Psychiatric Safety Signals of GLP-1 Receptor Agonists: A FAERS-Based Pharmacovigilance Study with Explainable Machine Learning

Background: As glucagon-like peptide-1 (GLP-1) receptor agonist use expands globally, reports of psychiatric adverse events (AEs) have increased in spontaneous reporting databases. However, which case-level characteristics are associated with these reporting patterns remains insufficiently characterized. This study aimed to characterize case-level features associated with psychiatric AE reporting in GLP-1 receptor agonist users through pharmacovigilance and explainable machine learning methods. Methods: The FDA Adverse Event Reporting System (FAERS) data (2021 Q2–2025 Q3) were analyzed using a comparator-based approach comprising other antidiabetic and anti-obesity agents. Disproportionality analyses (PRR, ROR, and IC) were performed to detect consolidated safety signals at the Preferred Term (PT) level, with additional drug-specific analyses for individual GLP-1 receptor agonists. Three classification models (logistic regression, XGBoost, and LightGBM) were developed, and SHAP values were used to identify case-level features associated with psychiatric AE reporting patterns. Results: A total of 211,195 unique cases were included (111,105 for GLP-1 receptor agonists; 100,090 for comparators). Sixteen PTs met consolidated signal criteria, with suicidal ideation being the most frequently reported (ROR 2.95). Drug-specific analyses indicated that signal magnitudes were consistently higher for semaglutide than tirzepatide. The XGBoost model achieved an AUROC of 0.816. SHAP analysis showed that age ≥65 years had the highest mean |SHAP| value (0.57) with a negative direction, corresponding to a lower predicted probability of psychiatric AE reporting in older adults. Semaglutide use ranked second (0.35) and showed a positive direction. Absence of concomitant medications (0.20) and diabetes indication (0.10) showed negative directions, while younger age (19–44 years, 0.08) showed a positive direction. These patterns remained consistent in sensitivity analysis excluding concomitant psychotropic medication users (AUROC 0.797). Conclusions: Older age status was associated with decreased predicted probability of psychiatric AE reporting, while semaglutide use and younger age showed positive contributions. These case-level patterns, identified through pharmacovigilance analysis using a comparator-based approach and explainable machine learning, suggest that reporting patterns may differ across subgroups and that the observed reporting heterogeneity among younger adults and semaglutide users merits further investigation in prospective studies.