Abstract
Background: Hearing loss is the most common sensory loss in humans and carries an enhanced risk of depression. No prior studies have attempted a contemporary machine learning approach to predict depression using subjective and objective hearing loss predictors. Objective: To deploy supervised machine learning to predict scores on a validated depression scale using subjective and objective audiometric variables and other health determinant predictors. Design: We used a large predictor set of health determinants to forecast individuals’ scores on a validated instrument to screen for the presence and severity of depression (PHQ-9). After model training, the relative influence of individual predictors on depression scores was stratified and analyzed. Setting: National Health and Nutrition Examination Survey (NHANES) 2015-2016 database. Patients: Adult survey participants. Measurements: Model prediction error performance. Results: The test-set mean absolute error was 3.03 (CI 95%: 2.91-3.14) and 2.55 (CI 95%: 2.48-2.62) on datasets with audiology-only predictors and all predictors, respectively, on the PHQ-9’s 27-point scale. Participants’ self-reported frustration when talking to members of family or friends due to hearing loss was the fifth-most influential of all predictors. Of the top ten most influential audiometric predictors, five were related to social contexts, two for significant noise exposure, two objective audiometric parameters, and one presence of bothersome tinnitus. Conclusions: Machine learning algorithms can accurately predict PHQ-9 depression scale scores from NHANES data. The most influential audiometric predictors of higher scores on a validated depression scale were social dynamics of hearing loss and not objective measures. Such models could be useful in predicting depression scale scores at the point-of-care in conjunction with a standard audiologic assessment.
Introduction
Hearing loss is the most prevalent sensory deficit in the human population and has long been recognized as a significant risk factor for depression. Despite this known correlation, clinical interventions—specifically the use of hearing aids—have shown inconsistent results in alleviating depressive symptoms. This discrepancy suggests that traditional clinical metrics, such as the pure-tone average (PTA), may be insufficient for capturing the full psychological impact of hearing impairment.
Traditional research often relies on objective measures of hearing sensitivity in controlled environments. However, these metrics fail to account for the "total hearing loss experience," which encompasses social isolation, communication frustration, and environmental challenges. This study explores whether machine learning (ML) can bridge this gap by integrating objective audiometric data with subjective, functional reports to more accurately predict depression scores.
Methods
Data Source and Cohort
The study utilized data from the National Health and Nutrition Examination Survey (NHANES) 2015-2016 database. The final cohort consisted of 4,582 adult respondents who completed a comprehensive battery of audiometric testing and the Patient Health Questionnaire-9 (PHQ-9), a validated 27-point depression screening tool.
Machine Learning Framework
The researchers employed LightGBM, a gradient boosting decision tree algorithm known for its high performance and efficiency. To ensure the model's interpretability, SHAP (SHapley Additive exPlanations) values were used to rank the influence of individual variables on the model's predictions.
The study was designed as a regression task to predict PHQ-9 scores using two distinct feature sets:
- All Predictors: 151 variables including demographics, medical history, and all hearing-related data.
- Audiology-Only: A subset focusing exclusively on objective tests (e.g., frequency thresholds, eardrum pressure) and subjective hearing questions.
To maintain robustness and prevent overfitting, a nested cross-validation approach was implemented to separate hyperparameter tuning from the final performance evaluation.
Results
The ML models demonstrated high accuracy in predicting depression scores on the 27-point PHQ-9 scale. The All-Predictor Model achieved a Mean Absolute Error (MAE) of 2.55, while the Audiology-Only Model achieved an MAE of 3.03.
Feature Importance and Social Dynamics
The most significant finding was the dominance of functional and social variables over objective clinical measurements. In the audiology-only model, five of the top ten predictors were related to social contexts. Most notably, a participant's self-reported frustration when communicating with family or friends was the primary predictor in the audiology-only model and the fifth most influential predictor in the comprehensive model—outranking many traditional medical and demographic markers.
Other influential audiometric predictors included:
- Significant noise exposure history.
- Presence of bothersome tinnitus.
- Objective parameters such as tympanometric width and compliance (indicators of middle ear function).
Discussion
The results underscore a critical shift in understanding the link between hearing and mental health: social dynamics are more predictive of depression than decibel levels. The high ranking of "frustration" suggests that the psychological burden of hearing loss stems less from the inability to hear quiet sounds and more from the breakdown of meaningful social connections.
This insight may explain why standard hearing aids do not always improve mental health outcomes. If a device amplifies sound but fails to improve speech clarity in noisy social environments (the "cocktail party effect"), it may not alleviate the specific stressors that drive depression. Furthermore, the inclusion of objective measures like tympanometric width in the top predictors suggests a complex interplay between physiological ear function and psychological well-being that warrants further investigation.
Conclusion
This study demonstrates that machine learning can accurately predict depression scores using national health data, highlighting the profound impact of the social experience of hearing loss. These models offer a potential pathway for "point-of-care" depression screening within audiology clinics. By identifying patients whose hearing loss is causing significant social frustration, clinicians can provide more targeted referrals to mental health services and prioritize rehabilitation strategies that focus on social reintegration rather than just auditory threshold improvement.
Summary updated