An interpretable machine learning approach for predicting high-frequency hearing loss risk in occupational workers

Abstract Background High-frequency hearing loss (HFHL) is prevalent among noise-exposed workers, yet routine screening remains costly. This study develops and validates an interpretable machine learning model for predicting HFHL risk, aiming to provide a cost-effective tool for early detection and targeted intervention. Methods A retrospective analysis was conducted on occupational health records from 5,037 workers exposed to noise. Demographic data, occupational exposure history, and laboratory variables were analyzed. Multiple machine learning models were compared, with CatBoost selected due to its superior performance. Results The CatBoost model achieved an AUC of 0.76, with sensitivity of 0.71 and specificity of 0.68. Age, noise exposure, and red blood cell count were the most influential predictors. SHAP analysis provided individualized risk profiles, facilitating personalized risk assessment. Conclusions This interpretable machine learning model offers robust accuracy in predicting HFHL risk, supporting cost-effective screening and personalized occupational health strategies.