Development and Internal Validation of Machine-Learning Models for Individualized Prediction of Hearing Recovery in Sudden Sensorineural Hearing Loss

Objective: To develop and internally validate a pragmatic machine-learning model for predicting hearing recovery in sudden sensorineural hearing loss (SSNHL) using routinely available baseline clinical variables and standard audiometry. Study design: Retrospective prognostic study. Setting: Tertiary teaching hospital in China. Patients: Adults with idiopathic SSNHL evaluated between January 1, 2017, and August 31, 2025, who had pretreatment and follow-up audiometry. The final cohort included 1097 patients contributing 1205 affected ears. Intervention(s): Routine clinical management, including systemic and/or intratympanic corticosteroid therapy according to contemporary clinical practice and physician judgment. Main outcome measure(s): Hearing recovery on follow-up audiogram classified by the 2015 Chinese guideline and Siegel criteria and dichotomized at clinically relevant thresholds. Primary model selection was based on overall performance across the 3 Chinese guideline-defined binary outcomes. Results: Six algorithms were compared using patient-level 5-fold cross-validation, within-fold imputation, calibration assessment, decision-curve analysis, and SHAP interpretation. Tree-based models outperformed logistic models overall. In the primary models excluding treatment variables, LightGBM showed consistent performance across Chinese guideline-defined outcomes. ROC AUCs were 0.873 for full recovery, 0.797 for important recovery and better, and 0.753 for minor recovery and better, with corresponding accuracies of 89.07%, 78.99%, and 69.87%. Under Siegel criteria, LightGBM achieved ROC AUCs of 0.889 for complete recovery, 0.807 for partial recovery and better, and 0.782 for slight improvement and better. Performance was similar in sensitivity models incorporating treatment variables. The most influential predictors were age, baseline hearing thresholds, audiogram configuration, and vertigo. Conclusions: A LightGBM model based on easily collected baseline clinical and audiometric variables showed useful internal performance for predicting hearing recovery in SSNHL and may support early, individualized prognostic assessment, pending external and prospective validation.