Utilizing a bidirectional and multivariable Mendelian randomization analysis, we aimed to explore the potential causal relationship between genetically predicted TSH levels and FT4 levels and their association with SSNHL. A robust MR design relies on three fundamental assumptions: (1) The correlation hypothesis, which assumes a strong correlation between genetic variation and the exposure factors (thyroid function). (2) The independence hypothesis, which posits that genetic variation is independent of confounding factors that may influence both the exposure and outcome. (3) The exclusivity hypothesis, which suggests that genetic variation only impacts the outcome (SSNHL) through exposure and not through alternative pathways (23). Figure 1 provides an overview of the design employed in this thyroid function-SSNHL two-sample bidirectional MR study. As this study involves a reanalysis of previously published data, no additional ethical approval is required.

The genetic association of TSH within the reference range was obtained from the GWAS meta-analysis conducted by Zhou et al., encompassing a total of 119,715 subjects from the Nord-Trøndelag Health Study, Michigan Genomics Initiative, and the ThyroidOmics consortium, with over 22.4 million genetic markers. These data are available for download from the GWAS database (https://www.ebi.ac.uk/gwas) (24). The summary data of FT4 within the reference range was derived from the GWAS meta-analysis carried out by Teumer et al. that involved 49,269 individuals and more than 8 million genetic markers. These data can be accessed for download from the dbGaP website with the accession number phs000930 (25).

Genetic data pertaining to SSNHL was obtained from the publicly accessible GWAS database, specifically identified with the entry number “finn-b-H8_HL_IDIOP.” The study encompassed a total of 196,592 participants, consisting of 1,491 cases and 195,101 controls.

Based on the GWAS results for TSH and FT4, we conducted a rigorous screening of single nucleotide polymorphisms (SNPs) that exhibited close associations with TSH and FT4, achieving genome-wide significance (P < 5 × 10⁻⁸). SNPs closely associated with SSNHL were defined by the criterion of a P-value of <10⁻⁶. These selected SNPs were then utilized as instrumental variables (IVs) in the Mendelian randomization (MR) analysis. The IVs for MR analysis were chosen based on the following criteria: (1) To mitigate estimation bias resulting from weak IVs, we employed the equation F = (R² × (n – 2))/(1 – R²) to assess the correlation between instrument strength and exposure. A significant correlation was considered when F ≥ 10. The estimated R² of IVs was calculated using the equation 2EAF (1 – EAF)^*β², where EAF denotes the frequency of the effect allele and β represents the estimated genetic impact on FT4 (or TSH) (26). (2) To account for the influence of linkage disequilibrium, we ensured that the r²-value was less than 0.001 within a distance of 10MB (27). (3) In order to satisfy the exclusive hypothesis (that IV variants solely affect SSNHL through thyroid function), any hearing loss-related SNP outcomes (P < 1 × 10⁻⁵) were excluded from each analysis as well. Phenoscanner search was used to eliminate all known phenotypes associated with any genetic instruments considered in our analysis (28). In Supplementary Table 1, we summarized the association between exposure and SNPs and their relationship to outcome.

The primary analysis employed to assess the causal relationship between the exposure and outcomes is the inverse variance weighting (IVW) method. This method involves regressing the genetic variance (exposure) of TSH and FT4 against the genetic variance (outcome) of SSNHL, with each data point representing a conflict (29). However, it is important to note that the estimated effect obtained through IVW may be subject to bias. To address this, we conducted additional sensitivity analyses using MR-Egger and weighted medians as supplementary approaches to IVW (30). The intercept derived from the MR-Egger regression model serves as an indicator of directional pleiotropy, whereby a p-value below 0.05 suggests the presence of horizontal pleiotropy (31). Weighted median estimates generally provide robust estimates that are nearly as accurate as those obtained through IVW even in situations where more than half of the genetic variants violate assumptions (32). The MR-PRESSO method detects and eliminates outliers to yield relatively unbiased estimates while also identifying potential horizontal pleiotropic effects of SNPs through global testing (33). We utilized Cochran's Q-test to assess the heterogeneity of all SNPs. Additionally, employing the leave-one-out method, we systematically removed each SNP one at a time to evaluate whether bias in MR estimation is driven by a single SNP by calculating the causal effect of gene-predicted exposure on the outcomes using the remaining SNPs (33).The reverse MR analysis was conducted with the objective of exploring whether SSNHL might be a risk factor for FT4 levels or TSH levels. Given that genotypes are established at conception in accordance with Mendelian segregation laws, the potential for reverse causality is greatly reduced (34).

The function of multivariate MR is similar to the independent evaluation of the effect of several intervention modalities in a randomized controlled trial. For this approach, genetic tools may be associated with multiple risk factors, but they must meet the equivalent instrumental variable assumption (35, 36). To investigate the independent influence of FT4 and TSH on the risk of SSNHL, given their close correlation, we employed multivariate MR analysis. When MR analysis showed a causal relationship between FT4 (TSH) and SSNHL, multivariate MR analysis was performed to evaluate the role of TSH (FT4) as a risk factor for SSNHL. The SNPs used in the multivariate MR analysis were derived from the combination of instrumental variables (IVs) identified in the univariate MR analysis for each exposure (35). To ensure data quality, we limited our analysis to SNPs with a clumping threshold of r² <0.001 within a 10 MB region and removed any duplicates. A p-value less than 0.05 was considered statistically significant when estimating the causal effect of exposure. All statistical analyses were performed using the R package “TwoSampleMR2 (version 0.5.6)” in R (version 4.2.1). For further details, please refer to the following link: https://mrcieu.github.io/TwoSampleMR/ (37).

The summary statistics of SSNHL include all 19 SNPs associated with FT4 levels. The F-statistic for each of the SNPs included in the analysis exceeded 10 (FT4 F statistics ranged from 32.66 to 479.82). Phenoscanner analysis showed that there was no association between SNPs and any other traits that could confound the exposure-outcome relationship. The MR analysis using the IVW method revealed a significant causal relationship between FT4 levels and the risk of SSNHL (OR = 0.747, 95% CI = 0.565–0.987, P = 0.04). The forest plot illustrates that genetically predicted FT4 levels are significantly associated with SSNHL (Figure 2A). Similarly, risk estimation results in MR-Egger regression and weighted median methods demonstrate a similar trend although the associations did not reach statistical significance (Figure 2B; Table 1). The P-values obtained from the Cochran Q-test for MR-Egger (Cochrane's Q = 10.54, P = 0.88) and IVW (Cochrane's Q = 10.50, P = 0.91) were both greater than 0.05, indicating no heterogeneity in the results (Figure 2C). No abnormal instrumental variables were found to contribute to pleiotropic effects in the overall MR estimation, as indicated by the global test for MR-PRESSO (P Global Test = 0.88). The leave-one-out sensitivity analysis suggests that the overall impact of FT4 on SSNHL is not driven by a single SNP (Figure 2D). We conducted a reverse MR analysis to assess the causal impact of SSNHL on FT4 levels. Following the application of the aforementioned criteria, we identified 14 SNPs significantly associated with SSNHL (Supplementary Table 1). In our reverse MR analysis employing the IVW method, we found no significant evidence supporting a causal relationship between SSNHL and the risk of FT4 levels (OR = 1.026, 95% CI = 0.999–1.054, P = 0.056; Supplementary Table 2). Finally, the results of the multivariate MR analysis, with adjustments made for TSH levels using the IVW method, demonstrated a direct causal effect of FT4 levels on the risk of SSNHL (OR = 0.929, 95% CI = 0.867–0.995, P = 0.036).

The summary statistics of SSNHL include all 88 SNPs associated with TSH levels, each with an F-statistic greater than 10 (TSH F statistics ranged from 37.00 to 1541.78). However, we did not find significant evidence indicating a potential causal effect of TSH on the risk of SSNHL (IVW, OR = 1.409, 95% CI = 0.895–1.230, P = 0.547; Figure 3A). Similar risk estimation results were observed in the MR-Egger regression and weighted median methods. The Cochran Q-test results for MR-Egger (Cochrane's Q = 69.39, P = 0.90) and IVW (Cochrane's Q = 70.47, P = 0.90) yielded p-values greater than 0.05, indicating no heterogeneity in the results (Figure 3B; Table 1). No abnormal instrumental variables were identified that could lead to pleiotropic effects in the overall MR estimation, as indicated by the global test for MR-PRESSO (P Global Test = 0.88; Figure 3C). In the leave-one-out sensitivity analysis, it was determined that the association between TSH and SSNHL was not driven by a single SNP (Figure 3D). We conducted a reverse MR analysis using the IVW method, and the results did not yield significant evidence indicating a potential causal effect of SSNHL on TSH levels (OR = 1.002, 95% CI = 0.989–1.015, P = 0.702; Supplementary Table 2). Furthermore, even in the multivariate MR analysis, where adjustments were made for FT4 levels, we found no evidence supporting a direct causal relationship between TSH levels and the risk of SSNHL (IVW, OR = 1.011, 95% CI = 0.880–1.161, P = 0.867).