CGGF-NFHSTA Framework for Imbalance-Aware Stroke Risk Prediction with Explainable Hybrid Sampling

Abstract

Stroke is a leading cause of death and disability and early risk prediction tools that can support clinical decision making are needed. But the problem of class imbalance, lack of adequate pre-processing and lack of interpretability impact stroke prediction using machine learning. To solve these problems, this study introduces a framework, called CGGF-NFHSTA, for stroke risk prediction based on structured clinical information and taking imbalance into consideration. The framework involves the preprocessing stage, the hybrid approach for handling data imbalance with SMOTE, Tomek Links, and ADASYN and the proposed Curvature-Guided Gradient Forest (CGGF) classifier. To evaluate robustly, 5-fold cross validation is used and SHAP analysis is used to make the results interpretable and shed light on the most important stroke risk factors. As a result of the experiments conducted, the accuracy, precision, recall, F1 and roc-auc of the proposed framework are observed as 98.9%, 99.1%, 98.6%, 98.8% and 0.99 respectively, which outperform the conventional classifiers. The proposed framework offers a reliable and comprehensible solution to stroke risk prediction while taking into account imbalance.