Machine learning–based risk classification of depressive symptoms among patients with hearing loss: evidence from the Health and Retirement Study (HRS)

Depression is a common and debilitating condition among individuals with hearing loss, substantially impairing patients’ quality of life. This study aimed to develop a model to classify depressive symptoms among individuals with hearing loss. Data from 2,397 patients with hearing loss in the 2016–2020 Health and Retirement Study (HRS) were analyzed. Thirty-one behavioral, health, psychological, and sociodemographic indicators were assessed, and LASSO regression was used to select predictive variables. Seven machine learning algorithms—Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), and k-nearest neighbors (KNN)—were trained to identify the optimal model. SHAP analysis was employed to provide interpretable insights, and an individualized risk stratification tool was constructed. After excluding participants with incomplete data, 1,693 patients were included in the final analysis, of whom 510 (30.1%) had depressive symptoms. Eleven predictors were retained for model development. Among the models, XGBoost demonstrated the most balanced overall performance. In the training set, it achieved an AUROC of 0.854 (95% CI: 0.832–0.877), an accuracy of 81.1%, a sensitivity of 56.0%, a precision of 76.1%, a specificity of 92.2%, and an F1 score of 64.5%. The model was specifically developed for patients with hearing loss, and the clinical benefit of interventions guided by the risk remains to be validated. In conclusion, using data from the 2016–2020 HRS database, we developed seven machine learning models to identify risk factors among patients with hearing loss. XGBoost achieved the best predictive performance and stability, while SHAP analysis provided interpretable insights into key contributors. These findings support early risk stratification and targeted prevention strategies, with potential to enhance mental health care for this vulnerable population. • HRS 2016–2020 hearing-loss cohort analyzed (n=1,693; 30.1% depressed). • Seven ML models compared; XGBoost showed the most robust performance. • XGBoost AUROC 0.854 on training set and 0.825 on test set. • Top drivers: sleep trouble, physical function limits, poor self-rated health status, life satisfaction. • Calibration, Brier, and decision-curve analyses support clinical utility. • SHAP plots enable patient-level interpretability and transparent risk attribution.