Explainable machine learning characterizes nonlinear age-dependent systemic contributions to sensorineural hearing loss: derivation and external validation

Age-related Sensorineural Hearing Loss (SNHL) occurs within the broader context of systemic aging and is associated with metabolic abnormalities, renal dysfunction, and chronic low-grade inflammation. However, the nonlinear age-related patterns of these systemic factors remain incompletely characterized. This study used Explainable Artificial Intelligence (XAI) to characterize systemic features associated with age-related SNHL and estimate age-related breakpoints in their model contributions. This cross-sectional study included 1,460 participants from the National Health and Nutrition Examination Survey (NHANES), 2003–2020. Participants were classified as having SNHL or normal hearing. LASSO regression and Boruta were combined for feature selection from more than 40 clinical indicators, and Extreme Gradient Boosting (XGBoost) was used for prediction. SHapley Additive exPlanations (SHAP) and segmented regression were applied to characterize nonlinear age-related changes and estimate corresponding breakpoints. An independent clinical cohort ( N = 251) was used for external evaluation. XGBoost showed good discrimination in the derivation analysis (Area Under the Curve [AUC] = 0.890). Seven predictive features were retained: age, sex, marital status, diabetes history, fasting glucose, serum creatinine, and the Systemic Inflammation Response Index (SIRI). Three separate feature-specific segmented-regression analyses identified estimated age-related breakpoints in the SHAP contributions of fasting glucose, serum creatinine, and SIRI at 62.56, 69.45, and 76.00 years, respectively. In 1,000 bootstrap iterations, all models converged successfully; the corresponding percentile 95% confidence intervals were 53.00–72.02, 58.00–72.00, and 69.81–78.00 years, respectively. The model retained good discrimination in the independent clinical cohort (AUC = 0.844). The predictive contributions of selected metabolic, renal, and inflammatory features to SNHL varied nonlinearly with age. XAI combined with segmented regression provided a quantitative framework for characterizing these age-dependent model patterns and estimating their associated breakpoints. These findings may help refine age-informed SNHL risk assessment and provide a basis for future longitudinal validation.