MACHINE LEARNING MODELS FOR PREDICTING SERIOUS ADVERSE EVENT REPORTS IN GLP-1 RECEPTOR AGONISTS USING FAERS, 2015-2025
Author(s)
Zhouzhou Chu, MSc1, Daniel O. Umoru, BPharm1, Ang Li, PhD2, Ariel Tran, PharmD1, Bowei Tian, BSc2, Sherry Yun WANG, BSc, MPhil, PhD1.
1School of Pharmacy, Chapman University, Irvine, CA, USA, 2Department of Electrical and Computer Engineering, University of Maryland, College Park, MD, USA.
1School of Pharmacy, Chapman University, Irvine, CA, USA, 2Department of Electrical and Computer Engineering, University of Maryland, College Park, MD, USA.
OBJECTIVES: Glucagon-like peptide-1 receptor agonists (GLP-1RAs) generate substantial post-marketing safety reports in the U.S. FDA Adverse Event Reporting System (FAERS). This study evaluates machine learning models to predict serious outcomes and identify determinants of serious reporting across branded and generic GLP-1RA products (2015-2025).
METHODS: We processed 2015-2025 FAERS data using DuckDB and RxNorm mapping to normalize GLP-1RA exposures (e.g., semaglutide, tirzepatide, liraglutide). Cases were labeled serious if any outcome code (DE, LT, HO, DS, CA, RI, OT) was present. Features included demographics, reporter type, indication bucket, molecule identity/active ingredient, brand/generic/mixed drug classification, and adverse event categories (e.g., gastrointestinal, pancreatic, psychiatric). Logistic regression (for interpretability) and random forest models (to capture nonlinearity) were trained using a group-aware holdout split by primaryid (25% test). Within the training set, a separate group-based validation set was used for probability calibration (Platt scaling and isotonic regression) and threshold tuning. Performance on the held-out test set was assessed using ROC-AUC, PR-AUC, and Brier score.
RESULTS: Among 479,921 GLP-1RA-associated FAERS cases, both models achieved strong discrimination (logistic regression ROC-AUC = 0.876; random forest ROC-AUC = 0.916). Logistic regression identified pancreatic, psychiatric, and renal/urinary disorders as contributors to higher predicted seriousness. Meanwhile, type 2 diabetes, injection-site reactions and dosing/administration issues were associated with lower predicted seriousness. Molecule identity contributed modestly relative to reporting context and adverse-event phenotype. Random forest SHAP analysis emphasized dosing/administration issues, injection-site reactions, pancreatic effects, reporter type, and temporal variables as the most influential predictors of seriousness.
CONCLUSIONS: Machine learning models effectively distinguished serious from non-serious GLP-1RA reports in FAERS. Across models, adverse-event phenotype and reporting context exerted stronger impact on seriousness classification than molecule identity alone. These findings demonstrate how routinely collected pharmacovigilance data can support real-world safety monitoring and risk stratification for high-utilization metabolic therapies in HEOR practice.
METHODS: We processed 2015-2025 FAERS data using DuckDB and RxNorm mapping to normalize GLP-1RA exposures (e.g., semaglutide, tirzepatide, liraglutide). Cases were labeled serious if any outcome code (DE, LT, HO, DS, CA, RI, OT) was present. Features included demographics, reporter type, indication bucket, molecule identity/active ingredient, brand/generic/mixed drug classification, and adverse event categories (e.g., gastrointestinal, pancreatic, psychiatric). Logistic regression (for interpretability) and random forest models (to capture nonlinearity) were trained using a group-aware holdout split by primaryid (25% test). Within the training set, a separate group-based validation set was used for probability calibration (Platt scaling and isotonic regression) and threshold tuning. Performance on the held-out test set was assessed using ROC-AUC, PR-AUC, and Brier score.
RESULTS: Among 479,921 GLP-1RA-associated FAERS cases, both models achieved strong discrimination (logistic regression ROC-AUC = 0.876; random forest ROC-AUC = 0.916). Logistic regression identified pancreatic, psychiatric, and renal/urinary disorders as contributors to higher predicted seriousness. Meanwhile, type 2 diabetes, injection-site reactions and dosing/administration issues were associated with lower predicted seriousness. Molecule identity contributed modestly relative to reporting context and adverse-event phenotype. Random forest SHAP analysis emphasized dosing/administration issues, injection-site reactions, pancreatic effects, reporter type, and temporal variables as the most influential predictors of seriousness.
CONCLUSIONS: Machine learning models effectively distinguished serious from non-serious GLP-1RA reports in FAERS. Across models, adverse-event phenotype and reporting context exerted stronger impact on seriousness classification than molecule identity alone. These findings demonstrate how routinely collected pharmacovigilance data can support real-world safety monitoring and risk stratification for high-utilization metabolic therapies in HEOR practice.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
RWD148
Topic
Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, SDC: Diabetes/Endocrine/Metabolic Disorders (including obesity)