Risk of Hearing Loss in Patients Using Exendin-4 Based Diabetes Drugs Compared to Other Glucose-Lowering Treatments

We conducted comprehensive searches across multiple databases, including PubMed, Embase, ClinicalKey, Cochrane CENTRAL, ProQuest, ScienceDirect, Web of Science, and ClinicalTrials.gov (Table S2), covering publications up to 1 February 2025. Regarding the sources of unpublished trials, we actually used several databases, including ScienceDirect (in which several important poster abstract will be posted here), ProQuest (in which some thesis would be posted here), and ClinicalTrials.gov (in which some gray literature would be uploaded here). Two independent researchers (PT Tseng and BY Zeng) carried out these searches, screened titles and abstracts, and resolved disagreements through discussion. We also manually reviewed reference lists from relevant review articles and meta-analyses for additional studies [18,19]. No language restrictions were applied.

This NMA was guided by the PICOS framework (Population, Intervention, Comparison, Outcome, Study Design): Population: Individuals without documented hearing loss at baseline; Intervention: Use of GLP-1 receptor agonists or SGLT2 inhibitors at any dose; Comparison: Placebo, standard care, or active comparator; Outcome: Incidence of hearing loss, either sensorineural hearing loss or conductive hearing loss; Study design: RCTs.

To minimize selective reporting bias, only RCTs with systematic adverse event assessments or targeted outcome evaluations were included [20]. Further, RCTs exclusively enrolling participants with documented hearing loss at baseline were excluded to maintain focus on causative or prophylactic evaluation. Therefore, to reduce the potential heterogeneity and bias, the inclusion criteria were strictly set as: (1) RCTs with participants free from documented hearing loss at baseline; (2) RCTs evaluating GLP-1 receptor agonists or SGLT2 inhibitors; (3) Studies involving human subjects; and (4) RCTs with systematic adverse event monitoring or direct evaluation of target outcomes.

Exclusion criteria included the following: (1) non-RCT studies; (2) RCTs specifically enrolling participants with documented hearing loss at baseline; (3) RCTs lacking direct comparisons of GLP-1 receptor agonists or SGLT2 inhibitors; (4) RCTs lacking information on target outcomes; and (5) animal studies.

Two independent reviewers assessed the risk of bias using the Cochrane Risk of Bias Tool 1.0 [21], achieving an inter-rater reliability score of 0.86. Discrepancies were resolved by consulting a third reviewer.

Due to variations in event reporting methods, the primary outcome was defined as "total hearing loss events recorded in registry systems". Event counts rather than individual patient numbers were considered. The hearing loss included sensorineural hearing loss or conductive hearing loss because they were frequently comorbid. Since the cut-off point between complete hearing loss and hearing impairment varied widely across different regions and years, we pooled these two diagnoses into one group but did not separate them. We calculated the dropout rates, indicating study withdrawals for any reason.

In order to provide more clinically relevant information to clinicians (i.e., the potential dose-dependent risk of hearing loss), we conducted a subgroup analysis by grouping RCTs by various dosages of individual medication, which procedures had been widely accepted in various fields of research [22,23,24]. The dose definitions followed original RCT classifications [10,11,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50]. However, if there was only one uniform dose used in the recruited RCTs, we would not subdivide them into subgroups. Rather, we will count them as single groups, such as dapagliflozin, lixisenatide, and liraglutide.

If RCTs applied an escalating titrating dosage strategy in the study design, we would use the "highest dosage (the final dosage)" as our recording. However, if RCTs consisted of a parallel low-dose arm and high-dose arm at the same time, we would use both dosages in our recording.

Two independent authors (PT Tseng and BY Zeng) extracted relevant data, including demographics, study design, treatment details, primary outcomes, and drop-out data. Missing information was requested by the corresponding authors. Data extraction followed protocols from the Cochrane Handbook for Systematic Reviews of Interventions and other medical literature standards [51].

We conducted network meta-analyses using MetaInsight (version 4.0.2, Complex Reviews Support Unit, National Institute for Health Research, London, UK), implementing a random-effects model within a Bayesian framework. The analyses were performed using the package of gemtc, BUGSNET, and bnma in R. For categorical data management, the procedures of corrections to single-zero-event studies would be operated by the MetaInsight programs [52]. We would exclude studies with zero events in both treatment and control arms to minimize bias [53,54]. Effect sizes were calculated using odds ratios (ORs) with 95% credible intervals (95% CrIs) and presented in forest plots. We performed Bayesian analysis using four Markov Chain Monte Carlo (MCMC) chains, each with a total of 25,000 iterations, discarding the first 5000 as burn-in. A thinning interval of 1 was applied, resulting in a total of 80,000 posterior samples for inference. Similarly, the relative effects of treatments compared to placebo were assigned normal priors with a mean of zero and the same precision. Model fit and consistency were assessed using residual deviance plots, per-arm residual deviance, and leverage plots. Convergence was visually inspected using Gelman convergence assessment plots for all studies.

Treatment rankings and effect sizes for both direct and indirect comparisons were generated and systematically tabulated. We assessed the consistency between direct and indirect evidence using the 'node-splitting' method, which is particularly valuable for network meta-analyses with available trial-level data. To evaluate the relative superiority of individual regimens, we conducted Bayesian-based surface under the cumulative ranking (SUCRA) evaluations, visualized through Litmus Rank-O-Gram and radial SUCRA plots.

Finally, we evaluated the heterogeneity with the tau value, I², and related p value. Publication bias would be assessed with a funnel plot.

To validate the reliability and convergence of treatment estimates, we employed a deviation-model analysis [55]. Specifically, each bar represents the residual deviance for an individual study arm; larger residual deviance values indicate poorer model fit for that arm in the per-arm models. Points positioned towards the upper-right corner suggest study arms with potentially poorer model fit and/or higher influence on the network meta-analysis results in the leverage plot. Further, we evaluated the overall quality of evidence in our NMA with the GRADE, which was a recognized framework to evaluate the certainty of evidence [56]. Finally, to re-validate the results of our NMA, we re-analyze our data with risk ratio (RR) and related 95% CrIs. To reduce the potential bias related to heterogeneous baseline illnesses across the recruited RCTs, we arranged subgroup analysis focusing on those RCTs recruiting subjects with a baseline illness of diabetes, which was the major indication of GLP-1 receptor agonists and SGLT2 inhibitors.

This study complies with the principles outlined in the Declaration of Helsinki.

Figure 1 illustrates the flowchart summarizing the literature search and screening process for this NMA. After excluding 118 articles for various reasons (Table S3), a total of 28 articles encompassing 29 RCTs were included in the analysis (Table S4) [10,11,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50]. The selected studies involved 145,895 participants (mean age = 64.0 years, range: 50.8 to 71.9 years; mean female proportion = 35.3%, range: 20.8% to 53.6%). The average study duration was 146.6 weeks (range: 13 to 281 weeks). The investigated GLP-1 receptor agonists included tirzepatide, efpeglenatide, liraglutide, albiglutide, dulaglutide, exenatide, semaglutide, and lixisenatide. The investigated SGLT2 inhibitors included bexagliflozin, canagliflozin, empagliflozin, ertugliflozin, dapagliflozin, and sotagliflozin.

The main results of the primary outcome revealed that lixisenatide (number needed to harm = 757.75) was associated with increased events of hearing loss in comparison with the controls. In contrast, only albiglutide was associated with fewer hearing loss events than the controls. Among the investigated regimens, the risk of lixisenatide ranked the highest (Figure 2A, Figure 3A and Figure S3A, and Table 1A). There was no significant heterogeneity detected (tau < 0.001, I² = 0%, p value = 0.9934). The funnel plot revealed symmetry (Figure S7A).

In our subgroup analysis of various dosages, lixisenatide and high-dose efpeglenatide (i.e., 6 mg/week) (number needed to harm = 757.75 and 679.00, respectively) were associated with increased events of hearing loss in comparison with the controls, whereas oral semaglutide revealed only a borderline association. In contrast, only low-dose tirzepatide, sotagliflozin, albiglutide, and low-dose injectable semaglutide were associated with fewer hearing loss events than the controls. Among the investigated regimens, the risk of lixisenatide ranked the highest (Figure 2B, Figure 3B and Figure S3B, and Table 1B). There was no significant heterogeneity detected (tau < 0.001, I² = 0%, p value = 0.8874). The funnel plot revealed symmetry (Figure S7B).

In our sensitivity test with re-validation by RR, the main results of the current NMA would not be changed. Specifically, lixisenatide was still the only regimen associated with increased events of hearing loss in comparison with the controls (). When subgrouping according to various dosages, the lixisenatide and high-dose efpeglenatide (i.e., 6 mg/week) were still associated with increased events of hearing loss in comparison with the controls (). Figure S2A and Table S5A Figure S2B and Table S5B

In this part of the subgroup analysis, the main results of the current NMA would not be changed. Specifically, lixisenatide was still the only regimen associated with increased events of hearing loss in comparison with the controls (). When subgrouping according to various dosages, the lixisenatide and high-dose efpeglenatide (i.e., 6 mg/week) were still associated with increased events of hearing loss in comparison with the controls (). Figures S1A and S2C and Table S5C Figures S1B and S2D and Table S5D

None of the investigated regimens were associated with different drop-out rates compared to controls (). Figures S1C, S2E and S3C, and Table S5E

The Bayesian-based SUCRA ranking list is depicted in. The deviation-model assessment did not demonstrate significant deviation among the current NMA (). Table S6A–C and Figure S4A–F Figure S5A–I

We identified that 81.8% (166/203 items), 17.2% (35/203 items), and 1.0% (2/203 items) of the included studies had low, unclear, and high risks of bias, respectively (). The inconsistency test, evaluating the assumption of consistency, showed no significant inconsistencies in the subgroup analysis of various dosages. In contrast, the inconsistency could not be evaluated in the main NMA because of its star-shaped in-network geometry (). The overall quality of evidence of this NMA falls within the moderate-high range (). Figure S6A,B Table S7A–C Table S8A–C

This NMA offers several methodological strengths that enhance the reliability and clinical utility of our findings. First, the NMA designation could enable statistical comparisons between different GLP-1 receptor agonists and SGLT2 inhibitors, providing more comprehensive evidence than RCTs or traditional pairwise meta-analyses. Our rigorous methodology included exclusively focusing on RCTs, ensuring high-quality evidence while minimizing potential bias. By specifically excluding participants with pre-existing hearing loss at baseline, we were able to isolate true causative or prophylactic effects. Furthermore, our detailed subgroup analyses across various dosages offer clinicians granular evidence to select the optimal dosage for specific patient populations who already had risk factors of hearing loss at baseline.

Some limitations warrant consideration. First, there were only a few studies addressing hearing loss related to GLP-1 receptor agonists or SGLT2 inhibitor use so that the included RCTs and recruited participants were relatively few (29 RCTs involving 145,895 participants). Second, our stringent focus on RCTs, while ensuring methodological rigor, potentially excluded valuable observational data from long-term cohort studies. Third, the variation in diagnostic approaches across multi-country trials presents a notable limitation. The lack of structural and uniform diagnostic methods to detect hearing loss may have introduced heterogeneity in case identification, potentially affecting the precision of our effect estimates. Although meta-analyses may raise concerns about "inconsistencies in diagnostic tools and baseline illnesses", the nature of these studies is rooted in large databases or numerous clinical studies, making them more reflective of clinical experience. Therefore, many meta-analyses on drug side effects [,] have become a benchmark for guiding future research [,]. Fourth, the star-shaped network geometry in the main outcome would limit the evaluation of inconsistency. Fifth, the main results of this NMA mainly came from a few RCTs [,] so that the clinicians should pay special attention when interpreting the results of our study in their clinical practice. Finally, all the recruited RCTs were designed based on various underlying diseases (i.e., diabetes or obesity) but not specifically to examine hearing loss occurrence, which might lead to ignorance of the occurrence of mild hearing loss. Because adverse effects data are often handled with less rigor than the primary beneficial outcomes of a study, incomplete reporting might still exist when synthesizing data. Therefore, future studies using various audiometric tests to routinely follow patients' hearing function should be warranted. These limitations suggest the need for large-scale studies specifically designed to assess causative or preventive effects in hearing loss. ^{15 16 64 65 10 11}