinical, laboratory, and treatment-related characteristics to estimate patient-specific drug response. The present computational proof-of-concept study developed and compared machine-learning models for predicting glycemic response to commonly used antidiabetic pharmacotherapies. A structured synthetic patient dataset representing clinically relevant characteristics was generated for model-development demonstration. Predictor variables included age, sex, body mass index, diabetes duration, baseline glycated hemoglobin (HbA1c), fasting plasma glucose, renal function, previous antidiabetic therapy, treatment class, and selected comorbidities. Logistic regression, random forest, support vector machine, XGBoost, and artificial neural network models were evaluated using stratified training/testing procedures and cross-validation. The primary outcome was prediction of a clinically meaningful glycemic response, defined as a reduction in HbA1c of at least 1.0 percentage point during follow-up. Among the illustrative models, XGBoost demonstrated the strongest overall discrimination, followed by random forest and artificial neural network models. Baseline HbA1c, diabetes duration, fasting plasma glucose, treatment class, body mass index, and renal function were among the most influential predictors. SHAP-based interpretation demonstrated that the direction and magnitude of individual feature contributions varied between patient profiles. The findings demonstrate the methodological feasibility of an explainable AI framework for patient-specific prediction of antidiabetic treatment response. However, the results are illustrative and cannot establish clinical validity. External validation using large, diverse real-world cohorts and prospective clinical evaluation are required before clinical implementation.