What this is
- This study assesses the risk of suicidal ideation and self-injury (SIS) in individuals with obesity and type 2 diabetes using glucagon-like peptide-1 receptor agonists (GLP-1RA) compared to sodium-glucose cotransporter 2 inhibitors (SGLT-2i).
- It utilizes a population-based cohort from the Valencia Health System, covering over five million residents.
- The study employs propensity score weighting to balance treatment groups and aims to provide insights into the safety of GLP-1RA regarding mental health outcomes.
Essence
- No increased risk of suicidal ideation or self-injury was found in patients using GLP-1RA compared to those using SGLT-2i. The findings suggest that GLP-1RA does not contribute to SIS in this population.
Key takeaways
- 3040 patients initiated treatment with GLP-1RA, while 11,627 started SGLT-2i. The GLP-1RA group was younger and had higher rates of anxiety, sleep disorders, and depression.
- The main analysis showed a hazard ratio (HR) of 1.04 for SIS in GLP-1RA users vs. SGLT-2i users, indicating no evidence of increased SIS risk. Intention-to-treat analysis yielded an HR of 1.36.
- The rarity of SIS events and wide confidence intervals indicate that while no increase in risk was found, caution is warranted as the effect could range up to threefold.
Caveats
- Data collection was based on routine clinical practice, which may introduce biases such as misclassification and under-registration of SIS events.
- The study's observational design limits causal inferences, and residual confounding cannot be ruled out despite statistical adjustments.
- The low incidence of SIS events means that the findings should be interpreted cautiously, as they do not exclude the possibility of a threefold increase in SIS risk.
AI simplified
Introduction
In the last 20 years, glucagon-like peptide-1 receptor agonists (GLP-1RA) have been used as one of the recommended pharmacotherapy options for individuals with type 2 diabetes mellitus, especially for large subgroups of patients with high cardiovascular risk and other comorbidities such as chronic kidney disease [1]. In several studies, GLP-1RA have been proved to reduce glucose levels, to ameliorate the metabolic syndrome and to provide cardiorenal benefits [2–5]. Furthermore, GLP-1RA have shown substantial weight-reduction effects [6], and in recent years several compounds have been authorised for weight management in people with overweight or obesity. Since then, GLP-1RA have received considerable media and influencer attention which, combined with intensive commercial practices, have led to a popularity storm and an explosion in the use of these drugs, resulting in global shortages of GLP-1RA drugs [7–9].
In contrast, there is a growing concern that GLP-1RA may be linked to increased risk of suicidal ideation and self-injury (SIS), following reports of the Icelandic Medicines Agency in July 2023 identifying a disproportionate number of cases of SIS in individuals taking liraglutide and semaglutide. Previously, other weight-loss drugs have been removed from the market for similar reasons [10], and regulators worldwide are currently reviewing safety data on GLP-1RA [11–13]. Although preliminary regulatory-led analyses of post-marketing data have not found conclusive evidence of a causal link [14], and in the absence of evidence from large, randomised clinical trials, observational studies adequately designed to assess the comparative safety of treatments, employing an active comparator, new user design, are warranted to evaluate the risk of SIS in individuals treated with GLP-1RA.
In this study, we aimed to assess the risk of SIS in new users of GLP-1RA to treat type 2 diabetes mellitus in individuals with obesity, when compared with new users of sodium-glucose cotransporter 2 inhibitors (SGLT-2i).
Methods
Design, setting and data sources
This is a real-world cohort study combining several population-wide databases from the Valencia Health System Integrated Database (VID). The Valencia Health System (VHS) is a comprehensive structure of hospitals, primary care facilities and other public resources managed by the government of the region of Valencia in Spain providing free, universal healthcare services (besides drug cost‐sharing) to 98% of the region's five million inhabitants. VID is a set of publicly owned, population-based healthcare, clinical and administrative electronic databases in the region that can be linked by means of a single personal identification number, and provides comprehensive information for the region’s population covered by the VHS since 2008. VID includes sociodemographic and administrative data (sex assigned at birth, age, nationality) as well as healthcare information such as inpatient and outpatient diagnoses, procedures, laboratory data, pharmaceutical prescriptions and dispensing linked at the individual level (including brand and generic name, formulation, strength and dosing schedule/regimen), hospitalisations, primary care electronic medical record data, mortality, healthcare utilisation and public health data [15].
Participants
Study flow chart
Variables
| Characteristic | Total,(%)N | SGLT-2i,(%)n | GLP-1 RA,(%)n | Stand. diff. before | Stand. diff. after | valuep |
|---|---|---|---|---|---|---|
| 14,667 | 11,627 (79.27) | 3040 (20.73) | ||||
| Sex, men | 7877 (53.71) | 6491 (55.83) | 1386 (45.59) | 0.206 | 0.006 | <0.001 |
| Income, €/year | ||||||
| <18,000 | 9489 (64.70) | 7636 (65.67) | 1853 (60.95) | −0.098 | −0.031 | <0.001 |
| 18,000–100,000 | 3054 (20.82) | 2337 (20.10) | 717 (23.59) | 0.084 | 0.022 | |
| >100,000 | 49 (0.33) | 33 (0.28) | 16 (0.53) | 0.038 | 0.008 | |
| Low resources | 1970 (13.43) | 1531 (13.17) | 439 (14.44) | 0.037 | 0.015 | |
| Not available | 105 (0.72) | 90 (0.77) | 15 (0.49) | −0.035 | 0.003 | |
| Age, years | ||||||
| 18–44 | 1787 (12.18) | 1174 (10.10) | 613 (20.16) | 0.284 | 0.004 | <0.001 |
| 45–64 | 8064 (54.98) | 6262 (53.86) | 1802 (59.28) | 0.109 | 0.007 | |
| 65–74 | 3516 (23.97) | 3006 (25.85) | 510 (16.78) | −0.223 | −0.009 | |
| ≥75 | 1300 (8.86) | 1185 (10.19) | 115 (3.78) | −0.253 | −0.003 | |
| Comorbidities | ||||||
| Heart failure | 1067 (7.27) | 908 (7.81) | 159 (5.23) | −0.105 | −0.026 | <0.001 |
| Dementia | 278 (1.90) | 234 (2.01) | 44 (1.45) | −0.043 | −0.017 | 0.05 |
| Hypertension | 10,440 (71.18) | 8376 (72.04) | 2064 (67.89) | −0.090 | −0.016 | <0.001 |
| Liver disease | 3346 (22.81) | 2626 (22.59) | 720 (23.68) | 0.026 | 0.007 | 0.207 |
| Renal disease | 805 (5.49) | 656 (5.64) | 149 (4.90) | −0.033 | 0.014 | 0.121 |
| Depression | 2950 (20.11) | 2207 (18.98) | 743 (24.44) | 0.133 | −0.003 | <0.001 |
| Coronary heart disease | 1701 (11.60) | 1465 (12.60) | 236 (7.76) | −0.160 | 0.006 | <0.001 |
| COPD | 1501 (10.23) | 1217 (10.47) | 284 (9.34) | −0.038 | 0.007 | 0.074 |
| Sleep disorders | 5276 (35.97) | 3962 (34.08) | 1314 (43.22) | 0.189 | 0.002 | <0.001 |
| Anxiety | 6324 (43.12) | 4821 (41.46) | 1503 (49.44) | 0.161 | 0.021 | <0.001 |
| Substance abuse | 221 (1.51) | 158 (1.36) | 63 (2.07) | 0.055 | −0.003 | 0.005 |
| Personality disorders | 200 (1.36) | 140 (1.20) | 60 (1.97) | 0.062 | −0.008 | 0.002 |
| Other psychiatric disorders | 598 (4.08) | 431 (3.71) | 167 (5.49) | 0.085 | −0.008 | <0.001 |
| Psychotic disorders | 634 (4.32) | 480 (4.13) | 154 (5.07) | 0.045 | −0.002 | 0.027 |
| Previous SIS | 29 (0.20) | 25 (0.22) | 4 (0.13) | −0.020 | −0.003 | 0.488 |
| Malignancies | 1503 (10.25) | 1221 (10.50) | 282 (9.28) | −0.041 | 0.016 | 0.051 |
| Lifestyle | ||||||
| Alcohol use | 656 (4.47) | 547 (4.70) | 109 (3.59) | −0.056 | 0.024 | 0.009 |
| Tobacco use | 3474 (23.69) | 2740 (23.57) | 734 (24.14) | 0.014 | 0.012 | 0.519 |
| BMI | ||||||
| 30–35 | 5148 (35.10) | 4604 (39.60) | 544 (17.89) | −0.494 | −0.007 | <0.001 |
| 35–40 | 3394 (23.14) | 2689 (23.13) | 705 (23.19) | 0.002 | 0.011 | |
| ≥40 | 2827 (19.27) | 1760 (15.14) | 1067 (35.10) | 0.473 | 0.014 | |
| Unknowna | 3298 (22.49) | 2574 (22.14) | 724 (23.82) | 0.04 | −0.017 | |
Outcomes
Primary outcomes were SIS. Outcomes were retrieved from different real-world databases in VID (primary care electronic records, in-hospital and outpatient hospital records and emergency room records), via ICD codes registered during healthcare encounters (see ESM Table). In the main per-protocol analyses, participants were followed from the index date to first mutually exclusive incidence of SIS, discontinuation of study drug (defined as more than 60 days without drug supply), initiation of other study drug, or switch to other glucose-lowering drug (defined as filling a prescription for any glucose-lowering drug other than the baseline treatment or metformin and not refilling the assigned drug for 60 days), death, or end of follow-up. In secondary, intention-to-treat analyses, participants were followed from the index date to first SIS event, death, or end of follow-up, and were analysed according to the treatment assigned at baseline, irrespective of actual exposure during follow-up. 1
Analysis
First, we described patient characteristics, overall and for GLP-1RA and SGLT-2i users (see Table 1). The p values were estimated using χ2 tests for categorical variables and t tests for continuous variables. Second, as the main analysis, we used a per-protocol approach to estimate the increased risk of SIS in GLP-1RA vs SGLT-2i users. The per-protocol approach evaluates the comparative effect of actual exposure to the treatment strategies assigned at baseline. We estimated the incidence rates by 1000 person-years of the first SIS event, overall and for each group. Second, we carried out multivariable Cox regression modelling using our set of covariates: age (as a categorical variable, reference [ref.] 18 to 44 years old), sex assigned at birth (ref. men), all comorbidities and lifestyle variables included as categorical variables (yes/no), and BMI (ref. 30–35 kg/m2). To adjust for potential confounding, we used inverse probability of treatment weighting (IPTW) based on propensity scores, a method that allows balancing without losing generalisability [17]. Propensity scores for each outcome were calculated based on the probability of initiating treatment with GLP-1RA taking into account the same covariates, to generate patient-specific stabilised weights. Covariate balance between the weighted exposure cohorts was assessed using standardised mean differences, with standardised differences <0.10 suggesting adequate balance [18]. Third, we performed several sensitivity analyses. We provided an intention-to-treat analysis, where we compared the effect of being assigned to the treatment strategies at baseline, regardless of whether or not the individuals continued to follow the strategies during follow-up. We also conducted several per-protocol stratified analyses: by the more prevalent psychiatric comorbidities in our cohort (depression, anxiety and sleep disorders); by sex; and by obesity levels; and we provided p values for interaction. Finally, to deal with participants with no BMI information at baseline, we also carried out analyses excluding individuals with no BMI information, in addition to an analysis using multivariate imputation using chained equations for BMI missing values (see ESM Table 3 for code employed for analysis). Statistical significance was defined as p<0.05. Ethics approval was obtained from the Clinic University Hospital of Valencia research ethics board (2022/164) with a waiver of informed consent. All analyses were performed using STATA version 14 and R version 3.6.0.). The investigators were granted access to the databases used to create the study population and performed several quality checks (consistency assessments, range and logic checks) before the linkage of the databases using a pseudonymised single identification number. The study was reported using the RECORD checklist for observational studies using routinely collected health data.
Results
We included 3040 individuals initiating treatment with GLP-1RA and 11,627 with SGLT-2i (Fig. 1). Mean age was 59 years, 53.7% were men, 64.7% earned less than €18,000/year, 71.2% had hypertension, 43.1% had anxiety, 36.0% had sleep disorders and 23.7% had registered tobacco use (Table 1). When compared with participants initiating treatment with SGLT-2i, those in the GLP-1RA group were younger (55 vs 60 years, p<0.001), were more likely to be women (54.4% vs 44.2%, p<0.001), had more anxiety (49.4% vs 41.5%, p<0.001), sleep disorders (43.2% vs 34.1%, p<0.001) and depression (24.4% vs 19.0%, p<0.001), and were more obese (35.1% of individuals with BMI ≥40 vs 15.1%, p<0.001). After propensity score weighting, standardised mean differences between groups were <0.1 for all covariates, showing adequate balance between groups at baseline after adjustment (Table 1).
| Person-years | Crude outcomes | Crude rate/1000 person-years | Weighted outcomes | Weighted rate/1000 person-years (95% CI) | HR (95% CI) | ||
|---|---|---|---|---|---|---|---|
| Main analysis | SGLT-2i | 12,575.93 | 17 | 1.35 | 18.11 | 1.44 (0.88, 2.52) | Ref. |
| GLP-1RA | 3785.42 | 10 | 2.64 | 6.66 | 1.76 (0.88, 4.09) | 1.04 (0.35, 3.14) | |
| Intention-to-treat analysis | SGLT-2i | 31,823.39 | 17 | 0.53 | 18.11 | 0.57 (0.35, 1.00) | Ref. |
| GLP-1RA | 7562.3 | 10 | 1.32 | 6.66 | 0.88 (0.44, 2.05) | 1.36 (0.51, 3.61) | |
| Analysis excluding patients with BMI missing data | SGLT-2i | 9966.08 | 13 | 1.3 | 14.76 | 1.48 (0.83, 2.93) | Ref. |
| GLP-1RA | 2714.9 | 7 | 2.58 | 4.54 | 1.67 (0.71, 4.95) | 0.89 (0.26, 3.14) | |
| Analysis with multiple imputation for BMI missing data | SGLT-2i | 12,575.93 | 17 | 1.35 | 18.11 | 1.44 (0.88, 2.52) | Ref. |
| GLP-1RA | 3785.42 | 10 | 2.64 | 6.66 | 1.76 (0.88, 4.09) | 1.29 (0.42, 3.92) |
Discussion
In this population-based, propensity-weighted cohort study, we found no evidence of an association between GLP-1RA and increased risk for SIS when compared with SGLT-2i, when prescribed to treat type 2 diabetes in patients with concomitant obesity. Our findings do not support an increased risk of SIS associated with GLP-1RA. All 95% CIs were compatible with no increase in risk. However, CIs of the observed associations are wide because of the small number of events. As a result of this, protective or adverse effects cannot be fully excluded. Because of the low frequency of SIS outcomes and their potential severity, and based on our estimates, we cannot rule out either a threefold increase or a decrease in SIS rates among people with type 2 diabetes and obesity treated with GLP-1RA. Large trials randomising tens of thousands of patients would be required to accurately detect an effect of GLP-1RA on SIS risk; in the absence of such trials, evidence from high quality pharmacoepidemiological studies employing population-based, observational data, such as this study, is warranted.
GLP-1RA are indicated for diabetes and obesity, which in turn are concomitant risk factors for depression and SIS. Available evidence, however, is difficult to interpret. For instance, a reciprocal association between obesity and overweight and depression has been observed [19], whereas in most studies elevated BMI is consistently associated with lower suicide completion [20]. Whether obesity is associated with suicide attempts or ideation remains unclear, as only a few, methodologically flawed studies have assessed these associations, providing heterogeneous findings, both between studies and in key risk groups [21]. Plausible mechanisms have been proposed that could explain an association of GLP-1RA with suicidality; other authors support a potential anti-depressant effect of GLP-1RA, others a lack of association. For instance, GLP-1RA have been tested in people with serious mental illness, who could be at a high suicidality risk, showing metabolic benefits with no increase in psychopathology or occurrence of SIS [22].
To date, evidence for an association between GLP-1RA and SIS is limited to a few preliminary reports from pharmacovigilance systems and pharmacoepidemiological studies. Randomised controlled trials involving GLP-1RA have not detected suicidality signals, but these studies were underpowered to detect rare events such as SIS, and they selectively exclude individuals with psychiatric comorbidity [23–26], which are prevalent in daily practice. Studies from the US Food and Drug Administration (FDA) have found a disproportionate reporting of suicidal ideation and ‘depression/suicidal’ thoughts with semaglutide and liraglutide, but not for suicidal behaviour, suicide attempts and completed suicide [27–30]. Reports from the European Medicines Agency (EMA) were due in November 2023, but the EMA asked for more data from manufacturers in December 2023, and final results are still unreleased [11, 31]. To date, only one study into this subject has been published, with more suicidal events being reported associated with semaglutide and liraglutide when compared to other GLP-1RA [32]. However, pharmacovigilance reporting in public databases is subject to many major limitations, such as heterogeneous reporting, including notoriety and ripple bias, double reporting, the lack of a population denominator or enough clinical detail to enable comparison between agents and over time [20, 22], therefore no causal association link can be inferred from such studies.
Two propensity score matched observational studies have been conducted to date, either evaluating suicidality risk in patients using semaglutide vs non-GLP-1RA glucose-lowering agents, or in those using any GLP-1RA vs users of dipeptidyl peptidase-4 (DPP-4) inhibitors. Both studies found a protective effect of GLP-1RA, which was consistent in several analyses for different indications (obesity or diabetes) and patient profiles (with or without previous depression or suicidal ideation) [33, 34]. In this way, available observational evidence to date refutes concerns of an increase in suicidality risk related to GLP-1RA treatment. However, both studies are based on a large but not representative sample of the US population, and generalisability of results is limited. Also, both provide intention-to-treat estimates only, and were unable to evaluate either medication adherence or the effect of actual exposure to GLP-1RA [33]. Our study adds to the scarce available evidence on the association between GLP-1RA and risk of SIS by providing evidence from a South European population of five million inhabitants. We employed a new user, active comparator design; we chose a suitable comparator group; we employed propensity scores based on IPTW to obtain balanced pseudo-populations at baseline; we provided both per-protocol and intention-to-treat estimates; and we performed several stratified analyses. In all analyses including stratified analyses per sex assigned at birth, we did not detect an association between SIS risk and GLP-1RA when compared with SGLT-2i. Our findings do not support an increased risk of SIS when taking GLP-1RA; however, rarity of SIS events calls for a cautious interpretation of our results.
Our study is subject to some limitations. First, the VID databases gather several sources of real-world clinical practice data and contain information as registered by health professionals during routine clinical practice, but data are not specifically prepared for research. In this sense, studies based on real-world clinical information like the VID are at risk of well-known biases such as differential recording, misclassification bias or missing data. For instance, identification of suicidal outcomes by means of ICD codes registered during routine clinical practice databases has been associated with a slight under-registration of cases, which may result in underestimation of the association between drug exposure and SIS outcomes [35], and studies validating suicidal classification have concluded that ICD codes have good positive predictive value but low specificity [36]. Second, despite accounting for many relevant covariates in the adjustment and employing IPTW techniques, we may have missed some potentially relevant information and thus we cannot rule out residual confounding. For instance, covariates such as income or BMI strata included categories for missing data, which may introduce bias in observational studies. However, results of sensitivity analyses were, overall, comparable to those of our main analysis, suggesting robustness of the observed associations. Also, given the observational nature of the study, we could expect the presence of indication bias, a type of confounding bias that occurs when a symptom or sign of disease is judged as an indication for a given therapy, and is therefore associated both with the use of a drug and with a higher probability of an outcome related to the disease for which the agent is indicated [37]. In this sense, our results must be interpreted with caution. Third, suicidal outcomes are rare events; because of the low incidence of SIS, and based on our findings, it is not possible to rule out either a threefold increase or a decrease in its occurrence among people treated with GLP-1RA. Therefore, association estimates should be interpreted with extreme caution. Fourth, coding decision-making in real-world practice may be subject to variability, a potential bias frequent in real-world studies. Fifth, information on inpatient medication is not available in VID, which may result in marginal misestimation. Sixth, stratified analyses resulted in some groups with a relatively low number of participants and events, where risk estimates should be considered as exploratory findings. Finally, the generalisation of our results to other settings outside Spain, or even to other Spanish regions, should be approached with caution. However, even if findings in one population cannot be generalised to a new setting if the association of interest is modified by patient characteristics, settings, or treatment variations which differ in the new setting, our findings are comparable to those observed in pharmacovigilance and pharmacoepidemiologic studies that do not find increased risk of SIS associated with GLP-1RA.
In view of the steady market growth of GLP-1RA drugs and prospects of potential expansion to hugely profitable and self-selling indications such as weight management or Parkinson’s disease [38], research is urgently needed to clarify the relationship between the use of GLP-1RA and mental health, as well as to identify individuals at a higher risk of suicidal ideation and self-injury. Further studies, including final evaluations from regulatory bodies, are warranted to discard a causal link between GLP-1RA and suicidality (shortly before publication of this article, EMA released a news report in line with our findings [39]).
Supplementary Information
Below is the link to the electronic supplementary material. ESM 1 (PDF 258 KB)