An explainable machine learning model for prognosis prediction in sudden sensorineural hearing loss under integrated therapy.

Background: Sudden sensorineural hearing loss (SSNHL) is a common otologic emergency with variable prognosis. Reliable prognostic tools, particularly for patients undergoing integrated therapy, remain scarce. Methods: This retrospective study enrolled 227 patients with unilateral SSNHL of the Qi stagnation and blood stasis syndrome who received standardized integrated therapy. Patients were partitioned into training (70%) and validation (30%) sets via outcome-based stratified sampling. To prevent information leakage, all preprocessing and feature selection were conducted strictly within the training set. Eight machine learning models were optimized through five-fold cross-validation (primary metric: AUC). Model performance was systematically evaluated across discrimination, calibration, and clinical utility, with SHapley Additive ExPlanations (SHAP) analysis applied to quantify interpretability and identify key prognostic predictors. Results: The XGBoost model demonstrated superior performance in the validation set (AUC: 0.718; 95% CI: 0.590-0.846), with robust calibration and positive clinical net benefit. SHAP analysis identified four key predictors: activated partial thromboplastin time (APTT), disease duration, platelet count (PLT), and total protein (TP). Non-linear interpretability revealed distinct prognostic thresholds: a critical intervention window was identified within 10 days of symptom onset; optimal recovery was associated with PLT counts of 200-250 × 10 Conclusion: This study developed an explainable XGBoost-based model, providing an initial framework for prognostic prediction in SSNHL patients. It demonstrates the integration of machine learning with TCM-informed therapy, suggesting potential pathways for personalized care. While promising, the results should be interpreted with caution due to the lack of external validation. Further multicenter prospective studies are crucial to confirm the model's generalizability and validate the identified biological predictors.