This was a single-centre, retrospective observational study conducted at Qinghai University Affiliated Hospital between August 2024 and May 2025. A total of 301 patients with hearing loss were included.

Inclusion criteria were as follows: (1) age ≥18 years; (2) completed audiological assessments, medical history, and psychological questionnaires; (3) completion of the Tinnitus Handicap Inventory (THI) for evaluating tinnitus severity; and (4) availability of all key clinical variables required for the analysis. Exclusion criteria included: (1) active acute ear diseases (e.g., otitis externa or media, sudden sensorineural hearing loss); (2) history of cranial trauma or otologic surgery; (3) diagnosed neurodegenerative or severe psychiatric disorders; and (4) incomplete or missing data.

All eligible patients were randomly divided into a training set (n = 210) and a validation set (n = 91) in a 7:3 ratio. The study was approved by the Ethics Committee of Qinghai University Affiliated Hospital (Approval No. P-SL-2024-056), and written informed consent was obtained from all participants.

A total of 18 candidate variables were included in the analysis to develop the predictive model. The primary outcome was moderate-to-severe tinnitus in patients with hearing loss. According to the THI, tinnitus severity is categorized as follows: slight (0–16), mild (18–36), moderate (38–56), severe (58–76), and catastrophic (78–100) (18, 19). In this study, patients with a THI score ≥ 38—corresponding to moderate, severe, or catastrophic tinnitus—were classified as having moderate-to-severe tinnitus and coded as 1. Those with a THI score < 38 (i.e., slight or mild tinnitus), or with no tinnitus symptoms, were coded as 0. This binary classification was used as the outcome variable in model construction.

The predictors encompassed five domains: demographic, otological, clinical, psychological, and sleep-related factors. Demographic variables included age, gender, and BMI, which was categorized as underweight (<18.5), normal (18.5–23.9), overweight (24.0–27.9), or obese (≥28.0) based on Chinese criteria (20). Otological indicators included hearing loss severity—classified as mild (26–40 dB), moderate (41–60 dB), or severe (61–80 dB)—as well as the duration of hearing loss (<12 vs. ≥12 months), laterality (unilateral vs. bilateral), family history of hearing loss or tinnitus, history of ototoxic drug exposure, and occupational noise exposure (≥85 dB for ≥8 h/day for ≥1 year).

Clinical comorbidities included hypertension, diabetes, hyperlipidemia, and laryngopharyngeal reflux (LPR), the latter defined by a Reflux Symptom Index (RSI) ≥ 13 and a Reflux Finding Score (RFS) ≥ 7 (21, 22). Psychological variables included anxiety and depression, assessed using the Zung Self-Rating Anxiety Scale (SAS) and the Patient Health Questionnaire-9 (PHQ-9), respectively. Anxiety was defined as SAS ≥ 50, and depression as PHQ-9 ≥ 10 (23, 24). Sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI), with a score >5 indicating poor sleep quality (25).

These variables were selected based on prior literature and clinical relevance, and were included in subsequent univariate and multivariable logistic regression analyses to construct the final prediction model.

All statistical analyses were conducted using R software (version 4.4.2). Continuous variables were tested for normality. Those with a normal distribution were presented as mean ± standard deviation and compared using the independent samples t-test, while non-normally distributed variables were expressed as median and interquartile range (IQR) and compared using the Mann–Whitney U test. Categorical variables were summarized as frequencies and percentages and compared using the chi-square test.

To identify predictors of moderate-to-severe tinnitus, univariate logistic regression analysis was first performed. Variables with a p-value < 0.05 in univariate analysis were subsequently included in a multivariate logistic regression model to identify independent risk factors. A predictive nomogram was constructed based on the final multivariable model.

Model performance was evaluated using several metrics: discrimination was assessed by the area under the receiver operating characteristic (ROC) curve (AUC) and concordance index (C-index); calibration was assessed via Brier score and calibration curves; and clinical utility was evaluated using decision curve analysis (DCA). Internal validation was performed using 1,000 bootstrap resamples. The Hosmer–Lemeshow goodness-of-fit test was used to assess model calibration. A two-sided p-value < 0.05 was considered statistically significant. All analyses and visualizations were implemented using R packages including rms, pROC, and rmda.

There were no statistically significant differences between the moderate-to-severe tinnitus group (THI ≥ 38, n = 186) and the no or slight-to-mild tinnitus group (THI < 38, n = 115), based on chi-square tests, in terms of gender (p = 0.702), diabetes (p = 0.283), smoking (p = 0.206), alcohol consumption (p = 0.471), duration of hearing loss ≥12 months (p = 0.778), history of LPR (p = 0.152), ototoxic drug use (p = 0.234), noise exposure (p = 0.081), family history (p = 0.226), hyperlipidemia (p = 0.896), or BMI category (p = 0.104).

In contrast, patients with moderate-to-severe tinnitus were significantly more likely to be of older age (≥60 years) (78.5% vs. 60.9%, p = 0.002), have hypertension (45.2% vs. 27.8%, p = 0.004), poor sleep quality (66.7% vs. 49.6%, p = 0.005), anxiety symptoms (57.5% vs. 40.9%, p = 0.007), depressive symptoms (55.4% vs. 35.7%, p = 0.001), severe hearing loss (39.8% vs. 22.6%, p < 0.001), and bilateral hearing loss (59.1% vs. 44.3%, p = 0.017).

Detailed comparisons of basic patient characteristics between the two groups are summarized in Table 1.

Univariate logistic regression analysis was performed on all candidate variables in the training set. The results indicated that age, hypertension, poor sleep quality, anxiety, depression, noise exposure history, and hearing loss severity were significantly associated with the presence of moderate-to-severe tinnitus among patients with hearing loss (all p < 0.05). These significant variables were subsequently entered into a multivariable logistic regression model. The analysis revealed that age (OR = 2.415, 95% CI: 1.243–4.746, p = 0.010), hypertension (OR = 2.120, 95% CI: 1.113–4.134, p = 0.024), poor sleep quality (OR = 2.821, 95% CI: 1.477–5.533, p = 0.002), anxiety (OR = 1.967, 95% CI: 1.044–3.771, p = 0.038), and severe hearing loss (OR = 3.452, 95% CI: 1.629–7.631, p = 0.002) were independent predictors of moderate-to-severe tinnitus. Details of the univariate and multivariable analyses are presented in Tables 2, 3. These five variables were incorporated into the final prediction model and visualized using a nomogram.

The predictive model was constructed based on the training set. Variables that were statistically significant in the univariate analysis (p < 0.05) were entered into a multivariable logistic regression model. Ultimately, five independent predictors of moderate-to-severe tinnitus were identified: age, hearing loss severity, poor sleep quality, anxiety, and hypertension (all p < 0.05). These variables formed the basis of the final prediction model. A nomogram was subsequently developed to provide individualized risk estimates for moderate-to-severe tinnitus in patients with hearing loss (Supplementary Figure 1), serving as a practical and visual tool for clinical risk assessment.

The predictive model based on five independent variables—age, hearing loss severity, poor sleep quality, anxiety, and hypertension—demonstrated good discriminatory ability in both the training and validation sets. The area under the AUC was 0.734 (95% CI: 0.664–0.806) in the training set and 0.760 (95% CI: 0.647–0.859) in the validation set, demonstrating acceptable discriminative performance. The corresponding Brier scores were 0.196 and 0.186, respectively, suggesting low prediction error and good model accuracy. Calibration curves based on 1,000 bootstrap resampling iterations showed excellent agreement between predicted probabilities and observed outcomes. The Hosmer–Lemeshow goodness-of-fit test also supported adequate model fit in both the training (χ² = 6.50, p = 0.591) and validation (χ² = 7.38, p = 0.496) DCA further indicated that the model provided a positive net clinical benefit across a wide range of threshold probabilities in both sets, supporting its potential utility for individualized risk stratification. To assess the model’s stratification ability in practical settings, patients in the validation set were divided into three groups—low-, medium-, and high-risk—based on tertiles of predicted probabilities. The actual incidence of moderate-to-severe tinnitus increased progressively across these groups, confirming the model’s capacity to effectively stratify risk levels. These findings are comprehensively illustrated in Figure 1.

This study has several limitations. First, its cross-sectional design precludes any inference of causality; it is unclear whether poor sleep and anxiety precede tinnitus onset or result from it. Longitudinal studies are needed to clarify temporal relationships. Second, our depression measure may have been underpowered; structured diagnostic interviews or longer-term assessments might reveal more nuanced effects. Third, the study was conducted at a single tertiary care centre in western China, which may limit generalizability to other populations. Cultural, genetic, and environmental differences should be considered when interpreting these findings.

We also lacked objective audiometric and sleep data. Detailed hearing profiles (e.g., tinnitus pitch matching, residual inhibition tests) and objective sleep assessments such as polysomnography would enhance the accuracy of tinnitus phenotyping and allow better mechanistic insight. Biomarkers such as cortisol levels or heart rate variability could also clarify the role of neuroendocrine and autonomic dysregulation. Future research should incorporate multimodal approaches—including neuroimaging, genetics, and longitudinal follow-up—to build a more comprehensive model of tinnitus risk.