Safety Signals of GLP‐1 Receptor Agonists: A Multi‐Method Pharmacovigilance Analysis of FAERS (2018–2025) With Sensitivity‐Stratified Prioritisation, Notoriety‐Bias Assessment, and Cross‐Database Validation

ABSTRACT

Background

GLP‐1 receptor agonists (GLP‐1 RAs) are widely prescribed for type 2 diabetes and obesity, but their post‐marketing safety profile—particularly psychiatric, vascular, neoplastic and rare ophthalmic events—remains incompletely characterised and is vulnerable to confounding by indication, secondary‐suspect attribution and notoriety bias.

Methods

We analysed 32 quarterly FAERS files (2018Q1–2025Q4): 288408 unique GLP‐1 RA reports versus 11 319 150 non‐GLP‐1 RA reports. Disproportionality used four methods (reporting odds ratio [ROR]; proportional reporting ratio [PRR]; Bayesian Confidence Propagation Neural Network information component, lower 95% bound [BCPNN IC025]; Multi‐item Gamma Poisson Shrinker empirical Bayes lower 5th‐percentile bound [MGPS EB05]; signal = ≥ 3 of 4). The top 30 ROR‐ranked signals were stratified by primary‐suspect (PS) sensitivity. Triangulation included active‐comparator, intra‐class psychiatric comparison, interrupted time‐series assessment of reporting around the 2023 EMA announcement, Weber‐effect evaluation with Cochran–Armitage trend testing, and external validation in JADER.

Results

Of 22 059 PT–class pairs, 544 met the ≥ 3‐of‐4 criterion. PS‐stratification retained 21 of the top 30 signals and demoted 9 as secondary‐suspect‐driven (e.g., capillaritis ROR 18.6 → 0.27; heart sounds 84.3 → 0.55). PS‐stable headline signals were dominated by gastrointestinal/hepatobiliary effects: cyclic vomiting syndrome (ROR 64.8; semaglutide‐dominant), gallbladder injury (84.2), impaired gastric emptying (45.1), plus liraglutide‐concentrated medullary thyroid cancer (41.1) and semaglutide‐concentrated optic ischaemic neuropathy (26.5). Only semaglutide carried a class‐level suicidal‐ideation signal (ROR 2.30). Interrupted time series detected no level (p = 0.39) or slope change (p = 0.75) at the 2023 EMA announcement; psychiatric time‐to‐onset (median 47 days) aligned with early GI reactions. Sixteen of 17 priority signals (94.1%) replicated in JADER. As disproportionality analyses, these are hypothesis‐generating safety signals and cannot establish incidence, relative risk, or causality.

1. Introduction

Glucagon‐like peptide‐1 receptor agonists (GLP‐1 RAs) have transformed the management of type 2 diabetes mellitus (T2DM) and obesity over the past decade. Following liraglutide's approval in 2010, the class now spans dulaglutide (2014), semaglutide (2017, with subcutaneous, oral and high‐dose obesity formulations) and the dual GIP/GLP‐1 receptor agonist tirzepatide (2022). Acting on receptors expressed in the pancreas, gastrointestinal tract, cardiovascular system and central nervous system, these agents produce glucose‐lowering, weight‐reducing and cardioprotective effects, and global prescriptions have expanded from a niche T2DM intervention to a class accounting for tens of millions of obesity prescriptions annually [1].

Despite robust efficacy and cardiovascular outcome data from LEADER [2], SUSTAIN‐6 [3], REWIND [4], SURPASS [5], SURMOUNT [6], SELECT [7] and STEP [8], the real‐world safety profile of GLP‐1 RAs—particularly for rare events, long‐latency outcomes, and signals emerging in populations under‐represented in trials—remains incompletely characterised. Spontaneous reporting systems such as the FDA Adverse Event Reporting System (FAERS) provide a complement to randomised data by enabling signal detection across millions of real‐world exposures. Disproportionality analysis (DPA) is the cornerstone methodology [9], but is susceptible to confounding by indication, channelling bias, selective reporting, and notoriety bias—the artificial inflation of reporting following regulatory or media announcements. A less‐discussed but practically important confound is secondary‐suspect (SS) attribution: when a drug appears in case reports as a concomitant medication rather than the primary suspect, disproportionality may be driven by the co‐administered drug, not by the drug of interest.

Recent regulatory attention has focused on potential psychiatric adverse events. EMA initiated a review in July 2023 [10]; PRAC later concluded in April 2024 that available evidence did not support a causal association. GLP‐1 receptors are expressed in limbic regions including the hypothalamus, hippocampus, and ventral tegmental area, providing a plausible neurobiological substrate for centrally mediated psychiatric effects. Yet whether such signals are class‐wide or drug‐specific, and whether they represent drug‐specific reporting patterns or post‐EMA reporting artefacts, remains unresolved. The introduction of tirzepatide adds further complexity: its dual GIP/GLP‐1 mechanism may produce additive or synergistic central effects, and its rapid market expansion (now the largest GLP‐1 RA share in FAERS by quarterly report volume) makes its safety profile a regulatory priority.

This study addresses three interrelated gaps. First, the absence of a systematic, multi‐method signal prioritisation framework spanning the entire 2018–2025 post‐marketing period, with explicit handling of secondary‐suspect attribution. Second, the unresolved question of whether psychiatric signals reflect drug‐specific reporting patterns or notoriety bias. Third, the lack of quantitative, network‐based characterisation of the GLP‐1 RA multi‐system safety fingerprint. Our pre‐specified objectives were: (i) to identify robust adverse‐event signals for GLP‐1 RAs as a class; (ii) to triangulate signals through primary‐suspect sensitivity, three‐group active comparator analysis, intra‐class comparison, and external database validation; (iii) to characterise drug‐specific profiles—especially psychiatric and rare neoplastic/ophthalmic signals; (iv) to assess temporal dynamics including Weber effects with formal trend testing; and (v) to test notoriety bias formally with interrupted time series regression.

2. Methods

2.1. Data Sources and Study Period

We extracted data from 32 quarterly FAERS ASCII files spanning 2018Q1–2025Q4. After applying the FDA‐recommended caseid/caseversion deduplication algorithm and restricting each case to its most recent version, the analytic dataset comprised 288 408 unique deduplicated reports mentioning at least one GLP‐1 RA. The reference population was the 11 319 150 non‐GLP‐1 RA reports in the same period, yielding 11 607 558 total deduplicated FAERS reports overall. For drug‐specific analyses, each drug had its own contingency table; cases mentioning multiple GLP‐1 RAs (e.g., a patient switched from liraglutide to semaglutide) were counted in each drug‐specific stratum (Figure S1), so the sum of drug‐specific exposure totals (semaglutide 78 971 + liraglutide 20 467 + dulaglutide 72 848 + tirzepatide 121 069 = 293 355) exceeds the 288 408 unique‐case count. For external validation, we used the Japanese Adverse Drug Event Report (JADER) database (1 027 194 deduplicated reports through April 2026). All adverse events were coded using MedDRA v26.0 Preferred Terms (PTs).

2.2. Drug Identification

Four GLP‐1 RA agents were investigated: semaglutide (Ozempic, Wegovy, Rybelsus), liraglutide (Victoza, Saxenda), dulaglutide (Trulicity), and tirzepatide (Mounjaro, Zepbound). Identification used case‐insensitive matching of both drugname and prod_ai fields against a pre‐specified synonym dictionary of international non‐proprietary names and brand names. Active comparator drugs were SGLT2 inhibitors (empagliflozin, dapagliflozin, canagliflozin, ertugliflozin) and DPP‐4 inhibitors (sitagliptin, saxagliptin, alogliptin, linagliptin), identified by the same procedure. SGLT2 inhibitors and DPP‐4 inhibitors were chosen as active comparators because both share the type 2 diabetes indication with GLP‐1 RAs but act through distinct mechanisms; a signal that remains robust against two mechanistically distinct, indication‐matched comparators is less likely to reflect confounding by indication alone.

2.3. Disproportionality Analysis (3‐of‐4 Composite Criterion)

For every drug–PT pair we constructed a 2 × 2 contingency table [11] and computed: (1) Reporting Odds Ratio (ROR) with 95% CI; threshold ROR_low > 1 [12]. (2) Proportional Reporting Ratio (PRR) with χ ² statistic; threshold PRR ≥ 2 and χ ² ≥ 4 with cell a ≥ 3 [13]. (3) Bayesian Confidence Propagation Neural Network Information Component IC025; threshold IC025 > 0 [14]. (4) Multi‐item Gamma Poisson Shrinker Empirical Bayes EB05; threshold EB05 ≥ 2 [15, 16]. A signal was considered robust when at least three of four criteria were met [17]. PTs with fewer than three co‐reported cases in the drug‐of‐interest stratum were excluded. The 3‐of‐4 criterion is more conservative than any single method and substantially reduces false‐positive discovery.

2.4. Primary‐Suspect (PS) Sensitivity Stratification (Methodological Pillar)

To distinguish signals primarily attributable to GLP‐1 RAs from those reflecting concomitant‐drug (secondary‐suspect) attribution, we re‐computed disproportionality on the PS‐only subset, in which only reports listing a GLP‐1 RA as the primary‐suspect drug (drug_role = “PS”) were retained. For each top‐30 ROR‐ranked signal we then classified outcomes into three categories: PS‐stable (significance and approximate magnitude preserved—Δlog ROR within 30%); PS‐unstable (significance lost in PS‐only despite Main_sig = TRUE); and PS‐unstable signals were demoted from the headline analysis and reassigned to the Supporting Informations S1 with explicit annotation that the disproportionality was driven by reports in which the GLP‐1 RA appeared as a secondary‐suspect drug. PS‐only sensitivity stratification was applied to the top 30 ROR‐ranked signals, which captured all candidates expected to feature in regulatory priority lists; lower‐ranked signals (ranks 31–544) were not individually retested under PS‐only restriction. This sensitivity stratification mitigates the SS‐attribution confounder, which is a known but rarely operationalised limitation of FAERS DPA. The top‐30 ROR band was chosen because it both captures the extreme‐magnitude signals most likely to feature on a regulatory priority list and represents the region in which secondary‐suspect attribution artefacts are most concentrated, since such artefacts tend to produce extreme reporting odds ratios; in addition, every signal that entered clinical prioritisation was re‐tested under PS‐only restriction irrespective of its ROR rank, including three lower‐ranked signals (Cholelithiasis, Pancreatitis acute and the semaglutide‐specific Suicidal ideation signal).

2.5. Three‐Group Active Comparator Analysis

To control for indication and channelling bias, two active comparator groups were employed: SGLT2 inhibitors and DPP‐4 inhibitors, both sharing the T2DM indication but with mechanistically distinct pharmacology. For each PT, ROR was calculated separately in three analyses: GLP‐1 RA versus all non‐GLP‐1 (main); GLP‐1 RA versus SGLT2i; GLP‐1 RA versus DPP‐4i. A signal was considered comparator‐robust if ROR_low > 1 in all three analyses.

2.6. Drug‐Specific and Intra‐Class Psychiatric Comparison

Individual drug‐level disproportionality analyses were conducted for each of the four agents against the full non‐GLP‐1 RA reference. For psychiatric adverse events specifically, an intra‐class comparison was performed: each drug was evaluated against the pooled reports of the other three GLP‐1 RAs as the reference population. This isolates within‐class drug‐specific psychiatric profiles after removing the class‐shared indication, channelling and expectation pressures that affect all GLP‐1 RAs uniformly. Signal threshold: ROR_low > 1 with cell a ≥ 3.

2.7. Interrupted Time Series Analysis (Notoriety‐Bias Test)

To test directly for notoriety bias surrounding the 2023Q3 EMA psychiatric review announcement, we fitted a segmented regression interrupted time series (ITS) model [18] to quarterly psychiatric AE reporting proportions:

$$Y_t={\beta }_0+{\beta }_1·{\mathit{\text{Time}}}_t+{\beta }_2·{\mathit{EMA}}_t+{\beta }_3·\left({\mathit{\text{Time}}}_t\times {\mathit{EMA}}_t\right)+{\epsilon }_t$$

where Y _t is the psychiatric AE proportion in quarter t (psychiatric AE count divided by total GLP‐1 RA reports per quarter); Time _t is sequential quarter number; and EMA _t is a binary indicator (0 = pre‐announcement, 1 = post‐announcement). The pre‐announcement period comprised 22 quarters (2018Q1–2023Q2); the post‐announcement period comprised 6 quarters (2023Q3–2024Q4). Autocorrelation was assessed using the Durbin–Watson statistic. A significant positive β ₂ would indicate an immediate level change consistent with notoriety bias; β ₃ tests whether the slope changed thereafter.

2.8. Time‐to‐Onset (TTO) Analysis

TTO was computed as the interval in days between drug start date and event onset; reports with TTO < 1 or > 3650 days were excluded. TTO data were available for 14 023 report–event records from 8605 reports across the analytic dataset. TTO distributions were compared across three pre‐specified groups: early gastrointestinal reactions (nausea, vomiting, diarrhoea; n = 11 347), psychiatric events (suicidal ideation, depression, anxiety, self‐injurious ideation, major depression, psychotic disorder, depressed mood; n = 205) and late gastrointestinal events (acute pancreatitis, pancreatitis, cholelithiasis; n = 1486). The remaining reports fell outside the three pre‐specified groups. Non‐parametric Mann–Whitney U tests compared groups. Drug‐specific psychiatric TTO profiles were computed for semaglutide, liraglutide and dulaglutide (tirzepatide had insufficient TTO records owing to its later approval). As a sensitivity analysis, the three‐group comparison was repeated applying a stricter upper bound of 730 days (2 years) to reduce the influence of implausible long latencies and possible data‐entry errors. To characterise the completeness of TTO data, reports with versus without a usable TTO were compared across drug, sex, age group, reporter source and seriousness.

2.9. Adverse‐Event Co‐Occurrence Network Analysis

For the co‐occurrence network analysis, GLP‐1 RA‐exposed cases were identified by scanning the 32 quarterly DRUG files for any record mentioning a GLP‐1 RA (substring match against drugname and prod_ai for the brand and INN names listed in Section 2.2). For each identified caseid, FDA dedup (max caseversion) was applied to obtain a canonical primaryid, and PT mentions were accumulated across all caseversion records of the case. This case‐set construction yielded per‐drug counts within ±2% of the primaryid‐deduplicated counts in Section 2.1 (semaglutide 78 971; liraglutide 20 467; dulaglutide 72 848; tirzepatide 121 069); the small variance reflects multi‐version PT capture. Of the resulting cases, 168 151 had ≥ 2 reported AEs and were used for the network analysis. We constructed a co‐occurrence matrix among 1421 PTs each appearing ≥ 30 times. The observed/expected (O/E) ratio for CNS–GI co‐occurrence was O/E = P(CNS ∩ GI)/[P(CNS) × P(GI)] [19]; values > 1 indicate supra‐additive co‐activation. Drug‐specific O/E ratios characterised the multi‐system safety fingerprint of each agent.

2.10. Weber Effect With Cochran–Armitage Trend Testing

Quarterly report counts were tabulated 2018–2025. Weber‐effect was assessed by computing ROR separately for three post‐approval epochs (Year 1–2, Year 3–4, Year 5+) for each GLP‐1 RA agent, with contemporaneous non‐GLP‐1 RA reports in each epoch as reference [20, 21, 22]. A statistically significant declining trend across epochs (Cochran–Armitage trend test, P _trend < 0.05) was interpreted as Weber‐effect evidence. A stable or increasing trend was interpreted as a reporting signal whose strength does not depend on reporting novelty (i.e., not attributable to the Weber effect). Where epoch data were available for only one period (typically Year 5+), trend testing was not applied.

2.11. Outcome Severity Assessment

For each signal, the worst‐case outcome was extracted from FAERS OUTC (death DE, life‐threatening LT, hospitalisation HO, disability DS, congenital anomaly CA, other OT). Serious‐outcome rate was the proportion of reports with any of DE/LT/HO/DS/CA. Signal‐level case fatality rate (CFR) was the proportion of DE outcomes.

2.12. Cross‐Database Validation

Seventeen priority FAERS signals were re‐tested in JADER (n = 1 027 194 reports) using identical disproportionality criteria (ROR_low > 1, cell a ≥ 3). Concordance was defined as a positive signal in both databases.

2.13. Subgroup and Sensitivity Analyses

Pre‐specified subgroup analyses were stratified by sex (male/female) and age group (18–44, 45–64, ≥ 65 years; the pre‐specified < 18 stratum contained too few cases (n = 365) to yield ≥ 3 co‐reports for any priority Preferred Term and was therefore not analysable) for top GI and metabolic signals. Method‐level concordance among the four DPA metrics (ROR, PRR, BCPNN IC025, MGPS EB05) was confirmed by tabulating signals meeting 4‐of‐4, 3‐of‐4, 2‐of‐4, 1‐of‐4, and 0‐of‐4 criteria across all 22 059 PT–drug pairs [23].

2.14. Statistical Software

Disproportionality, Weber, and outcome analyses were performed in R 4.5.2 (data.table, ggplot2, openxlsx). ITS, co‐occurrence network, and case‐level data construction were performed in Python 3.11 (statsmodels, scipy, networkx, pandas). All analyses used publicly available data and did not require ethical approval.

3. Results

3.1. Study Population and Temporal Trends

The analytic dataset comprised 288 408 unique deduplicated GLP‐1 RA primary reports. Drug‐specific exposure counts (cases counted in each stratum a drug was reported in) were tirzepatide 121 069 (42.0%), semaglutide 78 971 (27.4%), dulaglutide 72 848 (25.3%), and liraglutide 20 467 (7.1%); these exceed 288 408 because cases mentioning multiple GLP‐1 RAs are counted in each drug‐specific stratum. Of analytic dataset cases, 169 116 (58.6%) were female, with median age 58 years (IQR 48–68). Tirzepatide users were younger (median age 54, IQR 45–61) (Table S1). Quarterly report volumes (Figure 1A) showed dulaglutide steady through 2018–2024, liraglutide declining steadily, semaglutide accelerating from 2021 (after Wegovy approval), and tirzepatide—approved May 2022—exceeding all other GLP‐1 RAs by 2023Q3 and reaching 17 965 reports in 2025Q4, the largest single‐quarter volume in the class.

FIGURE 1.

Overall signal landscape and active‐comparator validation. (A) Quarterly FAERS report counts (2018Q1–2025Q4) for the four GLP‐1 RA agents. (B) Bubble plot of all 544 robust signals (3‐of‐4 criterion) detected across the GLP‐1 RA class. X‐axis: number of co‐reports (log scale); Y‐axis: ROR (log scale); bubble size: number of co‐reports. Top signals labelled. (C) Forest plot of the top 20 signals by ROR magnitude across three analyses: GLP‐1 RA main analysis (red circles), SGLT2i comparator (blue squares), DPP‐4i comparator (green triangles); ROR x‐axis on log scale; dashed vertical line at ROR = 1.

3.2. Overall Signal Detection

Disproportionality analysis identified 544 unique PT–class pairs meeting the 3‐of‐4 criterion across the class, from 22 059 unique PT–class pairs with sufficient counts (signal positivity rate 2.47%; Figure 1B; Table S2). Signals spanned 11 system organ classes. Gastrointestinal/hepatobiliary disorders constituted the largest category (approximately 90 signals; max ROR 84.2), followed by metabolism and nutrition (47 signals; median ROR 6.17), drug administration events (46 signals), endocrine/neoplasms (36 signals; median ROR 6.63), eye disorders (16 signals), and—prior to PS‐stratification—12 vascular disorders with median ROR 18.23 (this category contracted substantially after PS‐stratification; see Section 3.3).

3.3. PS‐Stratified Prioritisation: Signal Triage

Of the 30 signals ranked by ROR, 21 were PS‐stable and 9 were PS‐unstable under PS‐only restriction (the 9 PS‐unstable signals in Table S3, all 21 PS‐stable signals are listed in Table S4). Under PS‐only restriction, disproportionality was eliminated or rendered non‐significant under PS‐only restriction (main → PS‐only ROR: Capillaritis 18.6 → 0.27; Pulmonary vasculitis 25.4 and Total lung capacity abnormal 21.8 → no primary‐suspect cases; Heart sounds 84.3 → 0.55; Prolonged expiration 15.2 → 0.25; Filariasis lymphatic 176.6 and Inner ear operation 52.3 → no primary‐suspect cases; Gastric ischaemia 15.0 → 3.43 and Autoimmune vasculitis 10.7 → 3.56, both non‐significant with confidence intervals spanning unity). All nine demoted signals reflect cases in which a GLP‐1 RA appears as a secondary‐suspect drug—typically liraglutide as a concomitant medication while another drug is the primary‐suspect drug. The vasculitis cluster is particularly instructive: capillaritis cases were largely concentrated in the liraglutide‐secondary‐suspect stratum (drug‐specific n = 67 in liraglutide versus 0 in semaglutide, 0 in dulaglutide, 1 in tirzepatide), but PS‐only sema‐, dula‐ and tirz‐restriction left zero capillaritis cases with the GLP‐1 RA as the primary cause. Without PS‐stratification these would have been the headline regulatory signals; with PS‐stratification, they are correctly deprioritised. Full data for the demoted signals are in Table S3, with explicit annotation. Signals were prioritised using a transparent, pre‐specified qualitative triage that combined ROR magnitude, absolute case count, primary‐suspect stability, active‐comparator consistency, clinical seriousness and external (JADER) validation; this prioritisation is a heuristic to order signals for clinical and regulatory attention rather than a validated scoring system, and no result or conclusion depends on any particular weighting of these criteria.

3.4. Active Comparator Validation

Compared with both SGLT2 inhibitor and DPP‐4 inhibitor users, the dominant GLP‐1 RA gastrointestinal signals (nausea, vomiting, acute pancreatitis, cholelithiasis, gastroparesis), metabolic signals (weight decreased, glycosylated haemoglobin increased, decreased appetite), and cyclic vomiting syndrome (Main 64.8/SGLT2 435.4/DPP‐4 72.9) all retained ROR_low > 1 (Figure 1C). Several SS‐driven respiratory and vascular signals which appeared significant in the main analysis paradoxically also showed strong signals in the SGLT2/DPP‐4 comparator analyses, reflecting a similar SS‐attribution mechanism in those antidiabetic comparator drugs. This further reinforces our decision to prioritise PS‐stable signals.

3.5. Drug‐Specific Signal Profiles

Drug‐level signal counts varied substantially: semaglutide 829, liraglutide 669, dulaglutide 267, tirzepatide 218 (Figure 2A). Higher counts for semaglutide and liraglutide largely reflect their longer post‐marketing exposure and broader approved indications, rather than greater toxicity. Tirzepatide demonstrated 50 unique signals not shared with any other GLP‐1 RA (Figure 2B; Table S5). The most substantial by case volume were injection site reaction (n = 1662, ROR 5.27), therapeutic response changed (n = 1205, ROR 37.0), and product tampering (n = 465, ROR 42.3). A separate group met statistical thresholds on very few reports and is therefore treated as sparse‐count and exploratory rather than prioritised: stress polycythaemia (ROR 51.1), non‐proliferative retinopathy (ROR 20.1), and Hürthle cell carcinoma (ROR 17.9), each n = 3 (flagged as such in Table S5). Some tirzepatide‐unique signals are plausibly related to GIP receptor co‐agonism, although this remains hypothesis‐generating.

FIGURE 2.

Drug‐specific signal profiles. (A) Total robust signal counts per agent (semaglutide 829; liraglutide 669; dulaglutide 267; tirzepatide 218). (B) Top 20 tirzepatide‐unique signals (n = 50 total) by ROR. (C) Heatmap of intra‐class ROR values (each drug vs. pooled other three GLP‐1 RAs) for 18 psychiatric preferred terms; cells annotated with numeric ROR; white stars (★) indicate statistically significant signals (ROR_low > 1 and IC025 > 0, where IC025 is the lower bound of the 95% credible interval of the Bayesian information component); values capped at 5.5 for display. The pattern shows semaglutide's dominant intra‐class psychiatric burden, with notable contributions from liraglutide and a striking absence in tirzepatide.

3.6. Psychiatric Adverse Events: Intra‐Class Comparison

In the intra‐class design (each drug vs. pooled other three GLP‐1 RAs; Figure 2C), semaglutide showed elevated signals across multiple psychiatric PTs (Table S6). The most pronounced were psychotic disorder (intra‐class ROR 4.03), major depression (ROR 4.82), self‐injurious ideation (ROR 4.03), binge eating (ROR 3.85), suicidal ideation (ROR 3.41), and suicide attempt (ROR 2.65). Liraglutide showed significant signals for hallucination (ROR 3.88), suicide attempt (ROR 3.47), completed suicide (ROR 3.09), anxiety (ROR 2.35), and depression (ROR 1.99). Dulaglutide showed only one elevated psychiatric signal (nervousness, ROR 1.25). Tirzepatide showed only depressed mood (ROR 2.19) as a borderline‐elevated signal, with all other psychiatric PTs at or below the class average—a striking absence given tirzepatide's exposure volume.

At the class level, primary‐suspect GLP‐1 RA reports with suicidal ideation (n = 934) had a serious‐outcome rate of 96.7% and case‐fatality rate 1.2%, with hospitalisation in 10.6%; completed suicide (n = 66) reports were almost uniformly fatal (CFR 97.0%); psychotic disorder (n = 76) was associated with hospitalisation in 47.4% of cases (Table S7). Importantly, only semaglutide carried a class‐level Suicidal ideation signal (n = 621, ROR 2.30 vs. all‐FAERS, signal_3of4 = TRUE); pooled class‐level Suicidal ideation was not a robust signal (ROR 1.11, signal_3of4 = FALSE). These intra‐class differences should be interpreted as drug‐specific reporting patterns rather than as evidence of differential causal risk (see Sections 4.5 and 4.11).

3.7. Time‐to‐Onset: Psychiatric Reports Cluster Early and Align With Early Gastrointestinal Reactions

TTO data were available for 14 023 reports–event records from 8605 reports (Figure 3A). The psychiatric group (n = 205) had a median TTO of 47 days (IQR 15–127), with 42.4% occurring within 30 days. This distribution was statistically indistinguishable from early GI reactions (median 42 days; Mann–Whitney U comparison p = 0.771) but significantly earlier than late GI events such as acute pancreatitis and cholelithiasis (median 145 days; p = 5.5 × 10⁻²¹). Drug‐specific analysis (Figure 3B) showed semaglutide had the earliest psychiatric TTO (median 42 days; 44.4% ≤ 30 days), followed by liraglutide (median 48 days) and dulaglutide (median 84 days), consistent with semaglutide's relatively higher CNS penetrance. Tirzepatide‐specific psychiatric TTO had insufficient records for separate plotting.

FIGURE 3.

Psychiatric signal validity. (A) Box‐and‐whisker plots of time‐to‐onset (days, log scale) for early GI reactions (n = 11 347), psychiatric events (n = 205) and late GI events (n = 1486); Mann–Whitney U pairwise p‐values shown. (B) Drug‐specific psychiatric TTO violin/box plots for semaglutide, liraglutide and dulaglutide (tirzepatide omitted owing to insufficient TTO records). (C) Interrupted time series regression of quarterly psychiatric AE proportion on time, with the 2023Q3 EMA announcement as the intervention (vertical red dashed line). Solid blue lines show fitted segments. β ₂ (level change) and β ₃ (slope change) coefficients with p‐values are annotated.

3.8. Interrupted Time Series: No Statistically Detectable EMA‐Associated Reporting Shift

The ITS model explained 62.6% of variance in quarterly psychiatric reporting proportions (R ² = 0.626; Durbin–Watson = 1.84, no significant autocorrelation). Prior to the EMA announcement, psychiatric reporting showed a modest but significant upward secular trend (β₁ = +0.000502 per quarter, SE 0.000154, p = 0.003), consistent with growing prescribing and organic awareness. Critically, the EMA announcement was not associated with a significant immediate level change (β ₂ = +0.0034, SE 0.0039, p = 0.388) or slope acceleration (β₃ = +0.00036, SE 0.0011, p = 0.751) (Table 1; Figure 3C). These findings indicate no statistically detectable EMA‐associated level or slope change, and do not exclude delayed, brand‐ or indication‐specific, or nonlinear stimulated reporting that the present model cannot rule out (see Section 4.5 and 4.11): had the EMA review artificially inflated reporting, β ₂ would be significantly positive.

TABLE 1.

Interrupted time series regression (outcome: psychiatric AE proportion per quarter).

Parameter	Coefficient	SE	p‐value	Interpretation
Intercept (β 0)	+0.02656	0.00202	< 0.001***	Baseline psychiatric proportion
Time trend (β 1)	+0.000502	0.000154	0.003**	Pre‐EMA secular increase per quarter
EMA level change (β 2)	+0.00341	0.00388	0.388 ns	No detectable EMA‐associated level change
EMA slope change (β 3)	+0.000355	0.00110	0.751 ns	No post‐EMA acceleration
Model fit	R 2 = 0.626	n = 28 quarters	—	Pre‐EMA = 22 q; post‐EMA = 6 q

Note: Significance codes: ***p < 0.001; **p < 0.01; ns, not significant. Pre‐EMA period = 2018Q1–2023Q2 (22 quarters); post‐EMA period = 2023Q3–2024Q4 (6 quarters). Durbin–Watson = 1.84 (no significant autocorrelation).

3.9. AE Co‐Occurrence Network: The CNS–GI Safety Fingerprint

Co‐occurrence analysis across 168 151 GLP‐1 RA cases with ≥ 2 reported AEs yielded 32 931 unique high‐frequency PT pairs (≥ 10 co‐reports) among 1421 PTs each appearing ≥ 30 times (Table 2). The strongest pair was Nausea–Vomiting (n = 10 527; O/E 3.55), followed by Diarrhoea–Nausea (n = 7005; O/E 2.43) and Diarrhoea–Vomiting (n = 6189; O/E 3.60), confirming the expected GI cluster (Figure 4A). Top non‐GI co‐occurrences included Headache–Nausea (n = 3065; O/E 2.42; the strongest CNS–GI pair) and Incorrect dose administered–Injection site pain (n = 3783; O/E 2.77; an administration‐error cluster reflecting the injectable‐pen dosing form factor). CNS–GI co‐occurrence exceeded additive expectation for all four agents (Table 2; Figure 4B), with values increasing in the order semaglutide (O/E 1.13) < liraglutide (1.24) < dulaglutide (1.35) < tirzepatide (1.39). The graded supra‐additive pattern, with the dual GIP/GLP‐1 agonist tirzepatide showing the highest co‐activation, suggests progressively pronounced shared central engagement of GI and CNS pathways across the class—consistent with GLP‐1 (and GIP) receptor expression in the area postrema, nucleus tractus solitarius and limbic structures.

TABLE 2.

CNS–GI co‐occurrence analysis by GLP‐1 RA agent.

Drug	Cases (≥ 2 AEs)	Psychiatric AEs	GI AEs	CNS ∩ GI	O/E ratio
Semaglutide	56 156	13 889	26 322	7353	1.13
Liraglutide	13 965	3408	5069	1537	1.24
Dulaglutide	42 556	7054	14 247	3191	1.35
Tirzepatide	59 524	8945	21 844	4552	1.39

Note: Cases (≥ 2 AEs) = number of caseid‐deduplicated reports with at least two reported Preferred Terms in the per‐drug stratum. O/E ratio = P(CNS ∩ GI)/[P(CNS) × P(GI)]; values > 1 indicate supra‐additive co‐activation.

FIGURE 4.

Co‐occurrence and cross‐database validation. (A) Top 10 AE co‐occurrence pairs across all GLP‐1 RA cases with ≥ 2 reported AEs (n = 168 151); colour by pair type (GI–GI orange; CNS–GI blue; GI–Other/Other–Other grey). Co‐occurrence counts annotated to the right of each bar. (B) CNS–GI O/E ratios by GLP‐1 RA agent (liraglutide 1.24; semaglutide 1.13; dulaglutide 1.35; tirzepatide 1.39); n = number of multi‐AE cases per drug stratum (annotated within bars); dashed line at O/E = 1 indicates additive expectation. (C) Scatter of FAERS ROR vs. JADER ROR for the 17 priority signals submitted to cross‐database validation; diagonal y = x line shown; concordant signals in green (16/17); discordant signal (acute kidney injury) in red.

3.10. Weber Effect Across Post‐Approval Epochs

Weber analyses were performed across three post‐approval epochs (Year 1–2, Year 3–4, Year 5+) for semaglutide and tirzepatide, the two agents with sufficient temporal coverage (Figure 5; Table S8). Three patterns emerged. First, GI signals (nausea, cholelithiasis, pancreatitis acute, impaired gastric emptying) remained broadly stable across epochs for both agents, with Cochran–Armitage P _trend values 0.26–0.72—consistent with persistent reporting associations rather than reporting‐novelty (Weber) artefacts. Second, tirzepatide cyclic vomiting syndrome ROR escalated across post‐approval periods (Year 3–4 ROR 76.1, 95% CI 33.9–170.7; Year 5+ ROR 118.1, 49.9–279.4), the opposite of a Weber effect. Third, tirzepatide weight‐decreased ROR declined across epochs (Year 1–2 ROR 3.66 → Year 3–4 ROR 3.47 → Year 5+ ROR 3.30; P _trend = 0.011), the only signal in our analysis showing classical Weber‐effect attenuation. This dissociation—a pharmacodynamic‐related signal (weight decreased) attenuating while AE signals (cyclic vomiting, pancreatitis) do not—suggests Weber‐effect dynamics may be more pronounced for expected pharmacological effects than for unexpected adverse events, an observation we discuss in Section 4.6. Full epoch‐stratified ROR estimates (95% CI) with Cochran–Armitage trend statistics for both agents are provided in Table S8.

FIGURE 5.

Weber‐effect analysis: ROR across post‐approval epochs. ROR with 95% CI for six gastrointestinal/metabolic preferred terms (cholelithiasis, cyclic vomiting syndrome, impaired gastric emptying, nausea, pancreatitis acute, weight decreased) across three post‐approval epochs (Year 1–2, Year 3–4, Year 5+) for semaglutide (red) and tirzepatide (blue). Y‐axis log scale; dashed horizontal line at ROR = 1. Cochran–Armitage P _trend annotated where computable; * = p < 0.05. Note the divergent patterns: Tirzepatide weight‐decreased shows classical Weber decline (P _trend = 0.011), while tirzepatide cyclic vomiting syndrome escalates and other GI signals remain stable.

3.11. Cross‐Database Validation in JADER

Of 17 priority signals re‐tested in JADER (n = 1 027 194 total reports), 16 (94.1%) showed concordant signal direction (Figure 4C; Table S9). JADER ROR magnitudes were systematically 2.21‐fold higher than FAERS values (Spearman ρ = 0.78), reflecting the smaller Japanese reference population. The single non‐concordant signal (acute kidney injury) showed FAERS ROR < 1 (no signal) but JADER ROR > 1, plausibly reflecting differences in concomitant medication patterns and reporting culture. Notably, the GI/hepatobiliary cluster (Cholelithiasis, Pancreatitis acute, Cholecystitis acute) showed near‐perfect concordance with very high JADER RORs (27–30), strongly supporting the robustness and trans‐population generalizability of the GI safety fingerprint.

3.12. Subgroup Analyses

Pre‐specified subgroup analyses by sex (Table S10) and by age group (Table S11) showed broadly consistent signal patterns. The headline GI/metabolic signals (Pancreatitis acute, Cholelithiasis, Impaired gastric emptying, Weight decreased) retained ROR_low > 1 in both male and female reporters and across all three analysable age strata (18–44, 45–64, ≥ 65 years). Cyclic vomiting syndrome could be assessed only by age, because sex was recorded for just 18/254 (7%) of its reports; it remained significant (ROR_low > 1) in every age stratum, with point estimates rising with age, but rests on sparse per‐stratum counts and should be interpreted as exploratory. The complete sex‐ and age‐stratified ROR estimates (95% CI) for all priority signals are tabulated in Tables S10 and S11, respectively.

4. Discussion

4.1. Principal Findings

This is, to our knowledge, one of the most comprehensive multi‐method pharmacovigilance analyses of GLP‐1 RAs reported to date and, importantly, the first to operationalise primary‐suspect sensitivity stratification as a pre‐specified step to separate primary‐suspect‐stable signals from secondary‐suspect‐driven ones. From 288 408 unique GLP‐1 RA primary reports in FAERS, we identified 544 robust adverse‐event signals; PS‐stratification then triaged these into 21 PS‐stable headline signals and 9 PS‐unstable signals demoted to Supporting Information S1. The headline safety profile is dominated by gastrointestinal/hepatobiliary effects (cyclic vomiting syndrome, gallbladder injury, impaired gastric emptying, pancreatitis acute, cholelithiasis), supplemented by drug‐specific concentrations of medullary thyroid cancer (liraglutide), optic ischaemic neuropathy (semaglutide), and the only class‐level psychiatric signal (semaglutide‐driven Suicidal ideation). Three features of the psychiatric reporting patterns are relevant to their interpretation: early time‐to‐onset aligned with early GI reactions, absence of a statistically detectable change at the 2023 EMA announcement in the ITS model, and supra‐additive CNS–GI co‐occurrence. As detailed in Section 4.5, these observations are hypothesis‐generating and do not establish causality.

4.2. Gastrointestinal and Hepatobiliary Signals

The gastrointestinal landscape is consistent with known mechanism: GLP‐1 receptor activation in the enteric nervous system slows gastric emptying and intestinal motility; reduced gallbladder contractility and rapid weight loss together promote bile lithogenicity. The class signal for Pancreatitis (combined) of ROR 13.8 and the PS‐stable Cholelithiasis ROR 5.20 align with randomised trial meta‐analyses [24, 25]. The PS‐stable gallbladder injury ROR 84.2 and Impaired gastric emptying ROR 45.1 represent the specificity‐extreme end of the GI fingerprint. JADER confirmation at higher RORs (Cholelithiasis JADER ROR 30.5; Pancreatitis acute JADER ROR 27.2) substantially strengthens these conclusions.

4.3. Cyclic Vomiting Syndrome as a Sentinel Late‐Emerging Signal

Cyclic vomiting syndrome stands out as a highly specific, semaglutide‐dominant signal (drug‐specific n = 247, ROR 224.7 in semaglutide; class n = 254, ROR 64.8) with characteristics that warrant prospective study rather than dismissal as a reporting artefact. Three features are noteworthy. First, the signal is PS‐stable (Main 64.8/PS‐only 74.8) and comparator‐robust (vs. SGLT2i ROR 435; vs. DPP‐4i ROR 73), indicating a primary‐suspect‐stable, comparator‐robust reporting signal rather than one driven by secondary‐suspect attribution. Second, hospitalisation is the modal serious outcome. Third, while the class‐level signal is stable across post‐approval epochs (no Weber decline), the tirzepatide‐specific signal escalates (Year 3–4 ROR 76 → Year 5+ ROR 118), suggesting cumulative‐exposure‐dependent risk. Possible mechanisms include chronic GI motility dysregulation, autonomic reorganisation, or emerging hyperemesis‐syndrome‐like presentations in cannabinoid‐naïve patients. Prospective characterisation in dedicated cohort studies is now warranted, as is structured pharmacovigilance for hyperemesis‐pattern presentations in patients on chronic GLP‐1 RA therapy.

4.4. Why We Demoted the Vasculitis and Respiratory Signals: A Methodological Strength

Without primary‐suspect sensitivity stratification, our top‐priority signals would have included capillaritis (Main ROR 18.6), pulmonary vasculitis (25.4), total lung capacity abnormal (21.8), heart sounds (84.3), and prolonged expiration (15.2). Under PS‐only restriction, these signals collapsed to non‐significant levels (PS‐only RORs 0.25–0.55, with several signals having no remaining primary‐suspect cases), revealing that disproportionality was driven by reports in which a GLP‐1 RA—predominantly liraglutide—appeared as a secondary‐suspect drug, with another agent designated the primary cause. Capillaritis is the most striking example: drug‐specific case counts were 67 in liraglutide, 0 in semaglutide, 0 in dulaglutide, 1 in tirzepatide; in the liraglutide PS‐only stratum, zero capillarity cases remained. This pattern is most parsimoniously explained by patients receiving liraglutide concurrently with another drug (often immunomodulatory or oncologic) which is the actual cause of the Capillaritis. Without PS‐stratification, the FAERS signal mathematics would have produced a headline implicating GLP‐1 RAs in a serious vascular pathology that the data themselves do not support. We therefore consider PS‐only sensitivity stratification a methodological pillar that should be standard practice in spontaneous‐reporting pharmacovigilance, particularly for high‐volume drug classes used in patients with multiple concomitant medications.

4.5. Psychiatric Signals: Interpreting the Psychiatric Reporting Patterns

Three features of the psychiatric reporting patterns inform their interpretation, although none establishes causality. (i) The TTO of psychiatric events (median 47 days) matches early GI reactions (median 42 days; p = 0.77), not late GI events (median 145 days; p < 10⁻²⁰). Both peak within the first 30–47 days, when central GLP‐1 receptor occupancy reaches steady state. This temporal alignment fits a shared acute‐CNS‐activation model rather than slow‐onset toxicity. (ii) The ITS model found neither a significant level change nor a slope acceleration at the 2023Q3 EMA announcement, and reporting was already trending upward beforehand; however, this does not exclude other forms of stimulated reporting. (iii) The intra‐class design holds the class‐shared indication, channelling and expectation pressures constant, so the within‐class gradient (semaglutide ≫ liraglutide ≫ dulaglutide ≈ tirzepatide for psychotic, depressive and self‐injurious PTs) cannot be attributed solely to class‐wide media attention, although it may still reflect differential channelling, indication mix, or formulation/brand‐specific reporting rather than differential pharmacological risk. Importantly, only semaglutide carries a class‐level (vs. all‐FAERS) Suicidal ideation signal; the pooled class signal (ROR 1.11) is not robust. This suggests that the broader regulatory concern about GLP‐1 RAs and suicidality may be more accurately framed as a semaglutide‐specific concern within the class.

The mechanistic candidate for semaglutide's prominence is differential CNS access: the molecule's albumin‐binding fatty‐acid chain, longer half‐life. Preclinical studies with labelled semaglutide indicate access to selected brain regions via circumventricular organs and ventricle‐adjacent sites without crossing the blood–brain barrier [26], whereas human GLP‐1 receptor PET imaging has not demonstrated appreciable uptake in brain regions protected by the blood–brain barrier [27]. These observations are consistent with limited, region‐specific rather than global central access, and the link to the psychiatric reporting patterns remains hypothesis‐generating. GLP‐1 receptors in the ventral tegmental area, nucleus accumbens and prefrontal cortex modulate dopaminergic and serotonergic neurotransmission relevant to mood, reward and psychotic phenomena. An alternative or complementary mechanism is rapid weight loss itself: profound and sudden metabolic‐state change can precipitate decompensation in vulnerable individuals through leptin, ghrelin and energy‐homeostasis pathways. These hypotheses are not mutually exclusive.

Several sources of residual confounding cannot be resolved within FAERS and may contribute to the observed psychiatric reporting patterns. First, Ozempic, Wegovy and Rybelsus differ in indication, dose, route/formulation and user population, and we could not reliably separate these strata in the present workflow; the semaglutide signal therefore aggregates heterogeneous products. Second, obesity and diabetes populations differ in baseline psychiatric burden, healthcare‐contact frequency and reporting behaviour. Third, rapid weight loss may itself be associated with psychiatric symptoms or with their reporting, independently of any direct drug effect. Fourth, differential prescribing (channelling) and drug‐specific, brand‐specific media attention may not be captured by a single regulatory‐announcement ITS model. For these reasons, the within‐class gradient should be interpreted as a drug‐specific reporting pattern, not as a demonstration of differential causal risk.

4.6. The Tirzepatide Paradox: Few Specific Signals, Highest CNS–GI Co‐Activation

Tirzepatide simultaneously displays the lowest intra‐class psychiatric signal count (depressed mood only, with most other psychiatric PTs even showing intra‐class ROR < 1) and the highest CNS–GI co‐occurrence ratio (O/E 1.39). This pattern is notable given that SURPASS‐2 demonstrated superior glycaemic and weight outcomes for tirzepatide over semaglutide in T2DM patients [28], indicating that the differential psychiatric/CNS profile observed here is not simply an efficacy gradient. These two metrics measure different things. Disproportionality detects events whose reporting is over‐represented relative to a reference population, treating each PT independently. The co‐occurrence O/E asks how often two events are reported in the same case relative to chance. A drug can therefore have non‐elevated psychiatric DPA (because absolute psychiatric reporting frequency is in line with class expectations) yet relatively higher CNS–GI co‐occurrence (because when psychiatric AEs do occur, they tend to occur with GI AEs in the same case). The combination is most consistent with generalised central activation: tirzepatide patients who experience CNS effects experience them concurrently with GI effects, rather than as standalone psychiatric presentations.

Importantly, the CNS–GI O/E gradient is graded rather than dichotomous. The ordering semaglutide (1.13) < liraglutide (1.24) < dulaglutide (1.35) < tirzepatide (1.39) is monotonic and modest in absolute terms—closer to one and a half than to two—but consistent across all four agents. This graded pattern matches plausible mechanistic gradients: tirzepatide's dual GIP/GLP‐1 activity, dulaglutide's longer plasma half‐life favouring sustained central exposure, and semaglutide's higher CNS penetrance arguably allowing more isolated psychiatric presentations. Three corollaries follow. First, low intra‐class psychiatric DPA does not equal low psychiatric risk for tirzepatide. Second, the dual GIP/GLP‐1 mechanism may produce a quantitatively different, GI‐dominant CNS engagement profile, plausibly because GIP receptors modulate GLP‐1 receptor signalling through receptor cross‐talk. Third, surveillance for tirzepatide should pay particular attention to combined CNS/GI presentations rather than relying on isolated psychiatric coding. The escalating cyclic vomiting syndrome signal already provides a concrete instance of this pattern.

4.7. Optic Ischaemic Neuropathy: A New Focus

Optic ischaemic neuropathy emerged as a robust, PS‐stable, semaglutide‐concentrated signal (drug‐specific n = 528, ROR 75.9; class n = 624, ROR 26.5) supported by the broader eye‐disorder cluster (16 signals, median ROR 3.14; max ROR 31.5 for corrective lens users). This finding is consistent with a recent cohort study reporting a 4‐fold increase in NAION risk with semaglutide [29]. Mechanistically, GLP‐1 receptors in the retinal pigment epithelium and choroidal vasculature, combined with rapid weight‐loss‐induced fluid‐volume changes, may compromise optic nerve perfusion in susceptible patients (those with crowded optic discs or pre‐existing vasculopathy). We recommend that ophthalmology referral be considered for any GLP‐1 RA‐treated patient reporting acute visual change, and that prospective surveillance be implemented in dedicated cohorts.

4.8. Medullary Thyroid Cancer: A Liraglutide‐Concentrated Signal

Medullary thyroid cancer (MTC) appears as a PS‐stable signal (class n = 112, ROR 41.1; PS‐only ROR 48.4) with liraglutide carrying the highest drug‐specific magnitude (n = 19, ROR 53.8). This finding is biologically consistent with the FDA boxed warning derived from rodent C‐cell tumour studies [30]. JADER did not have sufficient MTC cases to validate, but the FAERS signal is among the most internally consistent in our analysis (4‐of‐4 method consistency in disprop_consistency). For prescribers, this suggests prudence in patients with personal or family history of MTC or multiple endocrine neoplasia type 2 (MEN‐2), particularly when initiating liraglutide.

4.9. Weber Findings: A Pharmacodynamic‐vs‐Adverse‐Event Dissociation

Our Weber analysis revealed an intriguing dissociation. Tirzepatide weight‐decreased—a pharmacodynamic effect closely tracking the drug's intended mechanism—showed a classical Weber decline (Year 1–2 ROR 3.66 → Year 5+ ROR 3.30; P _trend = 0.011). In contrast, GI adverse‐event signals (nausea, cholelithiasis, pancreatitis acute, impaired gastric emptying) remained stable or slightly increased across the same epochs, and cyclic vomiting syndrome escalated. This pattern suggests Weber‐effect dynamics—reporting‐novelty‐driven signal attenuation—operate primarily on expected, on‐mechanism findings rather than on unexpected adverse events. If true, this has methodological implications: a Weber‐stable AE signal is less consistent with a pure reporting‐novelty artefact than a Weber‐attenuating one, and Weber‐effect testing should be part of standard signal evaluation.

4.10. Methodological Contributions

Four methodological elements may be useful beyond this study. First, primary‐suspect sensitivity stratification (Section 2.4, Section 3.3, and Section 4.4) operationalises a previously informal practice of noting when signals depend on secondary‐suspect attribution; for high‐volume drug classes used alongside many concomitants, this triage is critical. Second, the ITS bias‐quantification template provides a standard test for notoriety bias following any high‐profile regulatory announcement. Third, the O/E co‐occurrence ratio is a simple, interpretable metric for the multi‐system safety fingerprint of a drug class, complementing single‐PT DPA. Fourth, the Weber + Cochran–Armitage trend‐testing combination distinguishes persistent pharmacological signals from reporting‐novelty artefacts.

4.11. Limitations

FAERS is subject to under‐reporting, stimulated reporting and incomplete data [31]. The absence of denominator data (prescription volumes) precludes true incidence estimation. Confounding by indication, BMI, baseline psychiatric history and concomitant medications cannot be fully eliminated by either active comparator or PS‐stratification, although both designs substantially attenuate them. PS‐only sensitivity stratification was applied to the top 30 ROR‐ranked signals; lower‐ranked signals (ranks 31–544) were not individually retested. Extending PS‐only re‐testing to all 164 robust signals supported by ≥ 100 reports confirmed this: 143 remained PS‐stable and 21 became PS‐unstable (lost significance or > 30% attenuation of ROR), all of them non‐prioritised terms (respiratory‐function, allergy, vascular and laboratory signals) demoted through the same secondary‐suspect mechanism, while every prioritised (Table 3) signal remained PS‐stable (Table S12). For the co‐occurrence network analysis we used FAERS caseid‐level deduplication to capture PT mentions across all caseversion records of each case (yielding per‐drug counts within ±2% of the primaryid‐level deduplication used in the main disproportionality analyses); the resulting case set is substantively consistent with the disproportionality dataset but better captures multi‐version cases. We did not stratify semaglutide by formulation (Ozempic vs. Wegovy vs. Rybelsus, which differ in dose and indication) or tirzepatide by dose (5/10/15 mg); within‐product heterogeneity may exist and is a high‐priority follow‐up. Usable time‐to‐onset data were available for 14 023 report–event records from 8605 reports (≈3% of analytic cases). A sensitivity analysis restricting TTO to ≤ 730 days (2 years) preserved the qualitative ordering, with psychiatric reports remaining statistically indistinguishable from early GI reactions and significantly earlier than late GI/hepatobiliary events (Table S13). Reports with versus without usable TTO data differed significantly by hdrug, reporter type, seriousness, sex and age (all p < 0.001; Table S14), so the TTO distributions should be read as descriptive onset patterns among reports with usable dates rather than as unbiased onset distributions for all GLP‐1 RA reports; differential completeness between the psychiatric and gastrointestinal subsets may also influence their comparison. The ITS post‐period is six quarters, limiting power to detect delayed notoriety effects beyond ~18 months. The intra‐class psychiatric comparison assumes exchangeability of patient populations across the four agents, which is imperfect given divergent approved indications—particularly the substantial weight‐loss‐without‐diabetes population on semaglutide (Wegovy) which is younger and more female than the diabetes‐population on dulaglutide. The active‐comparator design is similarly limited: although SGLT2 and DPP‐4 inhibitors share the diabetes indication, GLP‐1 RAs expanded over 2018–2025 into a younger, more predominantly female obesity population without diabetes, so the treated populations diverged over time—an indication/channelling drift that active‐comparator matching cannot fully remove.

TABLE 3.

Selected, clinically prioritised primary‐suspect (PS)‐stable signals for GLP‐1 receptor agonists (FAERS 2018Q1–2025Q4).

Preferred term	n (class)	ROR (95% CI)	PS‐only ROR	Drug‐dominant	Notes
Gallbladder injury	133	84.2 (62.3–113.9)	99.3	Sema (n = 117)/Lira (n = 27)	Hepatobiliary
Cyclic vomiting syndrome	254	64.8 (53.0–79.2)	74.8	Semaglutide (n = 247)	Sentinel signal (see Section 4.3)
Weight loss poor	1124	42.2 (38.8–45.9)	49.4	Class‐shared	Pharmacodynamic
Impaired gastric emptying	4146	45.1 (43.1–47.2)	50.0	Class‐shared	Core mechanism
Medullary thyroid cancer	112	41.1 (31.5–53.6)	48.4	Liraglutide (n = 19, ROR 53.8)	Boxed‐warning consistent
Optic ischaemic neuropathy	624	26.5 (23.9–29.3)	30.3	Semaglutide (n = 528, ROR 75.9)	New focus (see Section 4.7)
Allodynia	325	34.1 (29.4–39.6)	39.6	Class‐shared	Neuropathic
Lack of satiety	53	25.4 (18.0–35.8)	30.8	Class‐shared	Pharmacodynamic
Glycosylated haemoglobin increased	4477	17.6 (17.0–18.2)	15.7	Class‐shared	Treatment‐related
Cholelithiasis	1513	5.20 (4.93–5.49)	5.37 a	Liraglutide (n = 247, ROR 11.5)/Sema (n = 666, ROR 8.12)	Hepatobiliary
Pancreatitis acute	1045	4.13 (3.87–4.40)	3.60 a	Liraglutide (n = 173, ROR 9.10)	Pancreatitis combined ROR 13.8
Suicidal ideation (Sema‐only)	621	2.30 (2.11–2.48)	~2.3 a	Semaglutide‐specific	Class‐level not significant

Note: Class‐level statistics contains analytic dataset N = 288 408 GLP‐1 RA reports vs. 11 319 150 non‐GLP‐1 reference. Drug‐dominant column identifies the GLP‐1 RA agent contributing the largest share of cases; the corresponding drug‐specific n and ROR are given. Signals which lost significance are listed in Table S3.

^a Cholelithiasis, pancreatitis acute, and suicidal ideation PS‐only ROR estimated from drug‐specific PS analyses; signals retain significance and direction. PS‐only ROR shown is class‐aggregated.

4.12. Clinical and Regulatory Implications

We recommend (i) baseline psychiatric screening (e.g., PHQ‐9, Columbia Suicide Severity Rating Scale) before semaglutide initiation, with structured re‐assessment at 4, 8 and 12 weeks—the window of peak psychiatric TTO; (ii) explicit informed‐consent discussion of within‐class psychiatric risk gradient when patients face choice between GLP‐1 RA agents, particularly distinguishing semaglutide's higher signal load from tirzepatide's relative paucity of specific psychiatric signals; (iii) ophthalmology referral for any acute visual change in GLP‐1 RA users, especially those on semaglutide; (iv) thyroid history‐taking and ultrasound surveillance for liraglutide candidates with personal or family MTC/MEN‐2 history; (v) for tirzepatide, surveillance for combined CNS/GI presentations (the cyclic vomiting syndrome pattern), with dose–response analyses as longer‐term post‐marketing data accumulate; and (vi) regulatory consideration of whether current product information adequately reflects within‐class risk differentials and whether certain currently emphasised vasculitis warnings should be reconsidered in the light of PS‐stratified analyses.

5. Conclusions

This comprehensive multi‐method pharmacovigilance analysis of 288 408 GLP‐1 RA‐associated FAERS reports identifies 544 robust adverse‐event signals with a heterogeneous profile across agents. Primary‐suspect sensitivity stratification—a methodological pillar we recommend for routine adoption—reduces these to 21 PS‐stable headline signals and reassigns 9 secondary‐suspect‐driven signals (Capillaritis, pulmonary vasculitis, total lung capacity abnormal, and others) to Supporting Information S1. The prioritised GLP‐1 RA reporting profile is dominated by gastrointestinal/hepatobiliary effects with semaglutide‐concentrated psychiatric and ophthalmic signals, liraglutide‐concentrated medullary thyroid cancer, and an emerging tirzepatide‐cumulative cyclic vomiting syndrome pattern. Psychiatric reporting patterns showed several convergent observations—including time‐to‐onset alignment with early gastrointestinal reactions, the absence of a statistically detectable interrupted‐time‐series change at the 2023 EMA announcement, supra‐additive CNS–GI co‐occurrence with a graded gradient peaking at tirzepatide (O/E 1.39), and persistence of within‐class drug‐specific gradients. These convergent observations are hypothesis‐generating and do not establish causality or exclude residual reporting artefact; they require confirmation in pharmacoepidemiologic studies with denominators and patient‐level covariate adjustment. Within the class, semaglutide alone carries a class‐level Suicidal ideation signal warranting structured monitoring during the first 3 months, while tirzepatide displays a distinctive profile of fewer specific psychiatric signals but the highest CNS–GI co‐activation, which we interpret as generalised central engagement rather than reduced central liability. Optic ischaemic neuropathy and medullary thyroid cancer demand active case‐finding. The 94.1% cross‐database validation rate in JADER confirms robustness of the primary findings.

Author Contributions

Shiyi Xu: conceptualisation, methodology, data curation, formal analysis, visualisation, writing – original draft and editing.

Funding

The author has nothing to report.

Ethics Statement

This study used de‐identified, publicly available pharmacovigilance data and did not require ethical approval.

Conflicts of Interest

The author declares no conflicts of interest.

Supporting information

Figure S1: Study flow diagram: how de‐duplicated reports contribute to the different analyses.

Table S1: Baseline characteristics of the de‐duplicated GLP‐1 RA cohort (N = 288 408).

Table S2: The 30 preferred terms with the highest class‐level ROR (cohort‐corrected), with disproportionality metrics and 3‐of‐4 signal status.

Table S3: Nine PS‐unstable (demoted) signals that met the class‐level 3‐of‐4 criterion but did not remain signals under primary‐suspect (PS)‐only restriction.

Table S4: All 21 primary‐suspect (PS)‐stable signals retained under PS‐only restriction (complement of the 9 PS‐unstable signals in Table S3).

Table S5: Tirzepatide‐unique signals (n = 50): significant for tirzepatide and not for any other GLP‐1 RA, ranked by ROR (cohort‐corrected).

Table S6: Within‐class (intra‐class) psychiatric signals among GLP‐1 RAs (cohort‐corrected).

Table S7: Reported outcome severity for selected psychiatric preferred terms among primary‐suspect GLP‐1 RA reports (cohort‐corrected).

Table S8: Weber‐effect analysis: ROR (95% CI) by post‐approval epoch for semaglutide and tirzepatide, with Cochran–Armitage P‐trend.

Table S9: Cross‐database validation: FAERS vs. JADER concordance for 17 priority signals.

Table S10: Sex‐stratified subgroup analysis: ROR (95% CI) for priority GLP‐1 RA signals in male and female reporters (cohort‐corrected, N = 288 408).

Table S11: Age‐stratified subgroup analysis: ROR (95% CI) for priority GLP‐1 RA signals by age group (cohort‐corrected).

Table S12: Extended primary‐suspect (PS) sensitivity analysis: all robust signals supported by ≥ 100 reports, with main‐analysis and PS‐only ROR and PS‐stability.

Table S13: Time‐to‐onset under a stricter 730‐day (2‐year) upper cut‐off (sensitivity analysis).

Table S14: Completeness of time‐to‐onset data: characteristics of reports with versus without a usable TTO.

Xu S., “Safety Signals of GLP‐1 Receptor Agonists: A Multi‐Method Pharmacovigilance Analysis of FAERS (2018–2025) With Sensitivity‐Stratified Prioritisation, Notoriety‐Bias Assessment, and Cross‐Database Validation,” Diabetes, Obesity and Metabolism 28, no. 10 (2026): 9251–9264, 10.1111/dom.71138.

Data Availability Statement

FAERS data are publicly available from the U.S. FDA at https://fis.fda.gov/extensions/FPD‐QDE‐FAERS/. JADER data are publicly available from the Pharmaceuticals and Medical Devices Agency of Japan. Analytic R and Python code, intermediate datasets, and the complete signal tables are available from the corresponding author upon reasonable request.