Calibrated AI approach to pharmacovigilance using FAERS

Abstract We developed a transparent, calibrated AI framework using 242,312 FAERS reports on GLP-1 receptor agonists to predict adverse event seriousness. All models achieved strong discrimination, with area under the receiver operating characteristic curve (ROC-AUC)ranging from 0.877 to 0.930, and post hoc calibration improved probability reliability. Frequently reported gastrointestinal events showed low predictive importance, whereas dosing/administration, injection-site, and pancreatic events were among the strongest predictors, highlighting a disconnect between reporting frequency and predictive importance for reporter-classified seriousness.