Document resource
WHAT IS ALREADY KNOWN ON THIS TOPIC The U.S. Food and Drug Administration investigated the potential risk of suicidal ideation and behaviors for glucagon-like peptide 1 receptor agonists (GLP-1 RAs) on learning of post-marketing reports of suicidal ideation and behaviors in patients taking these products.WHAT THIS STUDY ADDS GLP-1 RA use was not associated with increased risk for intentional self-harm compared with sodium-glucose cotransporter-2 inhibitors or dipeptidyl peptidase-4 inhibitors. Subgroup analyses by comorbid obesity and diabetes, psychiatric history, prior intentional self-harm, age, and sex yielded similar results as the primary analyses.HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE, OR POLICY Use of GLP-1 RAs did not show increased risk of intentional self-harm, which provides some reassurance regarding the safety of these products.Glucagon-like peptide 1 receptor agonists (GLP-1 RAs) have been approved in the United States since 2005 to improve glycemic control in patients with type 2 diabetes and since 2014 to reduce excess body weight and maintain weight reduction long-term in patients overweight or with obesity. Because suicidal ideation and behaviors (SI/B) had been observed with certain central-acting drugs evaluated or approved for weight loss,1 since 2012, all centrally acting drugs with a long-term weight reduction indication approved in the United States have carried information about the potential risk of SI/B in the warnings and precautions section of the prescribing information, including GLP-1 RAs. However, the same GLP-1 RA products approved for indications in type 2 diabetes do not carry an SI/B warning.In July 2023, the U.S. Food and Drug Administration (FDA) began investigating the potential risk of SI/B for the GLP-1 RA products on learning of post-marketing reports of SI/B in patients taking GLP-1 RAs. Preliminary review of clinical trial and post-marketing data did not support an association between GLP-1 RA use and SI/B.2 Because of the small number of SI/B events in clinical trials and confounding in post-marketing data, FDA conducted a subject-level meta-analysis of clinical trials across GLP-1 RA development programs (Tran et al, submitted for publication) and an epidemiological study using claims data in the U.S. Sentinel System, which is the focus of this paper. Our objective was to compare the risk of intentional self-harm (an SI/B related outcome that can be measured in health administrative claims) with GLP-1 RA use compared with sodium-glucose cotransporter-2 inhibitors (SGLT-2is), or separately compared with dipeptidyl peptidase-4 inhibitors (DPP-4is), in a large population of commercially-insured and privately-insured individuals with type 2 diabetes in the United States.Research design and methods Data source The Sentinel System is a national, multisite, distributed data network established under the Sentinel Initiative, 3 4 which currently includes approximately 138.7 million active health plan members with over 1.5 billion person-years of longitudinal observation time. For this analysis, we included administrative claims data from October 1, 2015, to September 30, 2023, from ten Sentinel Data Partners, including large national insurers, integrated delivery systems, fee-for-service Medicare, and Medicaid. This study was conducted as part of the Sentinel surveillance activities under the auspices of the FDA and, therefore, was not under the purview of institutional review boards.Study population We conducted an active-comparator, new-user cohort study. The study population included patients with a diagnosis of type 2 diabetes newly initiating GLP-1 RAs (primary exposure), SGLT-2is (primary comparator), or DPP-4is (secondary comparator) who were at least 18 years of age and continuously enrolled in a health plan with medical and drug coverage for at least 183 days. To accommodate administrative gaps in enrollment, we included patients with up to a 45-day gap in between two consecutive enrollment periods into the study. SGLT-2is were chosen as the primary comparator as they have no known association with intentional self-harm and, like GLP-1 RAs, are used in patients with type 2 diabetes as second-line or third-line glucose-lowering drugs. A secondary comparison to DPP-4is, used for hyperglycemic control in patients with type 2 diabetes, was also conducted.All known single and combination products including GLP-1 RAs (albiglutide, dulaglutide, exenatide, liraglutide, lixisenatide, semaglutide, and the dual GLP-1/gastric inhibitory peptide RA tirzepatide), SGLT-2is (bexagliflozin, canagliflozin, dapagliflozin, empagliflozin, ertugliflozin, and the dual SGLT-1/SGLT-2 inhibitor sotagliflozin), and DPP-4is (alogliptin, linagliptin, saxagliptin, and sitagliptin) available in the United States during the study period were included in the analysis. New initiators were defined as patients with no evidence of GLP-1 RA, SGLT-2i, or DPP-4i dispensing in the 183 days prior to the first prescription fill (index date).Type 2 diabetes was defined as at least one International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) diabetes diagnosis code in any care setting and one dispensing for an oral antidiabetic agent in the 183 days prior to and including the index date. Exclusion criteria included evidence of other types of diabetes (ie, with diagnosis claims for type 1 diabetes, gestational diabetes, secondary diabetes, insulin use with no documented dispensing for other antidiabetic agents), pregnancy, or prior exposure to GLP-1 RAs, SGLT-2is, or DPP-4is (online supplemental table S1).SP110.1136/bmjdrc-2025-005840.supp1Supplementary dataStudy outcome and follow-up The primary outcome was intentional self-harm. We chose to evaluate the outcome of intentional self-harm rather than SI/B because of the known inadequacies of claims data to capture suicidal intent. 5 To identify intentional self-harm, we used an adapted validated algorithm6 using ICD-10-CM diagnosis codes for suicide attempts (T14.91) and intentional self-harm (select diagnosis codes within the following ranges: T36-T71 and X71-X83), which included intentional poisonings, toxic effects, asphyxiation events, drowning events, or self-harm events by other specified means (eg, by discharging gun, crashing of vehicle, jumping in front of moving object).5A patient’s follow-up time was based on the length of exposure episodes, created using the number of days of product supplied per dispensing in outpatient pharmacy dispensing data. We bridged together exposure episodes less than 30 days apart and added 30 days at the end of each exposure episode to create continuous treatment episodes. Follow-up started on the day after the patient’s first qualifying prescription fill for a GLP-1 RA or comparator drug until the first occurrence of intentional self-harm or a censoring event.Censoring events included hospitalization, treatment discontinuation, death, end date of available data from each Data Partner, health plan disenrollment, or switch to the comparator product. Censoring at hospital admission was necessary because we lacked information on inpatient drug use, but intentional self-harm events occurring on the day of hospital admission were included as outcomes. Treatment discontinuation was defined as no prescription in the 30 days following the last dispensing date plus the days of product supplied. Censoring events were defined using National Drug Codes.Potential confounders Potential confounders that could influence the choice of anti-diabetic medication and risk of intentional self-harm events included demographic and medical characteristics. Demographic characteristics (ie, age, sex, and calendar year of index medication) were captured on the index date. Medical characteristics (ie, general and diabetes composite measures of health, psychiatric history, obesity diagnosis or treatment, relevant medical product use, select chronic conditions, and health service utilization metrics) were captured using both ICD-9-CM and ICD-10-CM diagnosis codes in the 183 days prior to and including the index date ( online supplemental table S1).Statistical analysis We estimated propensity scores (PS) using a logistic regression model controlling for potential confounders discussed above. After trimming of patients in non-overlapping regions of the PS distributions to achieve better balance on baseline characteristics between the exposure groups, 7 we used the PS to calculate stabilized inverse probability treatment weights (IPTW) for the estimation of the average treatment effect.8 We applied a 1% truncation to handle extreme IPTW, that is, weights >99th (and <1st) percentile were replaced with values at the 99th (and 1st) percentile; sensitivity analyses applied 0% and 2.5% truncation thresholds to assess the impact of extreme weights on effect estimates. We used the absolute standardized difference9 to compare baseline variables between the exposure groups, with differences of <0.10 indicating balance, before and after applying IPTW.Weighted and unweighted (with adjustment for Data Partner site only) incidence rate (IR) of intentional self-harm events per 1000 person-years accounted for differing follow-up times. Case-centered logistic regression10 estimated hazard ratio (HR) and 95% confidence interval (CI)11 for intentional self-harm events in patients on GLP-1 RAs compared with patients on SGLT-2is. Analysis methods were repeated for the comparison of GLP-1 RAs versus DPP-4is. The assumption of proportional hazards was assessed using log-log plots and plots of Schoenfeld residuals.12 Subgroup analyses were conducted by sex, age group (18–24 years, 25–44 years, 45–64 years, 65–79 years, and ≥80 years), comorbid obesity and diabetes, psychiatric history, and prior intentional self-harm. Sensitivity analyses included defining the intentional self-harm outcome as diagnosis codes recorded only in the inpatient or emergency department setting, limiting the allowable exposure episode gap and extension to 15 days, and requiring two or more dispensings of the study drug.Patients missing data on race or ethnicity were included in the analyses, and they were assigned a value of ‘unknown’; those with a missing value for sex were excluded. All other variables were operationalized from health insurance claims data and were considered complete, as data were collected as part of routine billing for services administered.Prior to conducting our analysis, we estimated that 945 intentional self-harm events were needed to detect an HR=1.2 with 80% statistical power, assuming 5% type I error, 1:1 GLP-1 RA to SGLT-2i sample size ratio, and a null HR=1.0. The study design and analysis methods were specified prior to analyzing the outcome data, and analyses used two-sided tests. There was no multiplicity adjustment. All analyses were performed using SAS software V.9.4 (SAS Institute, Cary, North Carolina, USA).Data and resource availability The data that support the findings of this study are available from CVS Health (Aetna), Carelon Research/Elevance Health, Duke University School of Medicine, Department of Population Health Sciences through the Centers for Medicare and Medicaid Services which provided data, HealthPartners Institute, Humana Healthcare Research, Kaiser Permanente Colorado Institute for Health Research, Kaiser Permanente Mid-Atlantic Permanente Medical Group, PC, Mid-Atlantic Permanente Research Institute, Marshfield Clinic Research Institute, and OptumInsight Life Sciences, but restrictions apply to the availability of these data, which were used under license for the current study and therefore are not publicly available.Results The study population included 1 161 983 GLP-1 RA users and 1 081 155 SGLT-2i users with type 2 diabetes. Before IPTW, the GLP-1 RA group was younger, had a higher percentage of females, and was more likely to have a psychiatric or obesity history compared with the SGLT-2i group. Medication use metrics (ie, mean number of filled prescriptions, mean number of generic drugs dispensed, mean number of unique drug classes dispensed) were also higher among GLP-1 RA users compared with SGLT-2i users. After controlling for confounders using IPTW, the baseline characteristics of the two groups (1 154 786 GLP-1 RA users and 1 118 691 SGLT-2i users) were balanced with no notable differences ( figure 1, table 1). With IPTW, the cohort had a mean age that was slightly above 60 years (61.6 years in the GLP-1 RA group and 62.4 years in the SGLT-2i group), similar percentages of females (51.1% of GLP-1 RA users and 49.9% of SGLT-2i users) to males, and with the majority presenting with a psychiatric history (61.1% of GLP-1 RA users and 59.7% of SGLT-2i users).Figure 1Covariate balance: unweighted and weighted potential confounders, GLP-1 RA versus SGLT-2i. GLP-1 RA, glucagon-like peptide-1 receptor agonist; SGLT-2i, sodium-glucose cotransporter-2 inhibitor.Table 1Weighted Baseline Characteristics, GLP-1 RA versus SGLT-2i*GLP-1 RASGLT-2iPatient Characteristicsn=1 154 786n=1 118 691Age, mean (SD), years61.6 (10.3)62.4 (10.7)Age group, No. (%), years 18–244638 (0.4)4378 (0.4) 25–44123 031 (10.7)112 473 (10.1) 45–64492 529 (42.7)453 322 (40.5) 65–79480 868 (41.6)478 848 (42.8) ≥8053 721 (4.7)69 670 (6.2)Sex, No. (%) Female589 853 (51.1)558 058 (49.9) Male564 932 (48.9)560 633 (50.1)Race, No. (%) American Indian or Alaska Native8291 (0.7)6747 (0.6) Asian18 497 (1.6)33 426 (3.0) Black or African American109 076 (9.4)114 761 (10.3) Multiracial5987 (0.5)5347 (0.5) Native Hawaiian or Other Pacific Islander3255 (0.3)3344 (0.3) White582 796 (50.5)565 850 (50.6) Unknown426 884 (37.0)389 218 (34.8)Hispanic origin, No. (%) Yes52 382 (4.5)68 494 (6.1) No652 618 (56.5)658 580 (58.9) Unknown449 786 (38.9)391 617 (35.0)Medical history, No. (%) Obesity500 059 (43.3)460 838 (41.2) Weight-related intervention816 (0.1)834 (0.1) Lifestyle intervention204 401 (17.7)186 328 (16.7) Cardiovascular disease243 859 (21.1)242 374 (21.7) Cognitive impairment26 503 (2.3)25 974 (2.3) Traumatic brain injury2486 (0.2)2455 (0.2) Migraine43 603 (3.8)39 770 (3.6) Epilepsy15 005 (1.3)14 233 (1.3) Chronic pain246 286 (21.3)230 378 (20.6)Psychiatric history, No. (%) Any706 011 (61.1)668 215 (59.7) Psychiatric drug dispensing517 241 (44.8)495 488 (44.3) Prior intentional self-harm encounter1201 (0.1)1060 (0.1) Psychiatric diagnosis578 447 (50.1)540 035 (48.3)Medical product use, No. (%) Insulin419, 616 (36.3)439 493 (39.3) Metformin1 023 028 (88.6)988 826 (88.4) Alpha-glucosidase inhibitors4812 (0.4)4941 (0.4) Thiazolidinediones93 706 (8.1)95 314 (8.5) Meglitinides12 732 (1.1)13 303 (1.2) Amylin analogs146 (0.0)179 (0.0) Sulfonylureas428 853 (37.1)432 254 (38.6) Bile acid sequestrants5159 (0.4)4685 (0.4) Analgesic opiates310 342 (26.9)287 908 (25.7) Antipsychotics74 028 (6.4)79 285 (7.1) Anxiolytics/hypnotics215 275 (18.6)214 861 (19.2) Lithium and other mood stabilizers62 919 (5.4)67 618 (6.0) Antidepressants420 861 (36.4)396 405 (35.4)Health characteristics, mean (SD) Combined comorbidity score1.8 (2.3)1.9 (2.4) Adapted Diabetes Complications Severity Index1.5 (1.7)1.5 (1.7)Health service utilization, mean (SD) Filled prescriptions27.5 (21.4)26.7 (19.9) Generic drugs dispensed11.5 (5.7)11.2 (5.2) Unique drug classes dispensed9.8 (4.8)9.7 (4.9) Ambulatory visits11.3 (10.6)11.1 (10.6) Emergency room visits0.4 (1.2)0.4 (1.1) Inpatient visits0.1 (0.5)0.1 (0.5)*The weighted (inverse probability treatment weights—1% truncation) counts do not represent the actual number of patients, but the sum of patients’ weights.GLP-1 RA, glucagon-like peptide-1 receptor agonist; SGLT-2i, sodium-glucose cotransporter-2 inhibitor.; GLP-1 RA users contributed an average of 0.59 person-year at risk and SGLT-2i users contributed an average of 0.65 person-year at risk. There were 840 intentional self-harm events in the GLP-1 RA group and 609 intentional self-harm events in the SGLT-2i group, corresponding to unweighted IRs of 1.22 events per 1000 person-years in the GLP-1 RA group and 0.87 events per 1000 person-years in the SGLT-2i group. Before confounder control, the unweighted HR comparing GLP-1 RA to SGLT-2i was 1.38 (95% CI 1.25 to 1.54). Weighted IRs were 1.13 events per 1000 person-years in the GLP-1 RA group and 1.22 events per 1000 person-years in the SGLT-2i group (figure 2). The HR for intentional self-harm events comparing GLP-1 RA to SGLT-2i was 0.93 (95% CI 0.81 to 1.08).Figure 2Hazard Ratio: intentional self-harm, GLP-1 RA versus SGLT-2i. IPTW adjusted analysis included 1% truncation and average treatment effect stabilization. a Incidence rate: number of events per 1000 person-years; hazard ratio estimates plotted on the natural log scale. b Analyses were conducted in a subset of data where Sentinel Data Partners had at least one event in each subgroup. ED, emergency department; GLP-1 RA, glucagon-like peptide-1 receptor agonist; IP, inpatient; IPTW, inverse probability of treatment weight; SGLT-2i, sodium-glucose cotransporter-2 inhibitor.Subgroup HR estimates comparing GLP-1 RA users to SGLT-2i users ranged from 0.64 to 1.10 with 95% CIs that included the null of 1.0 (figure 2). For both GLP-1 RA and SGLT-2i users: (1) those with psychiatric history showed a higher IR compared with those with no psychiatric history; (2) those with prior intentional self-harm events showed a higher IR compared with those with no prior intentional self-harm; (3) females showed a higher IR compared with males, and (4) younger age generally showed a higher IR. Sensitivity analyses with varying IPTW truncation thresholds, outcome definitions, and exposure definitions resulted in HRs close to 1.0 with 95% CI that included 1.0 (figure 2).For the GLP-1 RA versus DPP-4i comparison, the study population (unweighted) included 1 150 636 GLP-1 RA users and 1 396 382 DPP-4i users with type 2 diabetes. Weighted baseline characteristics were generally balanced (online supplemental table S2), (online supplemental figure S5). With IPTW, the cohort had a mean age that was slightly above 60 years (61.9 years in the GLP-1 RA group and 63.3 years in the DPP-4i group), similar percentages of females (53.7% of GLP-1 RA users and 54.0% of DPP-4i users) to males, and with the majority presenting with a psychiatric history (59.7% of GLP-1 RA users and 57.9% of DPP-4i users). There were 826 intentional self-harm events in the GLP-1 RA group and 1088 intentional self-harm events in the DPP-4i group. GLP-1 RA users contributed an average of 0.59 person-year at risk and DPP-4i users contributed an average of 0.73 person-year at risk. Weighted IRs were 1.12 events per 1000 person-years in the GLP-1 RA group and 1.37 events per 1000 person-years in the DPP-4i group (figure 3). The HR for intentional self-harm events comparing GLP-1 RA to DPP-4i was 0.94 (95% CI 0.82 to 1.07). Subgroup and sensitivity analyses of GLP-1 RA versus DPP-4i resulted in HRs that were either close to 1.0 or had a 95% CI that included 1.0 (figure 3). Inspection of diagnostics did not indicate gross departures from the proportional hazards assumption in either the comparison of GLP-1 RA versus SGLT-2i or the comparison of GLP-1 RA versus DPP-4i.Figure 3Hazard Ratio: intentional self-harm, GLP-1 RA versus DPP-4i. IPTW adjusted analysis included 1% truncation and average treatment effect stabilization. a Incidence rate: number of events per 1000 person-years; hazard ratio estimates plotted on the natural log scale. b Analyses conducted in a subset of data where Sentinel Data Partners had at least one event in each subgroup. DPP-4i, dipeptidyl peptidase-4 inhibitors; GLP-1 RA, glucagon-like peptide-1 receptor agonist; IPTW, inverse probability of treatment weight.Discussion In this analysis involving a large sample of publicly and privately insured type 2 diabetes patients in the United States, GLP-1 RA use did not show an increased risk of intentional self-harm compared with SGLT-2i use among patients with type 2 diabetes (1.13 events per 1000 person-years in the GLP-1 RA group; 1.22 events per 1000 person-years in the SGLT-2i group; HR 0.93; 95% CI 0.81 to 1.08). Analyses in subgroups by age, sex, comorbid obesity and diabetes, psychiatric history, or prior intentional self-harm and the comparison of GLP-1 RAs versus DPP-4is yielded similar results.We found a markedly higher IR of self-harm events among GLP-1 RA, SGLT-2i, and DPP-4i users with a history of intentional self-harm and prior psychiatric history compared with individuals without that respective history. This finding is consistent with previous longitudinal studies that found prior intentional self-harm events and SI/B events to be a robust predictor of future intentional self-harm or SI/B events.13 These findings highlight the importance of psychiatric history evaluations for patients with type 2 diabetes.Our findings in the type 2 diabetes population are consistent with several prior observational studies assessing the potential association between GLP-1 RA use and different outcomes related to suicidal ideation and behavior.14–18 For example, Shapiro et al’s active-comparator, new-user cohort study using primary practice, hospital, and national death registration data from the United Kingdom found no evidence of increased risk of suicidal ideation, self-harm, or suicide in patients with type 2 diabetes comparing GLP-1 RAs to SGLT-2is (HR 0.91; 95% CI 0.73 to 1.12) or DPP-4is (HR 1.02; 95% CI 0.85 to 1.23).15 Another active comparator, new user study comparing suicide death between GLP-1 RA and SGLT-2i using health and administrative registers in Sweden and Denmark found no definitive evidence of increased risk of suicide death (HR 1.25; 95% CI 0.83 to 1.88).18 In addition, a PS-matched study of Medicare insured adults ≥66 years old in the United States by Tang et al found no increased risk of suicidal ideation and behaviors among users of GLP-1 RAs compared with SGLT-2is (HR 1.07; 95% CI 0.80 to 1.45, rate difference 0.16; 95% CI −0.53 to 0.86 events per 1000 person-years) or compared with DPP-4is (HR 0.94; 95% CI 0.71 to 1.24, rate difference −0.18; 95% CI −0.92 to 0.57 events per 1000 person-years).16Our study had several strengths. We included a large number of patients from commercial health insurers, Medicare, Medicaid, and integrated delivery systems in the United States. Adjusted analyses with a comprehensive set of potential confounders helped minimize bias from confounding and resulted in a weighted HR that did not show an increased risk compared with the unweighted analysis (0.93 vs 1.38). We evaluated two comparators, SGLT-2is and DPP-4is, to assess the risk of self-harm; both comparisons were consistent in their results. The study also included several subgroup analyses and sensitivity analyses to evaluate the robustness of the primary analysis.There are also notable limitations. We lacked data to assess the risk with extended durations of use. Our study population was limited to patients with health insurance and did not include patients with type 1 diabetes. Unmeasured confounders, such as socioeconomic factors or educational level, could bias results. We aimed to account for socioeconomic factors that could influence healthcare resource access by controlling for multiple healthcare utilization metrics (ie, mean filled prescriptions, generic drugs dispensed, unique drugs dispensed, ambulatory encounters, emergency room visits, and inpatient visits). The healthcare claims data provided incomplete data on race and ethnicity, and there were no data on patients’ educational level. Nonetheless, with both the SGLT-2i and DPP-4i user reference groups, weighting resolved the association seen in the unweighted analyses. This suggests that weighting ameliorated the confounding present in the unweighted results, though we cannot exclude the possibility of some residual confounding remaining. Misclassification of intentional self-harm events is also a concern; events occurring outside of the healthcare system (such as completed suicides that did not receive medical care) or not labeled as ‘intentional’ would not be captured by claims data. Also, we did not include suicidal ideation as an outcome because of the aforementioned limitation of claims data to capture suicidal intent. Nonetheless, we observed considerably higher incidence rates among patients with a history of psychiatric illness or intentional self-harm, suggesting that the outcome measure was successful at identifying higher-risk groups of patients.Conclusions This study did not show an increased risk of intentional self-harm comparing GLP-1 RA to SGLT-2i or DPP-4i use and provides some reassurance regarding the safety of GLP-1 RA use in patients with type 2 diabetes.