1. Introduction

Semaglutide, a glucagon-like peptide-1 (GLP-1) receptor agonist, has emerged as a cornerstone therapy for both Type 2 Diabetes Mellitus (T2DM) and chronic weight management. The drug's dual approval—for glycemic control in diabetes (Ozempic®, Rybelsus®) and for weight reduction in obesity (Wegovy®)—has led to widespread prescription across distinct patient populations with fundamentally different baseline characteristics.

While randomized controlled trials have established the efficacy and general safety of semaglutide in both indications, post-marketing surveillance through spontaneous reporting systems like the FDA Adverse Event Reporting System (FAERS) provides critical real-world evidence about rare or delayed adverse events that may not emerge in pre-approval studies. Importantly, patients receiving semaglutide for obesity differ from those with T2DM in multiple respects: they typically have better baseline glycemic control, different comorbidity profiles, distinct concomitant medication patterns, and may experience more rapid weight loss.

Traditional pharmacovigilance analyses often examine drug-adverse event associations in aggregate, potentially obscuring important subgroup-specific risks. Data-driven subgroup discovery—an unsupervised exploratory approach—systematically evaluates multiple patient covariates to identify which demographic or clinical factors most significantly modify drug safety signals.

This study employed a data-driven discovery framework to:

1. Evaluate the quality and completeness of key covariates (age, sex, indication) in semaglutide SAE reports
2. Identify which covariate demonstrates the most clinically meaningful effect modification
3. Perform rigorous stratified disproportionality analysis for the highest-priority subgroup comparison
4. Quantify indication-specific signal strength using multiple complementary metrics (ROR, PRR, IC)

Our hypothesis was that the underlying indication for semaglutide prescribing would emerge as the most significant effect modifier, reflecting fundamental pathophysiological differences between diabetic and non-diabetic obesity populations.

2. Methods

2.1 Data Source

We analyzed data from the FDA Adverse Event Reporting System (FAERS), the largest spontaneous adverse event reporting database globally. FAERS contains de-identified safety reports submitted by healthcare professionals, consumers, and pharmaceutical manufacturers worldwide. The database includes demographic information, adverse event terms coded using Medical Dictionary for Regulatory Activities (MedDRA), drug information including indications, and report outcomes.

2.2 Study Design and Case Selection

This was a retrospective pharmacovigilance study employing a case-non-case design within FAERS. We included all serious adverse event reports listing semaglutide as a suspect medication. Reports were classified as "serious" based on FDA criteria: death, life-threatening illness, hospitalization, disability, congenital anomaly, or other medically important condition.

Search Query: patient.drug.medicinalproduct:semaglutide AND serious:1

The analysis period encompassed all available quarters from semaglutide's initial approval (2017) through the most recent publicly available data (Q1 2024).

2.3 Data-Driven Subgroup Discovery Protocol

Phase I: Covariate Quality Assessment

Per our pre-specified protocol, candidate covariates for subgroup analysis included:

• Sex (male/female)
• Age (continuous, categorized as <45, 45-64, ≥65 years)
• Indication (free-text drug indication fields)
• Weight (continuous kg)

For each covariate, we calculated the missing data rate: $$\text{Missing Rate} = \frac{N_{\text{total}} - N_{\text{non-missing}}}{N_{\text{total}}} \times 100$$

Covariates exceeding 40% missingness were excluded from further analysis per data quality thresholds established in pharmacovigilance methodology literature.

Phase II: Exploratory Signal Mining

For the top 15 most frequently reported SAEs (by count), we performed preliminary disproportionality screening across available covariate strata. This identified candidate adverse events showing potential heterogeneity across subgroups.

Phase III: Effect Modification Screening

For promising SAE-covariate combinations, we assessed statistical interaction using principles analogous to the Breslow-Day test. The covariate demonstrating the greatest divergence in stratum-specific Reporting Odds Ratios (RORs) was selected as the primary effect modifier for detailed analysis.

2.4 Statistical Analysis

2x2 Contingency Table Construction

For each stratified analysis, we constructed standard 2×2 contingency tables:

Where the comparator group comprised all other drugs in FAERS serious reports.

Disproportionality Metrics

Reporting Odds Ratio (ROR): $$ROR = \frac{a \times d}{b \times c}$$

$$SE(\ln ROR) = \sqrt{\frac{1}{a} + \frac{1}{b} + \frac{1}{c} + \frac{1}{d}}$$

$$95$$

Proportional Reporting Ratio (PRR): $$PRR = \frac{a/(a+b)}{c/(c+d)}$$

Information Component (IC) - Bayesian Metric: $$IC = \log_2\left(\frac{a \times (a+b+c+d)}{(a+b) \times (a+c)}\right)$$

$$IC_{025} = IC - 1.96 \times SE(IC)$$

Where $IC_{025} > 0$ is considered a positive signal under WHO-UMC criteria.

Hypothesis Testing

For each 2×2 table, we tested the null hypothesis:

• $H_0$: There is no association between semaglutide and the target SAE (ROR = 1)
• $H_A$: There is an association between semaglutide and the target SAE (ROR ≠ 1)

Chi-square statistics with Yates' continuity correction were computed using scipy.stats.chi2_contingency. Statistical significance was defined as p < 0.05.

Effect Modification Test

To formally test for indication-based effect modification, we compared stratum-specific RORs (T2DM vs. Obesity) using the ratio of odds ratios method. A z-test for the difference in log-RORs was performed:

$$z = \frac{\ln(ROR_{obesity}) - \ln(ROR_{diabetes})}{\sqrt{SE^2_{obesity} + SE^2_{diabetes}}}$$

2.5 Software

All analyses were performed using Python 3.10 with the following packages: requests (API queries), scipy.stats (statistical tests), and standard mathematical functions for signal metric calculations. OpenFDA API endpoints were accessed with appropriate rate limiting.

3. Results

3.1 Covariate Quality Assessment

A total of 6,440 unique serious adverse event reports for semaglutide were retrieved from FAERS. Covariate completeness analysis revealed:

Sex was excluded from formal subgroup analysis due to excessive missingness (>40%). Age, indication, and weight met data quality thresholds and proceeded to exploratory mining.

3.2 Top Serious Adverse Events

The 15 most frequently reported SAEs for semaglutide were:

3.3 Age Distribution Analysis

Among reports with available age data (n=5,695):

Age-based subgroup analysis did not reveal dramatic effect modification, with gastrointestinal SAEs predominating across all age strata.

3.4 Indication-Based Discovery: The Key Finding

Analysis of drug indication fields revealed heterogeneous prescribing patterns:

Exploratory SAE Mining by Indication:

When examining the top SAEs within diabetes versus obesity subgroups, a striking pattern emerged:

Diabetes (T2DM) Patients - Top SAEs:

• CHOLELITHIASIS: 14
• HYPOGLYCAEMIA: 13
• WEIGHT DECREASED: 13
• DIZZINESS: 11
• MUSCULAR WEAKNESS: 11

Obesity/Weight Management Patients - Top SAEs:

• CHOLELITHIASIS: 18
• CHOLECYSTITIS ACUTE: 16
• CHOLECYSTITIS: 11
• BILE DUCT STONE: 8
• ACUTE KIDNEY INJURY: 7

Discovery Insight: Cholelithiasis (gallstones) and acute cholecystitis appeared disproportionately represented in obesity patients relative to diabetes patients, despite the diabetes subgroup being larger overall. This suggested potential effect modification by indication.

3.5 Discovery Selection Matrix

Selection Rationale: Cholelithiasis was selected as the primary endpoint for detailed analysis because: (1) it represents a well-documented but incompletely characterized GLP-1 class effect; (2) it showed substantial absolute risk elevation in both subgroups; (3) the ROR ratio indicated clinically meaningful heterogeneity; and (4) the finding has direct clinical implications for patient monitoring.

3.6 Stratified Disproportionality Analysis: Cholelithiasis

Overall Analysis (All Indications Combined)

Signal Metrics:

• ROR: 3.71 (95% CI: 2.85-4.83)
• PRR: 3.68
• IC: 1.89 (IC₀₂₅: 1.45)
• χ² (Yates): 142.7, p < 0.0001

Interpretation: Strong positive signal for cholelithiasis with semaglutide in aggregate analysis.

Stratum-Specific Analysis: Type 2 Diabetes Mellitus

Signal Metrics (T2DM Stratum):

• ROR: 3.21 (95% CI: 1.85-5.57)
• PRR: 3.18
• IC: 1.68 (IC₀₂₅: 0.97)
• χ² (Yates): 18.4, p < 0.0001

Stratum-Specific Analysis: Obesity/Weight Management

Signal Metrics (Obesity Stratum):

• ROR: 8.42 (95% CI: 5.18-13.67)
• PRR: 8.21
• IC: 3.07 (IC₀₂₅: 2.15)
• χ² (Yates): 87.3, p < 0.0001

3.7 Stratified Disproportionality Analysis: Acute Cholecystitis

Overall Analysis

Signal Metrics:

• ROR: 4.09 (95% CI: 2.87-5.83)
• PRR: 4.05
• IC: 2.03 (IC₀₂₅: 1.42)
• χ² (Yates): 98.2, p < 0.0001

Stratum-Specific: Type 2 Diabetes Mellitus

Signal Metrics (T2DM):

• ROR: 2.89 (95% CI: 1.45-5.76)
• PRR: 2.85
• IC: 1.52 (IC₀₂₅: 0.53)
• χ²: 9.8, p = 0.002

Stratum-Specific: Obesity/Weight Management

Signal Metrics (Obesity):

• ROR: 12.35 (95% CI: 7.52-20.28)
• PRR: 12.01
• IC: 3.62 (IC₀₂₅: 2.58)
• χ²: 72.4, p < 0.0001

3.8 Formal Effect Modification Test

Cholelithiasis - Interaction Test:

$$z = \frac{\ln(8.42) - \ln(3.21)}{\sqrt{0.237^2 + 0.274^2}} = \frac{0.963}{0.362} = 2.66$$

Two-tailed p-value for interaction: 0.008

Acute Cholecystitis - Interaction Test:

$$z = \frac{\ln(12.35) - \ln(2.89)}{\sqrt{0.245^2 + 0.352^2}} = \frac{1.456}{0.429} = 3.39$$

Two-tailed p-value for interaction: 0.0007

Interpretation: Both cholelithiasis and acute cholecystitis demonstrate statistically significant effect modification by indication. The association between semaglutide and gallstone-related adverse events is significantly stronger in obesity patients compared to T2DM patients.

3.9 Summary of Key Statistical Findings

Interaction p-values: Cholelithiasis: 0.008; Acute Cholecystitis: 0.0007

4. Discussion

4.1 Principal Findings

This data-driven pharmacovigilance study systematically explored subgroup heterogeneity in semaglutide-associated serious adverse events using FAERS data. Our automated discovery algorithm identified indication for prescribing (Type 2 Diabetes vs. Obesity/Weight Management) as the most clinically meaningful effect modifier, surpassing age-based stratification.

The key finding is that cholelithiasis and acute cholecystitis demonstrate significantly elevated reporting odds in obesity patients compared to diabetes patients. Specifically:

1. Cholelithiasis: ROR of 8.42 (95% CI: 5.18-13.67) in obesity patients versus 3.21 (95% CI: 1.85-5.57) in T2DM patients (interaction p=0.008)

2. Acute Cholecystitis: ROR of 12.35 (95% CI: 7.52-20.28) in obesity patients versus 2.89 (95% CI: 1.45-5.76) in T2DM patients (interaction p=0.0007)

Both findings exceed traditional pharmacovigilance signal detection thresholds (ROR lower CI bound >1, IC₀₂₅ >0) and demonstrate highly statistically significant associations.

4.2 Biological Plausibility

The observed indication-specific disparity in gallstone risk is biologically plausible and consistent with known pathophysiology:

Rapid Weight Loss Hypothesis: Obesity patients treated with semaglutide often experience substantial and relatively rapid weight reduction (15-20% body weight in clinical trials). Rapid weight loss is a well-established risk factor for cholesterol gallstone formation due to:

• Increased biliary cholesterol secretion
• Reduced gallbladder motility and stasis
• Altered bile acid composition

In contrast, T2DM patients may experience more modest weight loss and often have concurrent medications (e.g., SGLT2 inhibitors, insulin) that modulate metabolic responses.

Baseline Risk Differences: Paradoxically, while obesity itself is a risk factor for cholelithiasis, T2DM patients in real-world cohorts are often older, have longer-standing metabolic dysfunction, and may already be on ursodeoxycholic acid or other hepatoprotective agents. Additionally, T2DM patients undergo more frequent laboratory monitoring, potentially leading to earlier detection and management of biliary symptoms before they become "serious."

GLP-1 Receptor Effects: GLP-1 receptors are expressed in the gallbladder epithelium. Direct effects on gallbladder smooth muscle relaxation and delayed emptying have been demonstrated in animal models. Whether this effect is dose-dependent (higher doses in obesity treatment) or modified by baseline glycemic status requires investigation.

4.3 Comparison with Previous Literature

Our findings align with and extend several published observations:

1. STEP Trial Program: In semaglutide obesity trials (STEP 1-5), cholelithiasis was reported in 2-4% of treated patients versus 1% in placebo groups. Our real-world ROR of 8.42 suggests the signal may be even stronger in routine practice.

2. SUSTAIN Trials: Diabetes trials showed lower rates of gallstone events (~1-2%), consistent with our lower ROR in the T2DM stratum.

3. FDA Safety Communications: The FDA has issued warnings about GLP-1-associated gallbladder disease, but indication-specific risks have not been formally quantified.

4. Meta-analyses: Recent systematic reviews have noted heterogeneity in gallstone risk across GLP-1 trials but lacked power for formal subgroup comparisons.

4.4 Clinical Implications

For Obesity Patients:

• Baseline abdominal ultrasound may be considered in high-risk patients (female, >40 years, family history)
• Patient education about biliary colic symptoms (right upper quadrant pain, nausea after fatty meals)
• Consider slower dose titration to mitigate rapid weight loss
• Low threshold for hepatobiliary imaging in symptomatic patients

For T2DM Patients:

• Standard monitoring appears adequate given lower relative risk
• Maintain awareness of gallstone symptoms, especially with concurrent weight loss

For Prescribers:

• Discuss gallbladder risk during informed consent, particularly for obesity indications
• Document baseline biliary history
• Consider prophylactic ursodeoxycholic acid in very high-risk patients (controversial, requires individualized decision-making)

4.5 Strengths of This Study

1. Data-Driven Approach: Unlike hypothesis-driven analyses that pre-specify subgroups, our exploratory algorithm objectively identified the most divergent covariate without researcher bias.

2. Rigorous Statistics: We employed multiple complementary signal detection metrics (frequentist ROR/PRR and Bayesian IC) with appropriate confidence intervals and formal interaction testing.

3. Real-World Evidence: FAERS captures diverse patient populations and prescribing patterns not represented in restrictive clinical trials.

4. Transparency: All analytical code and decision rules are reproducible.

4.6 Limitations

1. Spontaneous Reporting Bias: FAERS is subject to under-reporting, differential reporting by specialty (endocrinologists vs. obesity medicine), and media-stimulated reporting bursts.

2. Missing Data: Despite meeting our <40% threshold, 15.5% of reports lacked clear indication data. We could not reliably analyze sex-specific risks due to >50% missingness.

3. Confounding: We could not adjust for important confounders such as:

   • Concomitant medications (statins, fibrates, estrogen therapy)
   • Baseline BMI and weight loss velocity
   • Dietary factors
   • Prior cholecystectomy status
   • Duration of semaglutide exposure

4. Cannot Establish Causality: Disproportionality analysis identifies statistical associations, not causal relationships. The observed signals require confirmation in cohort studies with individual-level data.

5. Denominator Uncertainty: FAERS lacks true exposure denominators (total patients treated). We used all other drugs as the comparator, which may dilute signal strength.

6. Coding Variability: MedDRA terms may be inconsistently applied (e.g., "cholelithiasis" vs. "gallbladder disorder" vs. "cholecystitis").

4.7 Future Research Directions

1. Target Cohort Studies: Retrospective analysis of electronic health records or claims databases with confirmed exposure and outcome data.

2. Mechanistic Studies: Investigation of GLP-1 receptor effects on gallbladder motility and bile composition in humans.

3. Risk Prediction Modeling: Development of clinical tools to identify obesity patients at highest risk for semaglutide-associated cholelithiasis.

4. Comparative Effectiveness: Head-to-head comparison of gallstone risk across GLP-1 agonists (semaglutide vs. liraglutide vs. tirzepatide).

5. Pharmacogenomics: Exploration of genetic variants in bile acid transporters or GLP-1 receptors that may modify individual susceptibility.

5. Conclusion

This data-driven pharmacovigilance analysis of FAERS reveals that the risk of gallstone-related serious adverse events with semaglutide is significantly modified by treatment indication. Patients receiving semaglutide for obesity management demonstrate approximately 2.6-fold higher reporting odds for cholelithiasis and 4.3-fold higher odds for acute cholecystitis compared to patients treated for Type 2 Diabetes.

These findings have immediate clinical relevance:

• Obesity patients initiating semaglutide warrant enhanced counseling about biliary symptoms and consideration of baseline risk stratification
• Clinicians should maintain heightened vigilance for right upper quadrant symptoms, particularly during periods of rapid weight loss
• Regulatory authorities may consider indication-specific language in product labeling

While spontaneous reporting data cannot establish causality, the consistency of our findings with clinical trial observations and established pathophysiology strengthens confidence in the signal. Prospective cohort studies with individual-level data are needed to quantify absolute risks and validate these disproportionality findings.

As semaglutide and related incretin therapies continue to expand into new indications and patient populations, data-driven pharmacovigilance approaches will be essential for identifying subgroup-specific safety signals that might otherwise remain obscured in aggregate analyses.

References

1. Wilding JPH, Batterham RL, Calanna S, et al. Once-Weekly Semaglutide in Adults with Overweight or Obesity. N Engl J Med. 2021;384(11):989-1002.

2. Marso SP, Bain SC, Consoli A, et al. Semaglutide and Cardiovascular Outcomes in Patients with Type 2 Diabetes. N Engl J Med. 2016;375(19):1834-1844.

3. Davies M, Faerch L, Jeppesen OK, et al. Semaglutide 2·4 mg once a week in adults with overweight or obesity, and type 2 diabetes (STEP 2): a randomised, double-blind, double-dummy, placebo-controlled, phase 3 trial. Lancet. 2021;397(10278):971-984.

4. Taubes G. As weight-loss drugs surge in popularity, researchers worry about long-term effects. Science. 2023;380(6652):1346-1347.

5. Bate A, Evans SJW. Quantitative signal detection using spontaneous ADR reporting. Pharmacoepidemiol Drug Saf. 2009;18(6):427-436.

6. Zorych I, Madigan D, Ryan P, et al. Disproportionality methods for pharmacovigilance: an experimental evaluation. Drug Saf. 2013;36(1):39-48.

7. U.S. Food and Drug Administration. FDA Adverse Event Reporting System (FAERS) Public Dashboard. Available at: https://fis.fda.gov/sense/app/d10be6bb-494e-4cd2-82e4-0135608ddc13

8. OpenFDA. Drug Event API Documentation. Available at: https://open.fda.gov/apis/drug/event/

9. Norén GN, Hopstadius J, Bate A, Edwards IR. Information component measures in the light of gold standards. Pharmacoepidemiol Drug Saf. 2010;19(S2):S1-S292.

10. Shuster JJ, Jones LS, Salmon DA. Fixed vs random effects meta-analysis in rare event studies: the rosiglitazone link with myocardial infarction and cardiac death. Stat Med. 2007;26(24):4375-4385.

11. American Gastroenterological Association. AGA Clinical Practice Update on Management of Gallstones. Gastroenterology. 2023.

12. Porter KM, O'Connell TX. Gallstone disease associated with rapid weight loss. Am Surg. 1993;59(11):744-746.

Acknowledgments

This analysis was conducted using publicly available data from the FDA Adverse Event Reporting System (FAERS) via the OpenFDA API. The authors acknowledge the FDA's commitment to transparency in making these data accessible for independent research.

Funding

No external funding was received for this study.

Conflicts of Interest

The authors declare no conflicts of interest related to this work.

Data Availability

All data used in this study are publicly available through the OpenFDA API (https://open.fda.gov). Analytical code is available upon reasonable request from the corresponding author.