Presentation (CTI)

OBJECTIVES: Adverse drug reactions (ADRs) are preventable cause of morbidity, prolonged hospitalization, and increased healthcare costs in Indian hospitals. This study aimed to develop and validate an explainable, AI-powered Clinical Pharmacist Assistant (AI-CPA) to predict patient-level ADR-risk using large-scale clinical data and to support proactive pharmacist-led interventions.
METHODS: A supervised ML model based on XGBoost was trained on approximately 20.3 million de-identified inpatient records derived from MIMIC-IV clinical data, the FAERS, and a synthetic Indian-hospital dataset. Rigorous preprocessing included drug-normalization, clinically informed imputation, and feature encoding. Data were split into training (70%), validation (15%), and a held-out test set (15%). AUC-ROC with 95% CIs, sensitivity, specificity, precision, F1 score, and accuracy were calculated for all models. Model interpretability was provided using SHAP, and the system was designed to be HL7 FHIR-compliant.
RESULTS: On the held-out test-set, the model demonstrated excellent discrimination with an AUC-ROC of 0.99923 (95% CI: 0.99915-0.99931; p<0.0001). Sensitivity was 99.3%, specificity 98.7%, precision 85.4%, F1 score 91.8%, and overall accuracy approximately 98.8%. Performance remained stable across training, validation, and test splits. Key-predictors of ADRs were AKI, certain medications and their modes of administration, serum-creatinine and urea nitrogen-levels, diabetes, and duration of hospital stay. SHAP analysis provided transparent, patient-specific insights into the driving factors behind ADR-risk.
CONCLUSIONS: The AI-CPA showed exceptional predictive performance and transparency in a large, independent test set. Interoperable design coupled with clinician-interpretable outputs could make this tool very useful for the ADR-risk assessment tasks of pharmacists to support proactive pharmacovigilance in Indian-hospitals. Its impact on workflow-efficiency and burden of preventable ADRs should be measured in a real-world, prospective study.