Development and external validation of an interpretable machine learning model for high-frequency hearing loss screening

Early identification of high-frequency hearing loss (HFHL) remains difficult in community settings because conventional audiometric screening requires specialized equipment and trained personnel. We aimed to develop, interpret, and externally validate a machine-learning model using routinely collected information to support HFHL risk screening in Chinese adults. Data from 13,950 adults participating in the 2023–2025 China National Health Survey were randomly divided into training and internal validation sets (8:2). Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, and eight machine-learning algorithms were compared. Discrimination, classification performance, and probabilistic accuracy were evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, F1 score, specificity, and Brier score. SHapley Additive exPlanations (SHAP) were used for interpretation. Two sensitivity analyses assessed model performance for isolated HFHL and HFHL concurrent with speech-frequency hearing loss. External validation was performed in an independent hospital-based cohort of 223 adults. HFHL was present in 9,779 of 13,950 participants (70.10%). LASSO retained 11 predictors, with age having the largest coefficient. XGBoost showed the highest AUC (0.8728; 95% CI, 0.8589–0.8860) and the lowest Brier score (0.1276) among the eight models. In sensitivity analyses, the AUC was 0.7869 for isolated HFHL versus normal hearing and 0.9063 for concurrent HFHL and speech-frequency hearing loss versus normal hearing. SHAP analysis identified age, menopausal status, sex, and educational level as the leading contributors. In external validation, XGBoost achieved an AUC of 0.9196, a Brier score of 0.1126, an accuracy of 86.55%, a sensitivity of 81.52%, and a specificity of 90.08%. An interpretable XGBoost model based on routinely available variables showed good discrimination and probabilistic accuracy for HFHL screening, including in an independent hospital cohort. The accompanying web tool may support preliminary risk stratification and referral for formal audiometry.