Abstract
<title>Abstract</title> <p>Background Sudden sensorineural hearing loss (SSNHL) is a highly heterogeneous condition with unpredictable recovery. We aimed to develop and validate an interpretable, subtype-specific machine learning (ML) framework for prognostic assessment using the largest SSNHL clinical cohort reported to date to overcome the "black-box" nature of existing models. Method In this retrospective cohort study, 3,957 patients diagnosed with SSNHL between 2010 and 2017 were analyzed. Eight ML algorithms were trained and validated for five predefined audiogram subtypes: ascending, descending, atypical, profound, and flat. Model performance was evaluated using accuracy, area under the receiver operating characteristic curve (AUC), and Brier score. SHapley Additive exPlanations (SHAP) and partial dependence plots (PDPs) were used to interpret the models and quantify the nonlinear impact of predictors. Results In total, 3,957 patients were included in the analysis. The subtype-specific models demonstrated robust performance with test accuracies ranging from 0.706 to 0.774. Support Vector Machine and Logistic Regression emerged as the optimal classifiers. Five core prognostic factors were identified: disease duration, pure-tone average, age, WHO hearing classification, and systemic glucocorticoid therapy. Notably, the study quantified subtype-specific treatment windows, showing the disease duration threshold for a 50% recovery probability ranged from 7.5 to 13.1 days. Conclusion ML is a reliable tool for high-precision prognostic assessment in SSNHL. By quantifying actionable intervention thresholds, this interpretable framework moves beyond "black-box" predictions, providing a real-time decision support tool that reinforces the necessity of early systemic glucocorticoid therapy.</p>