Work

Temporal shifts in pancreatitis reporting with GLP-1 receptor agonists: active-comparator disproportionality analyses and machine-learning report prioritization in FAERS

Year: 2026
Type: preprint
Abstract: Background Spontaneous reporting changes over time in volume, reporters, and drugs. Using pancreatitis, a labelled event, with glucagon-like peptide-1 receptor agonists (GLP-1 RAs) as a case study, we assessed how these changes affect disproportionality analyses and machine-learning report prioritization in the US Food and Drug Administration Adverse Event Reporting System (FAERS). Methods FAERS files (2015–2025) were deduplicated. Pancreatitis was defined by the narrow Acute pancreatitis Standardised MedDRA Query. Proportional reporting ratios, reporting odds ratios (RORs), information components, and empirical Bayes geometric means were calculated for seven GLP-1 RAs with standard thresholds, without causality assessment. Exploratory analyses used dipeptidyl peptidase-4 (DPP-4) and sodium-glucose cotransporter-2 (SGLT2) inhibitors as active comparators, crude and adjusted for calendar year, reporter type, age, sex, and region. As a methodological demonstration, machine-learning models were compared with simple baselines under temporal validation and annual retraining. Results Of 14,930,529 reports, 359,192 involved a GLP-1 RA, 7,321 with pancreatitis. All agents met all signal criteria against the full database (class ROR 8.00; 95% confidence interval 7.80–8.21). Against DPP-4 inhibitors, the crude class ROR was 0.99 (0.94–1.04), although GLP-1 RA reports included pancreatitis more often in every year. The year-adjusted (primary) ROR was 1.38 (1.31–1.45) and the fully adjusted ROR 1.85 (1.74–1.96), but among the minority of reports with a recorded diabetes indication it was 0.85 (0.80–0.91). Against SGLT2 inhibitors, the class ROR was 1.86 (1.76–1.97) crude and 2.12 (2.01–2.25) adjusted for year. Under temporal validation, gradient boosting outperformed the best baseline (area under the receiver operating characteristic curve 0.723 vs 0.627); a frozen model deteriorated (0.646 in 2025), whereas annually retrained gradient boosting remained at 0.73–0.80. Fields completed after case assessment inflated performance (0.938). Conclusions Changes in calendar time and reporter type influenced the crude comparison with DPP-4 inhibitors and eroded machine-learning prioritization, which annual retraining only partly mitigated. Pancreatitis was consistently reported more often with GLP-1 RAs than with SGLT2 inhibitors, whereas the comparison with DPP-4 inhibitors was inconclusive, and agent-level estimates varied with the comparator and model. These are reporting patterns for a labelled event, not estimates of risk or causation.
Institution: Lund University
Language: English
Subfield: Toxicology
Open Access status: green
OpenAlex ID: W7225894336
Cites: 36
Cited by:
Related to: 10