Introduction

Noise pollution is a major public health concern worldwide1, which has negative impacts on different aspects of the human health including hypertension, ischemic heart disease, sleep disorders in addition to hearing impairment and tinnitus2. Noise exposure is a severe concern for heavy industrial workers. The prevalence of hazardous occupational noise exposure has been reported to be approximately 25% in USA3, 15% in Canada4, 20% in the European Union5, and 20% in Australia6. While some evidence indicates that occupational noise exposure may be decreasing in parts of the developed world, workplace noise is increasing in many developing countries as economies shift from an agricultural to a more industrial base7. Moreover, noise exposure is increasing in the general living environment due to multiple factors such as urbanization, entertainment, traffic, and the use of personal listening devices.

Noise-induced hearing loss (NIHL) is the most common noise-related health problems8, affecting approximately 20.3% of occupational individuals9,10. Severe NIHL impairs spoken communication potentially contributing to social isolation, depression and cognitive decline11,12,13. NIHL typically spreads from the high frequencies to the speech frequencies with accumulated noise exposure14,15,16. According to the Chinese national standard (GBZ 49-2014) 0.5, 1, and 2 kHz define the speech frequencies for the Chinese language (CN-standard)17,18,19, slightly different from the World Health Organization speech frequencies of 0.5, 1, 2, and 4 kHz (WHO-standard)20,21,22,23. Thus, significant communication disorder is not seen until the hearing loss spreads into the speech frequencies, i.e., NIHL at speech-frequency (SFHL)24,25. Although hearing loss above speech frequencies impacts speech perception, the effect is generally small26. Due to the importance of hearing with spoken language, hearing loss in speech frequency range is often the focus of hearing preservation. Therefore, regulations for occupational noise exposure focus on hearing loss in the speech frequency range27. In addition to communication difficulties, SFHL is a significant risk factor for other adverse events such as social isolation, decreased physical activity, dementia, cancer, and increased mortality28,29,30,31. Therefore, preventing NIHL has important implications for health care.

A significant characteristic of NIHL is its large variability even among subjects experiencing similar noise exposure. This extreme variability has given rise to the concept of “tender” versus “tough” ears to describe individuals that are highly susceptible versus extremely resistant to NIHL. Therefore, developing a method that could accurately identify subjects with “tender ears” and removing them from jobs with high levels of noise pollution would be extremely beneficial. Previous efforts have been made to identify “tender” and “tough” ears, including the use of single nucleotide polymorphisms (SNPs)32,33,34. However, there is currently no widely accepted method to screen out “tender ears” based solely on an individual’s genetic profiles and such a method is unlikely to be available in the near future35,36. This indicates the need for comprehensive approaches like multimodal data to mitigate the adverse effects of noise exposure for hearing.

Therefore, an alternative pragmatic approach for identifying subjects at high risk of developing NIHL is to attempt to develop prediction models. Many reports have utilized this approach37,38; however, they rarely focus on SFHL and genetic factors. Moreover, these models were developed using cross-sectional data rather than more tightly controlled longitudinal data suggesting that further improvements could be made.

To address some of these issues, we developed our prediction models using both cross-sectional and longitudinal multimodal data from a large pool of noise-exposed subjects. Machine learning was used to identify risk factors from the cross-sectional data. Then, risk prediction models were developed using the longitudinal data. The performance of the model was tested with the external dataset. SNP data were first incorporated into the model to determine if the genetic information would significantly improve the ability of the model to predict the risk of developing SFHL. From a public health perspective, these multimodal models provide an innovative tool for precision prevention of NIHL. By enabling individualized risk assessment, the models support stratified surveillance, targeted workplace interventions, and more effective allocation of occupational health resources. Furthermore, they provide evidence to inform regulatory policies, guide worker health promotion programs, and reduce the long-term societal and economic burden of NIHL. As industrialization and urbanization continue to accelerate worldwide, the integration of such predictive tools into occupational health practice will be essential for safeguarding hearing health and improving population well-being.

Methods

Ethics

This study was approved by the Institutional Ethics Review Board of the Shanghai Sixth People’s Hospital. The study was registered with the Chinese Clinical Trial Registry (ChiCTR-RPC-17012580). Written informed consent was obtained from every participant before the start of the study. This study adheres to the reporting guidelines of Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and TRIPOD Reporting Guidelines.

Study workflow

The whole workflow was summarized in Fig. 1. After excluding unqualified subjects, the variables that were associated with SFHL were verified by feeding the cross-sectional data to the Least Absolute Shrinkage and Selection Operator (LASSO) and Random Forrest (RF) machine learning algorithms39,40. The candidate factors were narrowed down by excluding the factors with high collinearity. In the final step, prediction models were established with longitudinal data using the template of the generalized estimating equation (GEE) model and the Nomogram Cox model. Considering the established role of sex in NIHL development41,42, separated models for males and females were developed for NIHL prediction to avoid confounding effect by the different sex distributions between training set and test set.

Fig. 1: The workflow of this study.
Fig. 1: The workflow of this study.The alternative text for this image may have been generated using AI.
Full size image

The figure summarizes the overall study design and analytical pipeline, including participant recruitment and eligibility screening, cross-sectional analysis for risk factor identification, longitudinal modeling for risk prediction, external validation, and evaluation of model performance.

Participants

A total of 9669 subjects were recruited from two shipyards (6966 from the Zhenhua shipyard, 2703 from the Jiangnan shipyard) between 15 June 2012 and 15 July 2024. Subjects were carefully evaluated by audiometry annually except for annually except for 2013–2014, 2020, and 2022–2023. The participating shipyards had implemented standardized occupational noise-control procedures, including strict adherence to 8-hour work limits in high-noise areas and mandatory use of certified noise-reduction earplugs. Afterwards, they filled out a questionnaire to obtain demographic data, including age, sex, noise exposure history, type of work, and pre-existing medical conditions. Sex generally refers to a set of biological attributes that are associated with physical and physiological features such as chromosomal genotype, hormonal levels, and internal and external anatomy.

Participants were excluded if they met any of the following criteria: (1) pre-existing conditions known to cause hearing loss (n = 651, including e.g., congenital deafness, chronic otitis media) to isolate noise-related effects; (2) abnormal middle ear impedance, ensuring audiometric reliability (n = 447); (3) change of work type during follow-up, to maintain stable noise exposure (n = 409); or (4) documented exposure to ototoxic agents, preventing confounding from chemical co-exposures (n = 1023). These exclusions aimed to enhance the specificity of the cohort for studying occupational noise-induced hearing loss. To minimize the potential confounding effect of hereditary or late-onset genetic hearing loss, we implemented a two-step screening and exclusion procedure. First, during enrollment, participants were asked to self-report any family history of childhood deafness or early-onset hearing impairment; individuals with a positive history were excluded. In China, the term ‘family history of deafness’ typically refers to congenital or early-onset hearing disability, and not to adult-onset hearing loss related to occupational noise exposure. Therefore, this item captures hereditary or developmental conditions rather than noise-related hearing loss. Second, during the initial clinical and audiometric evaluation, subjects whose audiometric profiles were inconsistent with the typical pattern of noise-induced hearing loss were also excluded, regardless of their reported family history.

Useful cross-sectional data were obtained from a total of 7139 subjects, 5053 from Zhenhua shipyard, and 2086 from Jiangnan shipyard. Useful longitudinal data were obtained from 1542 male and 241 female subjects from Zhenhua shipyard (training set), and from 601 male and 115 female subjects from Jiangnan shipyard (external test set). Importantly, the initial audiogram of subjects on which longitudinal data were collected did not have SFHL (defined as the binaural average hearing threshold <25 dB HL based on the CN-standard and WHO-standard). All subjects with repeated measurement (the crude cohort) formed two groups. The first group consisted of 1542 males and 241 females from the Zhenhua shipyard; their data formed the training set used for the longitudinal analysis. The second group consisted of 601 males and 115 females from the Jiangnan shipyard; their data formed the external test data set used for the longitudinal analysis. Since only 74 incident SFHL cases (60/2143 males; 14/356 females) were seen within the initial 1–2-year follow-up period with an overall incidence of approximately 3.0%, a reliable risk estimation was not warranted. Therefore, we extended the follow-up period to ≥3 years and constructed a “3-year (3-y) cohort” to obtain a sensitive analysis with adequate statistic power.

Audiometric evaluation

Qualified medical assistants conducted the hearing evaluation of workers on-site. Hearing thresholds (HT) were measured separately for each ear using a Type 1066 manual audiometer (Natus Hearing & Balance) with Sennheiser HDA-300 headphones capable of testing over the standard and extended high frequencies (EHF). SFHL was defined as a threshold of 25 dB or higher in the binaural pure-tone average (PTA) across the three frequencies of CN-standard and WHO-standard.

SNP selection and genotyping

Based on previous findings from the literature43,44,45,46,47, 17 NIHL SNPs were selected as candidate genetic variables. They incorporated homozygous mutations, heterozygous mutations, and no mutations into the model as graded variables, with no mutations serving as the control group. The SNPs were detailed in Table S7. Unfortunately, most of these SNPs were not derived from whole-genome sequencing. Clearer roles of genetic polymorphisms in NIHL can be identified via the genome-wide approaches. Among our included SNPs, the study by Niu et al. identified significant loci by using a GWAS chip in an additive model, and validation with an independent cohort to reach a higher reliability46. In another study, Wang et al. used whole-exome sequencing in a similar approach to identify significant NIHL loci. High reliability was warranted by the advantages of the genome-wide approach and the large sample size in validation47. Although the remaining SNPs were not derived from whole-genome sequencing data, they have been almost validated through meta-analysis43,44,45. Individuals possessing NIHL-associated susceptibility SNPs are said to have “tender” ears, while those with resistance SNPs are described as having “tough” ears.

Whole-blood samples were collected from 2499 subjects for longitudinal analysis. Genomic DNA was extracted from peripheral blood using the Magnetic Blood DNA Kit (Onrew, Fo Shan, Chinese). All the subjects were genotyped using Hi-SNP technology on the Illumina X Plus platform (Illumina, California, USA). Repeated control samples were set in every genotyping plate, and the concordance was >99%.

Noise exposure estimation

The noise exposure was measured according to the national standard of China (Measurement of physical agents in the workplace—part 8: noise)48. Briefly, an ASV5910-R digital recorder (Aihua Instruments) was used to measure the noise level on the dBA scale. Long-term equivalent noise level (Leq) was used as the primary exposure indicator and was measured three times at each location of concern, each measurement lasting for 8 hours. Cumulative noise exposure (CNE) was calculated as Leq-8h + 10 × log (T), where T represents the work duration in years.

Prediction model construction and statistical analysis

The normality of the data was evaluated by the Kolmogorov-Smirnov test. Categorical and continuous variables were assessed using the Chi-squared test and Student t-test. Continuous variables were described using the mean and standard deviation. Categorical variables were described by raw numbers and percentages. Statistical analyses were performed using R version 4.4.2. A two-sided P value < 0.05 was considered statistically significant.

The factors primarily evaluated for producing the prediction model included both demographic variables (e.g., sex, age, smoking, alcohol consumption, and CNE) and audiometric values (bilateral average of HT at 3, 4, 6, 10, and 12.5 kHz named as BHnkHz and combined frequencies, e.g., BH3_12.5 kHz, BH means binaural average hearing threshold at the corresponding frequency). The audiometric factors were filtered by LASSO and ranked by RF. The LASSO factors among the RF top 10 were further analyzed with logistic regression. The number of audiometric factors was narrowed down after eliminating the variables with high collinearity, as indicated by a variance inflation factor (VIF > 4)49. Candidate audiometric factors were compared the predictive performance using the C-index and the Hosmer-Lemeshow goodness-of-fit test. The power of each factor in risk discrimination was evaluated using the receiver operating characteristic (ROC). The cut-off value of each factor was generated as the optimal Youden index.

Considering the data trait of repeated measurement, the prediction models were established by GEE. To make the prediction model more robust, Cox models were also constructed to transform prediction risk values into graphic scores (nomograms). The probability of having SFHL for n (3, 4, and 5) years of noise exposure was based on the function of the total score. The 3-year cohort data in the external set were used for the sensitivity test. The subjects with a higher than 50% probability of developing SFHL were considered high-risk according to common practice50,51. The prediction power was evaluated by statistical calculation of the concordant (C) index, area under the curve (AUC), and calibration curve. Seventeen NIHL-associated SNPs were individually or in combination incorporated into the models, with the SNP or combination ultimately retaining the highest predictive performance being retained. To evaluate the improvement in model performance after incorporating genetic variants, supplementary analyses were performed. The Net Reclassification Index (NRI) and Integrated Discrimination Improvement (IDI) were used to assess the enhancement in risk classification and overall discriminatory ability, respectively. Additionally, the Delong test was performed to statistically compare the differences in the AUC between the models with and without SNPs.

Results

Baseline characteristics of enrolled participants

Table 1 summarizes the demographic characteristics, personal behavior, and history of diseases of 7139 included participants at their initial year (the largest sample size in NIHL field). The inclusion and exclusion of subjects were shown in the Fig. 2. In the total of 7139 participants, 2768 of 4310 males (64.3%) in the training set presented with noise‑induced SFHL at baseline, compared with 1119 of 1720 males (65.1%) in the test set. Among females, 502 of 743 (67.6%) in the training set and 251 of 366 (68.6%) in the test set had prevalent SFHL at baseline. There was no significant difference in any of the items between participants in the training set and the test set (P > 0.05) (Table S1), except for the sex distribution (P = 0.003). There were fewer female subjects than males in both sets. Table S2 compares the variables based on sex: males had higher values than females for age, CNE, smoking, and alcohol consumption. Meanwhile, regardless of male and female subjects, SFHL subjects were older and experienced prolonged noise exposure (P < 0.001) (Tables S3 and  S4). 1783 workers have the longitudinal data in the training cohort and 716 have them in the test set. The mean follow-up duration was 5.34 years in the male training cohort and 5.41 years in the male test cohort. For female participants, the mean follow-up was 4.15 years in the training cohort and 3.75 years in the test cohort. Incident cases of CN-SFHL were as follows: in the female training cohort, 23 of 241 participants (9.5%); in the female test cohort, 12 of 115 (10.4%); in the male training cohort, 236 of 1542 (15.3%); and in the male test cohort, 87 of 601 (14.5%).

Fig. 2: Outlines of the inclusion and exclusion procedures.
Fig. 2: Outlines of the inclusion and exclusion procedures.The alternative text for this image may have been generated using AI.
Full size image

The diagram shows the recruitment, exclusion, and final inclusion of participants from two shipyards for cross-sectional and longitudinal analyses.

Table 1 Baseline Information of Subjects from Two Shipyards
Full size table

Identification of risk factors using cross-sectional data

All the variables in the primary list were ranked separately for male and female subjects based on their contribution to decreasing Gini impurity (Fig. 3). An exceptionally high rank was seen at BH3kHz in both sexes. The AUC values were shown for males and females separately (AUCs > 0.85) (Fig. 4A, B). The optimal values generated from ROC of the selected factors are presented (Table S5) and were used for the calculation of risk values. Variables with VIF > 4 (collinearity) in the logistic regression analysis were excluded. The audiometric factors were narrowed down to 2 for males (BH3kHz, BH3_6 kHz) and 2 for females (BH3kHz and BH3_6kHz). The predictive performance was compared across the predictors of different frequency bands by using the C-index. As summarized in Table S8, non-EHF frequency audiometrics (BH3kHz and BH3–6kHz) showed numerically higher C-index values as compared with those in EHF range. Therefore, the use of predictors improved the predicting ability of the model, and the prediction by speech-frequency predictor was truly speech-frequency specific. In addition Hosmer-Lemeshow goodness-of-fit test was carried out to assess the agreement between predicted and observed outcome probabilities for each predictor. As shown in Table S8, the result of this test for every predictor is insignificant (P > 0.05), suggesting a reliable estimation of the observed data by each of the predictors.

Fig. 3: The audiometric factors associated with SFHL in the training sets.
Fig. 3: The audiometric factors associated with SFHL in the training sets.The alternative text for this image may have been generated using AI.
Full size image

A, B LASSO coefficient profiles. Shown are the trajectories of the coefficients (y-axis) for candidate audiometric variables across a sequence of regularization penalty values (log(λ), top x-axis). The vertical dashed line indicates the λ value selected via 10-fold cross-validation, which yielded the most parsimonious model. The shaded bands around each coefficient path represent the variability (standard error) of the coefficient estimates across the cross-validation folds. The analyses are descriptive of variable selection and no inferential statistical tests or P values are generated at this stage. Panels are shown for (A) male (n = 4310 biologically independent participants) and (B) female (n = 743 biologically independent participants) workers. C, D Random Forest feature importance ranking. Variables are ranked by their mean decrease in Gini impurity (x-axis), a measure of their contribution to classifying SFHL in the Random Forest ensemble. Error bars represent the standard deviation of the importance measure across all trees in the forest. Panels are shown for (C) male (n = 4,310) and (D) female (n = 743) workers.

Fig. 4: Discrimination and logistic regression of candidate audiometric factors.
Fig. 4: Discrimination and logistic regression of candidate audiometric factors.The alternative text for this image may have been generated using AI.
Full size image

A ROC curves depict the discriminative ability of individual audiometric factors for predicting SFHL in male workers from the training set (n = 4310 biologically independent participants). The AUC is indicated for each factor. B ROC analysis for female workers in the training set (n = 743 biologically independent participants). Forest plots display adjusted ORs with 95% CIs from multivariable logistic regression for the training set, analyzing the associations of (C) BH3kHz and (E) BH3_6kHz with SFHL, separately for males (n = 4310) and females (n = 743). All models were adjusted for age and CNE. The square represents the point estimate (OR), and the horizontal line represents the 95% CI. Exact two-sided P values are shown (all P values < 0.001); no adjustment for multiple comparisons was applied in this exploratory analysis. Corresponding forest plots from the external test set for (D) BH3kHz and (F) BH3_6kHz, for males (n = 1720) and females (n = 366 biologically independent participants). ORs and 95% CI were derived from multivariable logistic regression models. All P values are two-sided and exact values are provided. In this exploratory analysis comparing candidate audiometric factors, no statistical adjustment for multiple comparisons was applied. AUC area under the curve, BH binaural average hearing threshold, CI confidence interval, CNE cumulative noise exposure, FPR false positive rate, OR odds ratio, ROC receiver operating characteristic, TPR true positive rate.

In logistic regression, after adjusting age and CNE, odds ratios (ORs) of different measures in the training set were significant (all P < 0.001) (Fig. 4C, E). Higher PTA at the identified frequencies correlated with SFHL occurrence for male subjects (BH3kHz: OR, 8.52; 95%CI, 6.84–10.61, P < 0.001; BH3_6kHz: OR, 7.45; 95%CI, 6.05–9.17, P < 0.001) and female subjects (BH3kHz: OR, 20.38; 95%CI, 10.80–38.46, P < 0.001; BH3_6kHz: OR, 21.05; 95%CI, 10.95–40.49, P < 0.001). Using the same principle, all the variables were verified as risk factors for the external test set (all P < 0.001) (Fig. 4D, F).

Establishing the prediction model using longitudinal data

The HR values in Table 2 were obtained from the GEE model (the risk effect of all audiometric factors was adjusted by age and CNE). In the crude cohort for male subjects, higher PTA of BH3kHz and BH3_6kHz indicates a higher risk of developing SFHL (HR: 3.01, P < 0.001; HR: 3.00, P < 0.001). For female subjects, the frequencies identified as significant SFHL risk were also significant (HR: 6.01, P < 0.001; HR: 5.16, P < 0.001). All the variables were verified as SFHL risk predictive factors in the external test set (P = 0.006; P = 0.004; P = 0.009; P = 0.027) (Table 2). Therefore, for male and female subjects, age, CNE, BH3kHz, and BH3_6kHz were used to establish SFHL prediction models. Because SFHL is believed to be affected by genetic factors, SNPs were also input to the models.

Table 2 GEE analysis for risk factors adjusted age and CNE in crude and 3-y cohorts
Full size table

ROC analysis for male and female GEE models was done using the crude cohort data from the training set and external test set (Fig. S1). Overall, excellent discrimination performance was seen in training set for male (AUC for CN-standard is 0.80; 0.86 for WHO-standard) and female (AUC for CN-standard is 0.82; 0.86 for WHO-standard) subjects (Fig. 5A, B). In the external test set, the AUCs were all higher than 0.78 (Fig. S1). After incorporating NIHL SNPs, the AUCs seem to increase by 2 to 3 percentage points (Fig. 5; Fig. S1). To further quantify the incremental predictive value of incorporating genetic data, we evaluated the models using the Net Reclassification Index (NRI), Integrated Discrimination Improvement (IDI), and the Delong test for AUC comparison (Tables S10-S11). The Delong test revealed a statistically significant increase in AUC after adding SNPs in the male training set for both the CN standard (P = 0.008) and WHO standard (P = 0.037). However, this significance was not consistently replicated across all cohorts or standards; for instance, improvements in the female training set and the male test set under the CN-standard were not statistically significant (P > 0.05). The NRI indicated moderate improvement in risk classification in several cohorts—for example, 0.287 (P < 0.001) under the CN standard in the male training set—and reached 0.996 (P < 0.001) in the female test set under the same standard. Corresponding IDI values were generally small but statistically significant in most analyses. However, the magnitude of improvement was modest, and the results were not consistent across different definitions (CN vs. WHO standards) or between training and test sets. This variability, possibly influenced by sample size limitations, suggests that the added discriminative value of genetic factors in the current cohort is limited and not robust enough for generalizable application. Consequently, given the marginal and inconsistent gain in predictive performance, genetic variants were not incorporated into the final Nomogram‑Cox models intended for clinical visualization and diagnostic use.

Fig. 5: GEE prediction model discrimination and Nomogram-Cox model.
Fig. 5: GEE prediction model discrimination and Nomogram-Cox model.The alternative text for this image may have been generated using AI.
Full size image

A ROC analysis of GEE models for male workers. B ROC analysis of GEE models for female workers. C Nomogram-Cox model for male workers. D Nomogram-Cox model for female workers.

The risk value from the Cox model with crude cohorts is graphically demonstrated by nomograms into prediction scopes (Fig. 5C, D, C for male and D for female). The prediction results are displayed according to CS-standard. The contribution of each factor was presented by converting the range of the individual factor to the point scope (0 to 100); the scope points of other variables were converted proportionally and plotted on the point scope. The probability of developing SFHL was shown with total scope points with 3-, 4-, 5-year risk.

Evaluation of prediction models by sensitivity analyses

To further evaluate the sensitivity of the results, GEE models were constructed for the 3-y cohorts (Table 2). For male subjects, higher PTA of BH3kHz and BH3_6kHz indicated a higher risk of developing SFHL (HR: 3.01, P < 0.001; HR: 2.95, P < 0.001). For female subjects, the PTA of BH3kHz and BH3_6kHz also showed significant SFHL risk (HR: 5.16, P < 0.001; HR: 5.05, P < 0.001). Moreover, these variables were identified as SFHL risk predictive factors in the external test set (all P < 0.05) (Table 2). ROC analysis for male and female GEE models (Fig. S2) using the 3-y cohort data from the training set and external test set all had AUCs higher than 0.80.

For the nomogram based on the Cox model, among male subjects, the C-index of the nomogram in the training set of the crude cohort was 0.79 (95% CI, 0.77 to 0.80), very similar to the C-index of 0.78 (95% CI, 0.76 to 0.81) of the external set in the same cohort (Table S6). This similarity suggests good prediction power of the model, which was further supported by good model calibration in the calibration curves, which show the correlation between predicted and actual probability of SFHL and remarkable discrimination reflected by all AUCs that were higher than 0.7 (Fig. S3). The good prediction power of the model is supported by similar C-indexes between the training set and the external set in the 3-year cohort for male subjects and the result of model calibration and ROC analyses (Fig. S5).

For the female subjects in both the crude cohort and 3-y cohort, similar C-index values were obtained in both training sets, 0.82 (95%CI, 0.80 to 0.84) for the crude cohort, and 0.81 (95%CI, 0.79 to 0.83) for the 3-y cohort (Table S6). Good calibration and discrimination of models were also observed in the Fig. S4. These results were further verified good prediction power for the external set (Fig. S6).

Discussion

The present study aimed to establish robust prediction models for identifying noise-exposed individuals at high risk of developing SFHL, a condition of critical importance for occupational and public health. To our knowledge, this is the first investigation to integrate both cross-sectional and longitudinal multimodal data—including epidemiological, audiological, and genetic information—using the largest cohort reported to date in this field. While machine learning methods were applied for risk factor selection, traditional regression-based approaches demonstrated strong predictive efficacy, underscoring their value for practical application in occupational health settings. Our findings further revealed that the models were effective under both the CN and WHO standards for SFHL, with even higher predictive performance under the WHO criteria, thereby enhancing their international relevance and applicability. Incorporation of an external testing dataset improves the generalizability of the predictive model. Notably, incorporation of SNP into the GEE models significantly improved predictive performance, highlighting the added value of integrating genetic susceptibility into population-level risk assessments. Although such improvements were less pronounced in the Nomogram-Cox models, the overall results suggest that multimodal GEE models represent the optimal predictive framework. Moreover, we acknowledge that despite our screening procedures, a small number of individuals with unrecognized late-onset hereditary hearing loss might remain in the cohort. Importantly, the inclusion of such cases would not artifactually inflate the observed specificity of our model for noise-related factors. Since hereditary hearing loss is generally independent of occupational noise exposure, these individuals often present with elevated hearing thresholds even in the absence of significant noise exposure. Consequently, their presence in the dataset would introduce “hearing loss without high noise” instances, which would more likely dilute, rather than exaggerate, the estimated association between noise exposure and speech-frequency hearing loss. Therefore, any residual misclassification of this nature would likely bias our results toward the null, making our estimates of the predictive contribution of noise exposure conservative.

These models offer important advances in precision prevention in public health. By identifying individuals at elevated risk, they enable stratified surveillance, targeted interventions, and more efficient allocation of occupational health resources. Importantly, the inclusion of genetic information provides a foundation for personalized prevention strategies, which may be integrated into worker health promotion programs and regulatory policies. Beyond the workplace, this approach has broader implications for mitigating the growing global burden of NIHL as industrialization and urbanization intensify worldwide. The implementation of such predictive tools in occupational health practice could therefore play a vital role in reducing communication-related disability, preventing downstream mental and cognitive decline, and alleviating the substantial societal and economic costs associated with hearing loss. Risk prediction is not a new topic in the study of occupational NIHL. However, previous studies have only relied on cross-sectional data38,52. In a recent report, in which the prediction model was constructed from cross-sectional data, the authors claimed a sensitivity and specificity of over 90%38. But the validity and predictive power of the models are still uncertain; further assessment should be carried out.

Unlike most previous reports, the present study used audiometry with EHF in the hearing survey. In one of our previous studies, we reported that 12.5 kHz was the most sensitive frequency in evaluating the existence of NIHL53. However, the result of the present study indicated that 3 kHz was more impactful in predicting SFHL in both sexes, a result that is likely because 3 kHz is in much closer proximity to the speech frequencies than 12.5 kHz. Moreover, PTAs at both BH3kHz and BH3_6kHz appeared to be closely associated with SFHL risk, which is likely due two complementary mechanisms: (1) the inclusion of 3 kHz captures early basal cochlear damage that progresses apically into speech frequencies (consistent with the classical 4-6 kHz ‘noise notch’), (2) early extending of HL to 3 kHz region represents a significant loss of audible speech spectrum. Therefore, the combination of the two metrics (BH3kHz and BH3_6kHz) thus provides a more robust marker of both emerging and established cochlear dysfunction along the tonotopic gradient. This frequency is close to the traditional “notch” frequency used in the NIHL evaluation54 and suggests that subjects with NIHL extending to this frequency would have a higher risk of developing SFHL.

Age and noise are both extremely important risk factors for hearing loss. However, when both factors are present, they may have an additional interactive effect on hearing loss55,56. To test whether the SFHL model in our study population was affected by this confounding factor, we used professional statistical methods for analysis. The results showed that after introducing the interaction term age:CNE into the model, there was no significant multiplicative interaction (Table S9). Using indicators such as Relative Excess Risk of Interaction (RERI), Attributable Proportion (AP), and Synergy Index (S), we found that noise and age also did not exhibit additive interaction in relation to SFHL (For male workers, RERI:0.05[−0.05 to 0.16], AP:0.03[−0.02 to 0.09], S:1.10[1.01–1.21]; For female workers, RERI:0.002[−0.131 to 0.137], AP:0.003[−0.176 to 0.183], S:0.988[0.599–1.655]). Therefore, in our sample, we confirmed that noise and age have independent effects on SFHL, and the model is not influenced by their interaction.

Lifestyle factors such as alcohol consumption and smoking have been indicated as important risk factors for hearing loss in previous studies57,58. However, the risk effect of smoking on hearing loss remains controversial. For example, Win et al. did not obtain a significant correlation between alcohol consumption or smoking and NIHL (P > 0.05)59. In the present study, these two variables did not seem to be risk factors for NIHL. However, since the lifestyle information was collected by a self-report questionnaire with non-quantitative questions, the data may be biased.

To maintain etiological specificity for noise-related hearing loss, workers with documented exposure to ototoxic solvents (e.g., toluene, styrene) were excluded from the analytic cohort since the co-exposure to those chemicals could introduce synergistic effects and confound the isolation of a pure noise effect. In addition, the absence of standardized, quantitative exposure data precluded the reliable inclusion of this variable in the present study; a crude binary measure would have introduced misclassification without improving predictive validity. Furthermore, the effect of ototoxic solvents on NIHL has been addressed in previously studies60,61,62,63, which show that prospectively collected, quantitative solvent metrics are needed to develop integrated prediction models for workers in settings with mixed noise and chemical exposures.

Occupational noise pollution has become increasingly widespread in contemporary work environments; however, the associated health risks, particularly those affecting the auditory system, have not received sufficient attention, and effective strategies for early prevention remain inadequately developed, highlighting the need for comprehensive approaches to mitigate the adverse effects of noise exposure in occupational settings. Our prediction was established based on audiometric variables, genetic factors, and closely related or relevant commonly available variables such as CNE and age. Although NIHL is a type of acquired hearing loss, genetic contributions to NIHL have long been suspected as an important contributor. The use of longitudinal data and genetic factors is expected to improve the prediction performance, another important innovation of our current research. However, this study focused specifically on noise-exposed workers without other major auditory risk factors. Approximately 20% of potential participants were excluded due to congenital or early-onset hearing disorders, middle-ear disease, abnormal tympanometry, or exposure to ototoxic solvents. These groups may have different or heightened susceptibility to auditory injury, and therefore the current models should not be generalized to such populations. The predictive estimates presented here apply to workers with isolated occupational noise exposure. Additional research will be required to develop tailored prediction tools for individuals with mixed noise–ototoxic exposures or pre-existing auditory vulnerability.

Although this study focuses on predicting speech-frequency NIHL due to its direct impact on communication, any detectable high-frequency threshold shift represents early cochlear injury and should prompt immediate hearing preservation measures. Early NIHL warrants reinforcement of protective equipment use, review of noise exposure, and closer audiometric surveillance. Because the predictive factors identified in our model arise before SFHL occurs, the model may aid in earlier risk stratification and intervention, supporting both primary and secondary prevention strategies within occupational hearing conservation programs.

Although the present study collected data from a large initial sample, there are various limitations. Firstly, the longitudinal sample size was smaller than expected. The second limitation is that the time gap between the first survey and the final follow-up was short. This small gap likely reduced the number of subjects who developed SFHL during the follow-up. Extending the follow-up period would be an important factor in testing the model in the future. Thirdly, in the present study, we did not conduct targeted screening for known deafness-associated genes to exclude individuals with congenital HL. We only asked about the family history of deafness. Finally, the models still have not been commercialized. Further research and multicenter testing should be conducted to optimize the models.

Conclusion

Our study uses longitudinal data to develop and test the prognostic GEE multimodal models to predict SFHL, providing an effective strategy to personalize occupational noise assessment for hearing health.