Predicting Depression From Hearing Loss Using Machine Learning

Matthew G. Crowson, Kevin H. Franck, Laura C. Rosella, Timothy C. Y. ChanEar 2021 · DOI

Abstract

Machine learning algorithms can accurately predict PHQ-9 depression scale scores from National Health and Nutrition Examination Survey (NHANES) data. The most influential audiometric predictors of higher scores on a validated depression scale were social dynamics of hearing loss rather than objective audiometric testing. Such models could be useful in predicting depression scale scores at the point-of-care in conjunction with a standard audiologic assessment.

Introduction

The correlation between hearing loss and depression is well-documented, yet the underlying relationship remains non-linear and difficult to quantify. Clinical observations often show that improving a patient's hearing thresholds through amplification does not consistently mitigate depressive symptoms. This discrepancy suggests that objective hearing ability (how well one hears) and the subjective hearing experience (how hearing loss impacts life) are distinct contributors to mental health. This study leverages supervised machine learning to identify which specific aspects of the hearing loss experience most accurately predict depression severity.

Methodology

The researchers utilized the LightGBM algorithm, a gradient-boosting decision tree framework, to analyze data from the 2015-2016 National Health and Nutrition Examination Survey (NHANES). The final cohort included 4,582 adult respondents who underwent comprehensive audiometric testing.

The modeling pipeline included:

  • Feature Engineering: 151 variables were processed, encompassing demographics, medical history, and both objective and functional audiometric data.
  • Preprocessing: Missing values were imputed, and variables with high multicollinearity ($>0.90$) were removed. Data normalization was achieved through Yeo-Johnson and Box-Cox power transformations.
  • Validation: A nested cross-validation approach (10-fold inner loop for hyperparameter tuning; 5-fold outer loop for evaluation) ensured model robustness.
  • Explainability: SHAP (Shapley Additive exPlanations) values were used to determine the relative importance of each predictor.

Results

The predictive performance was measured using Mean Absolute Error (MAE) on the 27-point PHQ-9 scale:

  • Comprehensive Model (All Predictors): MAE of 2.55 (95% CI: 2.48–2.62).
  • Audiology-Only Model: MAE of 3.03 (95% CI: 2.91–3.14).

In the audiology-only model, five of the top ten predictors were related to social contexts. Only two objective parameters—tympanometric width and compliance—appeared in the top ten, highlighting the dominance of subjective experience over clinical measurements.

Discussion: The "Frustration Factor"

The most significant insight from the SHAP analysis was the weight of subjective social frustration. A participant’s self-reported frustration when communicating with family or friends was the fifth most influential predictor overall and the primary predictor in the audiology-specific model.

This finding provides a potential explanation for why hearing aids sometimes fail to alleviate depression: while they improve decibel thresholds, they may not fully resolve the complex social dynamics or the "cacophony" of noisy environments that lead to social isolation. The results suggest that "point-of-care" screening for depression could be effectively integrated into standard audiologic assessments by focusing on a few high-impact questions regarding social function.

Conclusion and Future Work

This study demonstrates that machine learning can effectively predict depression scores by prioritizing the functional and social dimensions of hearing loss. While the cross-sectional nature of the NHANES data precludes causal conclusions, the findings advocate for a shift in aural rehabilitation. Future research will focus on inferential analyses to determine if rehabilitation strategies targeting social dynamics, rather than just sound amplification, yield better mental health outcomes.

Summary updated

Abstract

Machine learning algorithms can accurately predict PHQ-9 depression scale scores from National Health and Nutrition Examination Survey data. The most influential audiometric predictors of higher scores on a validated depression scale were social dynamics of hearing loss and not objective audiometric testing. Such models could be useful in predicting depression scale scores at the point-of-care in conjunction with a standard audiologic assessment.
BibTeX
@inproceedings{crowson2021predicting,
  author = {Matthew G. Crowson and Kevin H. Franck and Laura C. Rosella and Timothy C. Y. Chan},
  title = {{Predicting Depression From Hearing Loss Using Machine Learning}},
  booktitle = {Ear},
  year = {2021},
  doi = {10.1097/aud.0000000000000993},
}
Figure