Integrating Tuned Ensemble Learning and Explainable AI for Reliable Stroke Prediction

Abstract

A stroke occurs due to a sudden interruption of blood flow to a specific region of the brain. Although extensiveresearch has already been conducted using this dataset, previous studies have not effectively addressed the issue of dataleakage during testing. The primary contribution of this study is to mitigate both the data leakage and class imbalanceproblems byapplying three different balancing techniques. We employed nine classification algorithms in this study: RandomForest (RF), AdaBoost (AB), Logistic Regression (LR), Gradient Boosting (GB), K-Nearest Neighbors (KN), Decision Tree(DT), Naive Bayes (NB), Voting (VT), and Stacking (ST). A key aspect of our approach involved using hyperparametertuning to determine the optimal configuration for each model. Additionally, we proposed a novel ensemble method, namedVoting, which combines hyperparameter-optimized LR, DT, GB, and RF classifiers. The performance of this method wascompared with other models using various evaluation metrics, including accuracy, precision, ROC curve, PR-AUC curve,and more. To enhance the interpretability of important features within the dataset, we also applied two explainable AI (XAI)techniques: Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP).