基于机器学习与SHAP可解释性的突发性聋预后预测模型构建及验证OA

Objective To construct machine learning-based prognostic models for sudden sensorineural hearing loss(SSNHL)to support clinical decision-making.Methods Clinical data from 1120 SSNHL patients(2020～2024)were retrospectively collected from two centers and divided into training(784 cases)and validation sets(336 cases).Key predictors were selected via LASSO regression,and four models-support vector machine(SVM),XGBoost,logistic regression(LR),and random forest(RF)-were developed.Model performance was evaluated using ROC curves,calibration curves,decision curve analysis(DCA),and shapley additive explanations(SHAP).Results Four predictors were identified:audiogram type,age,onset-to-treatment time,and vertigo.In the training set,RF and XGBoost achieved superior AUC(0.804,95%CI:0.774～0.835)compared to SVM(0.722)and LR(0.722)(P＜0.001).Validation set results confirmed robustness(RF AUC=0.773;XGBoost AUC=0.769).SHAP analysis revealed that age＞49.5 years,onset-to-treatment time＞5.5 days,profound hearing loss,and vertigo significantly increased prognostic risk.Calibration and DCA curves demonstrated high consistency between predicted and observed outcomes,with optimal clinical net benefit at 25%～50%threshold probabilities.Conclusions RF and XGBoost models accurately predict SSNHL prognosis,and SHAP-based interpretability enhances clinical utility.