Model Prediksi Penyakit Stroke Menggunakan Explainable Neural-XGBoost Berbasis SHAP

Abstract

Stroke is a major health concern that requires accurate prediction to support early prevention and intervention. This study aims to develop a stroke prediction model using a Hybrid Deep Neural Network–XGBoost approach integrated with SHapley Additive exPlanations (SHAP)-based Explainable Artificial Intelligence (XAI). The research procedures include data preprocessing, class imbalance handling using the Synthetic Minority Over-sampling Technique (SMOTE), feature extraction using a Deep Neural Network (DNN), classification using XGBoost, and model interpretation using SHAP. Model performance was evaluated using stratified 5-fold cross-validation based on accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC). The proposed model was compared with six benchmark models: Logistic Regression, Decision Tree, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, and baseline XGBoost. The results show that the Hybrid DNN–XGBoost model achieved an accuracy of 64.46% ± 0.80%, precision of 64.55% ± 0.87%, recall of 64.17% ± 0.76%, F1-score of 64.36% ± 0.76%, and ROC-AUC of 69.72% ± 0.99%. Its accuracy exceeded that of all benchmark models, which ranged from 48.96% to 50.62%. SHAP analysis identified DeepFeature_28, DeepFeature_31, DeepFeature_5, DeepFeature_6, and DeepFeature_26 as the features contributing most to the model's predictions. These features were associated with major clinical risk factors, including age, blood pressure, glucose levels, and body mass index (BMI). These findings indicate that the Hybrid DNN–XGBoost approach improves predictive performance compared with the benchmark models while providing insights into feature contributions to stroke predictions.