Presentation (CTI)

OBJECTIVES: To develop and validate an explainable AI-powered Clinical Pharmacist Assistant (AI-CPA) for predicting adverse drug reactions (ADRs) and supporting proactive pharmacist-led interventions in Indian hospitals.
METHODS: A retrospective dual-cohort validation study used approximately 20.3 million de-identified inpatient records from MIMIC-IV, FAERS, and a synthetic Indian hospital dataset. Data preprocessing included drug standardization, clinically guided imputation, feature engineering, and categorical encoding. An XGBoost model was trained using a 70/15/15 train-validation-test split. Performance was evaluated using AUC-ROC, sensitivity, specificity, precision, F1-score, and accuracy. Model interpretability was assessed using SHAP. External validation was performed using an independent dataset of 45 cases.
RESULTS: Internal validation demonstrated excellent discrimination (AUC-ROC 0.99923; 95% CI 0.99915-0.99931), with sensitivity of 99.3%, specificity of 98.7%, precision of 85.4%, F1-score of 0.918, and accuracy of approximately 98.8%. Important predictors included acute kidney injury, drug administration route, serum creatinine, blood urea nitrogen, diabetes, and length of hospital stay. External validation yielded an AUC-ROC of 0.8425, sensitivity of 85.1%, specificity of 82.4%, precision of 75.3%, F1-score of 0.799, and accuracy of 83.5%, indicating acceptable generalizability.
CONCLUSIONS: The AI-CPA demonstrated excellent internal performance and clinically acceptable external validation, with transparent prediction through SHAP. This approach has potential to strengthen proactive pharmacovigilance, improve clinical pharmacist workflow, and support safer medication management in Indian hospitals. Prospective multicenter studies are warranted to confirm real-world clinical impact.