Predicting Age-Related Hearing Loss in Community-Dwelling Older Adults: Multicenter Retrospective Cohort Study

Introduction

Age-related hearing loss (ARHL) is the third most prevalent health condition among older adults, after hypertension and arthritis, and has become a significant public health issue [1]. The World Health Organization estimates that more than 1.5 billion people worldwide experience hearing loss, a number projected to reach 2.5 billion by 2050 [2]. In China, hearing loss affects 30% of adults older than 60 years and rises to 50% for those older than 70 years, totaling at least 70 million cases [3]. ARHL can hinder communication and social interaction, leading to greater risks of social isolation, depression, cognitive decline [4-6], frailty, and falls [7,8].

Despite these negative health consequences, routine hearing screening remains significantly underused, with rates below 25% in older populations [9,10]. Key barriers include limited patient awareness and financial constraints; for example, a diagnostic audiogram in China can cost ¥100 (US $14.1) to ¥300 (US $42.3), a substantial out-of-pocket expense for many older adults; these challenges are further compounded by limited time, training, and equipment among primary care clinicians [11-14]. Furthermore, high false-positive rates from some hearing screening methods can increase costs, prompting policymakers to adopt a cautious approach [3,15].

Leveraging primary health care centers for the management of hearing loss is the most suitable and cost-effective approach [16,17]. A preliminary prediction tool designed for these centers could accurately identify individuals with high risk for targeted audiometric screening, thereby optimizing referrals and specialist resources. While previous models have been developed using known risk factors for hearing impairment, such as age, environmental noise [18], comorbidities [19], exercise, and nutrition [20], their clinical applicability is often limited by small sample sizes, a lack of external validation, and reliance on complex variables.

Therefore, this study aimed to develop a streamlined nomogram-based prediction model for ARHL among community-dwelling older adults, using a large sample derived from health check-ups conducted in primary health care institutions in Beijing, China. Additionally, the model was externally validated in 2 independent datasets. An online tool was used to deploy the model as an interactive application, enabling primary health care providers to perform risk prediction without requiring any software installation or programming expertise. This study adhered to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis guidelines [21].

Methods

Study Population

This multicenter retrospective cohort study used electronic health examination records from 3 community primary health care centers (center A, center B, and center C) in Beijing, China. Data were collected from older adults who participated in the annual health checkup under the National Public Health Service Project between January 1, 2020, and October 30, 2023. The dataset included comprehensive information on demographics, medical history, physical examination results, and chronic disease records. For this study, the dataset from center A, which had the largest sample size, was used for model development and internal validation, while data from center B and center C were used for external validation. A detailed study design and sample selection are shown in Figure 1.

From center A, a total of 20,923 patient records were initially screened. First, 3.94% (825/20,923) records with incomplete data were excluded. After subsequently excluding 0.26% (54/20,923) individuals with major ear conditions, 5.96% (1246/20,923) individuals aged younger than 65 years, and 0.43% (91/20,923) individuals with cognitive or neurological disorders, a final cohort of 89.41% (18,707/20,923) participants was included in the analysis. The external validation cohorts were derived from center B and center C. Initially, 11,454 records from center B and 5522 records from center C were screened. After applying the exclusion criteria, a total of 3.89% (446/11,454) participants were excluded from center B and 4.60% (254/5522) from center C. This process resulted in final validation cohorts of 96.11% (11,008/11,454) participants from center B and 95.40% (5268/5522) from center C.

‎

Assessment of Hearing Loss

The outcome variable for this study was hearing loss, which was assessed using the whispered voice test as specified by the standardized protocol in the National Basic Public Health Service Standards (Third Edition) [22], issued by China’s National Health Commission. This screening was performed by trained primary health care professionals to ensure consistency across all centers.

Specifically, the whispered voice test, a widely recognized and validated screening tool, was used due to its simplicity, cost-effectiveness, and reliability in detecting hearing loss in clinical and community settings [23,24]. The protocol dictates the following procedure: the examiner stands approximately 0.5 m from the participant’s tested ear, positioned slightly behind them to prevent lip-reading. The participant’s nontested ear is occluded (blocked) by the examiner to isolate the tested ear. After a full exhalation, the examiner whispers a simple, standardized phrase (eg, What is your name?). Failure to correctly repeat the phrase was classified as the presence of hearing loss. For the purpose of our model, this binary outcome (presence vs absence of hearing loss) was used as the dependent variable [22].

Sample Size Estimation

The determination of an adequate sample size for a multivariate predictive model requires a sufficient representation of individuals and outcome events relative to predictor parameters [25]. The combined sample size from 3 primary health care institutions provided robust support for using 33 predictive variables in both modeling and validation. Specifically, according to the widely applied 10 events per variable principle, at least 330 events (33 variables×10 events) are required [26]. Given an event rate of 0.063, this translates to approximately 5239 cases. The modeling dataset comprised 18,707 cases and 1177 events, which is more than 3 times the minimum requirement of 330 events. This large number of events per variable provides sufficient statistical power and minimizes the risk of model overfitting.

Model Development and Validation

Variable selection was conducted using the least absolute shrinkage and selection operator (LASSO) regression, which applies a Least Absolute Deviations penalty to shrink some regression coefficients to 0, simplifying the model while preserving variables with strong predictive power. Tenfold cross-validation determined the optimal penalty parameter λ (λ_min), and variables with nonzero coefficients in the final model were chosen as candidate predictors [27,28]. Next, univariate analysis was performed on the LASSO-selected variables. Variables with P values <0.05 were considered for inclusion in the multivariate logistic regression model. A forward stepwise regression approach was applied to identify the final predictors to be included in the model. To handle missing data, a complete case analysis was used. This approach, which excludes records with any missing data on key variables, was considered appropriate because the proportion of such records was low across all cohorts (3.94% in center A, 1.62% in center B, and 1.68% in center C), consistently remaining below 5%.

Statistical analysis was performed using R software (version 4.2.2; R Foundation for Statistical Computing). To facilitate clinical application of the prediction model, we developed an interactive online application using the Shiny framework (RStudio, PBC). The final logistic regression model and coefficients were embedded into the server logic using the rms package to ensure consistency with the nomogram. A user-friendly interface was created with Shiny, shiny Widgets, and shiny themes packages, allowing real-time input and calculation of predicted risk. Dynamic visualizations of individualized risk were implemented using plotly, providing immediate graphical feedback. The application was deployed on a cloud-based Shiny server to allow online access without requiring local installation.

Model validation included 3 aspects: discrimination, calibration, and clinical utility [29]. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and Matthews correlation coefficient to evaluate the model’s ability to distinguish between individuals with hearing loss and those without hearing loss. The calibration of the model was assessed using calibration plots generated through 1000 bootstrap resamples, in addition to the Brier score and mean absolute error, to evaluate the consistency between the predicted and actual observed hearing loss outcomes. Risk stratification strategy was based on the optimal cutoff value (the hearing loss risk probability when the Youden index is maximum, Youden=sensitivity+specificity−1) [30]. Finally, the clinical utility of the final model was evaluated using decision curve analysis.

Ethical Considerations

The study was approved by the ethics committee of Capital Medical University, Beijing, China (Z2024SY071). Informed consent was obtained from participants during routine health examinations, and no additional consent was required for this secondary analysis. All data were anonymized and securely stored with access restricted to the research team. No financial or other compensation was provided, as the data were derived from routine public health services.

Results

Population Characteristics

A total of 18,707 participants were included in the analysis in center A. Of these, 44.20% (8268/18,707) were male and 55.80% (10,439/18,707) were female, with a mean age of 72.89 (SD 7.15, range 65-99) years. A total of 6.29% (1177/18,707) participants were identified as having hearing loss, while 93.71% (17,530 /18,707) had normal hearing. The participants were then randomly divided into training (13,094/18,707, 70%) and internal validation sets (5613/18,707, 30%) in a 7:3 ratio. Additionally, 96.11% (11,008/11,454) and 95.40% (5268/5522) of participants were included from center B and center C, respectively, for external validation.

Model Development and Performance

The most significant features selected by LASSO from the training set were used to identify potential predictors for the model. The screening process is detailed in Figure 2. Initially, 33 variables were identified, guided by systematic reviews, meta-analyses, and expert clinical opinions. These factors encompass demographic characteristics, lifestyle factors, physiological measurements, and chronic diseases. The model’s bias showed minimal variation within the interval (λ_1se, λ_min). When λ_min was set to 0.002 in the LASSO model, 17 variables exhibited nonzero coefficients, as shown in Table 1.

‎

First, univariate analysis was conducted on the 17 variables identified through LASSO regression to evaluate their independent associations with hearing loss. The detailed results of this analysis, including odds ratios and P values for all variables, are presented in Table 1. The results indicated that age, education level, marital status, BMI, physical function, exercise frequency, dietary habits, smoking, alcohol consumption frequency, hypertension, left systolic blood pressure, and right systolic blood pressure were significantly associated with hearing loss.

Subsequently, a forward stepwise regression approach selected 6 predictors—age, education, exercise frequency, physical function, dietary habits, and hypertension—which were then included in the multivariate logistic regression model (Table 2). A user-friendly nomogram was developed, as illustrated in Figure 3. The optimal cutoff value was 0.805 based on the maximum Youden index. Accordingly, the population was classified into low (predicted probability <80.5%) and high (predicted probability≥80.5%) hearing loss risk groups. A comparison of the model’s predicted risk classifications against the observed outcomes is presented in Table 3. For example, in the training set, of the 814 individuals, the model correctly identified 639 (78.5%) with hearing loss (true positives) and of the 12,280 individuals, 10,099 (82.2%) with normal hearing (true negatives), corresponding to the sensitivity and specificity values reported.

In the training set, the model achieved an AUC of 0.828 (95% CI 0.813-0.842), sensitivity was 0.785 (639 out of 814 participants who actually developed hearing loss were correctly predicted as high-risk), specificity was 0.822 (10,099 of the 12,280 participants who did not develop hearing loss were correctly predicted as low-risk), and balanced accuracy was 0.804. In the internal validation set, AUC was 0.806 (95% CI 0.782-0.831), sensitivity was 0.774, specificity was 0.820, balanced accuracy was 0.797, and Matthews correlation coefficient was 0.350.

‎

External Validation

In external validation set A, the AUC was 0.720 (95% CI 0.670-0.771), with a sensitivity of 0.690, a specificity of 0.882, a balanced accuracy of 0.786, and a Matthews correlation coefficient of 0.204. In external validation set B, the model achieved an AUC of 0.747 (95% CI 0.712-0.782), with a sensitivity of 0.624, a specificity of 0.843, a balanced accuracy of 0.734, and a Matthews correlation coefficient of 0.270. These results highlight the model’s good predictive performance for identifying hearing loss in older adults. The receiver operating characteristic curves are shown in Figure 4.

The nomogram’s calibration plots are shown in Figure 5. Quantitative metrics supported the visual assessment of calibration. In the external validation set A, the Brier score was 0.125 and the mean absolute error was 0.002. In the external validation set B, these values were 0.221 and 0.003, respectively. Lower Brier scores and mean absolute error values suggest that the model provides probability estimates close to the true risk, helping clinicians avoid unnecessary interventions for patients with low risk and missed diagnoses in patients participants with high risk.

Figure 6 presents the decision curves for the nomogram. The high-risk threshold probability represents the predicted probability of hearing loss, above which clinicians would consider routine audiometric screening and early interventions, such as lifestyle modifications. In the decision curve analysis plots for both cohorts, the model (red line) consistently demonstrated greater net benefit than the default strategies (treat all or none) across a conservative overlapping threshold range of approximately 5% to 75%. Using the model to guide ARHL prediction and implement routine screening and early interventions provided an additional net benefit of up to 20% compared with alternative strategies that assume all patients are either severely or nonseverely affected.

‎

‎

‎

Visualization of Predictive Models: Development of a User-Friendly Web Page

To facilitate the clinical application of our nomogram, we developed an interactive and user-friendly web application using the R Shiny framework. The application features a simple, single-panel user interface where clinicians can input a patient’s 6 predictor variables—age, education, exercise frequency, physical function, dietary habits, and hypertension—using a combination of numeric input fields and dropdown menus (Figure 7).

Upon clicking a Predict button, the tool instantly computes and displays the patient’s predicted probability of hearing loss. The output is presented as a clear percentage, along with a corresponding risk stratification (High risk or Low risk based on the 80.5% cutoff) to aid in immediate clinical decision-making. The web application is publicly accessible and requires no software installation or programming knowledge, ensuring its broad utility for primary care providers [31].

‎

Discussion

Principal Findings

The principal finding of this large, multicenter study is the development and external validation of a practical, 6-predictor nomogram for predicting hearing loss in community-dwelling older adults. Our model, which uses readily available data from routine health checkups (age, education, exercise frequency, physical function, dietary habits, and hypertension), demonstrated good discriminative ability (AUC>0.720 in all cohorts) and calibration.

A key clinical implication of this finding is the model’s potential to shift primary care from reactive secondary screening to proactive primary prevention. The online tool, developed as a proof of concept, illustrates how this model could be integrated into clinical workflows. For example, if embedded within electronic medical record systems, it could provide real-time risk scores, creating a risk-stratified screening pathway. This would allow general practitioners to prioritize formal audiometry and targeted interventions for individuals at the highest risk, potentially improving the efficiency of health care resource allocation.

Comparison With Prior Work

Our risk prediction model incorporates demographic, lifestyle, and chronic disease factors, aligning broadly with previous studies. It is well established that age is strongly associated with hearing loss. Hearing decline along with aging is primarily attributed to progressive degeneration of cochlear sensory hair cells, stria vascularis, and spiral ganglion neurons, leading to impaired sound transmission [18,32]. Consistent with previous studies, we found a higher risk of hearing loss in individuals with lower education levels [33,34]. This may be attributed to its correlation with lower socioeconomic status [35], which can limit health care access and use, ultimately leading to poorer health status. Additionally, lower education levels are often associated with unhealthy lifestyle attributes (eg, smoking, alcohol consumption, and exercise habits) that were included in our model [36]. Interestingly, while recent studies have suggested a potential link between diabetes and hearing loss [37,38], our study did not identify a statistically significant association between the two. Instead, hypertension was included as a predictor in the model. It may contribute to hearing decline by reducing blood supply to the stria vascularis, a cochlear structure dependent on terminal arteries without collateral circulation [39,40].

Recent studies have focused on a prediction model for hearing loss. Ge et al [41] developed a logistic regression model for ARHL using 18 key factors based on 401 older adults. Their model achieved an AUC value of 0.777, and the sensitivity and specificity were 70.9% and 75.3%, respectively; however, due to limited sample size and absence of internal and external validation, further scrutiny is required regarding its predictive performance. Similarly, Wang et al [42] developed a predictive model comprising 22 features. While the model achieved an AUC of 0.864 and 0.844 during training and internal validation, the large number of variables and complex operation may hinder its clinical applicability and generalizability. Predictive models should strike a balance between performance and practicality, aiming to achieve the highest possible predictive accuracy with the fewest possible variables [43].

Strengths and Limitations

The primary strengths of this study include its large, multicenter sample, which enhances the robustness and generalizability of our findings, and a rigorous variable selection process. The resulting 6-predictor model is simple, and its variables are accessible from routine physical examinations without requiring specialized tests. Furthermore, the development of an online tool highlights its potential feasibility for real-world implementation. Finally, the selection of an interpretable white-box model (logistic regression) was a deliberate choice. While more complex black-box algorithms may offer marginal gains in accuracy, their lack of transparency is a significant barrier to clinical adoption. Our model achieves an optimal balance between its robust, externally validated performance and the essential need for interpretability, making it well suited for guiding real-world clinical decisions.

However, this study is not devoid of limitations. First, our study was conducted exclusively in primary health care centers within Beijing. While we attempted to mitigate this by validating across 3 distinct centers, the shared urban setting may influence the results by limiting the model’s generalizability. Second, our model intentionally omits key risk factors such as genetic predisposition, long-term noise exposure, and ototoxic medication use. This was a necessary trade-off to create a practical tool based solely on variables routinely collected in health checkups. However, we acknowledge that the absence of these predictors may limit the model’s accuracy in populations where such factors are highly prevalent. Third, hearing loss was assessed using the whispered voice test rather than pure-tone audiometry. To mitigate this, all professionals followed a standardized protocol. However, this method may influence the results by underestimating mild hearing loss and being subject to interexaminer variability. Fourth, several predictors (eg, physical function, exercise frequency, and dietary habits) were based on self-report, which may be prone to information bias and reduce predictive precision. Finally, as a retrospective cohort study, our analysis can suggest temporal relationships between exposures and outcomes but cannot establish definitive causal inferences due to potential confounding and selection biases.

Future Directions and Conclusions

Building on this work, several avenues for future research are recommended. The immediate priority is to validate the current nomogram in broader and more diverse geographic and demographic cohorts, using pure-tone audiometry as the gold standard outcome for a more precise assessment. Future prospective studies could also enhance the model’s accuracy by incorporating key variables not available in our dataset, such as genetic predisposition, long-term noise exposure, and objective measures of physical activity. Finally, as the current online tool is a proof of concept, dedicated implementation research is required. This should address the practical challenges of integrating the tool into diverse electronic medical record systems and evaluate its real-world impact on clinical decision-making, patient outcomes, and cost-effectiveness.

In conclusion, this study provides a validated, practical tool that shows promise for identifying older adults at high risk of hearing loss. If confirmed in future studies, this nomogram could become a valuable asset in the primary prevention of ARHL.