Stroke Risk Prediction Using CatBoost with an Explainable Artificial Intelligence Approach

Abstract

Stroke is among the main causes of death worldwide. According to the World Health Organization (WHO), strokes, including ischemic and hemorrhagic, account for around 11% of global mortality. Therefore, early prediction is crucial as part of efforts to prevent the risk of stroke and to assist healthcare professionals in clinical decision-making. This work aims to develop a stroke risk prediction model using the CatBoost algorithm, and to interpret the prediction results using an Explainable Artificial Intelligence (XAI) approach through the SHAP method. The CatBoost model's evaluation results demonstrate strong performance, with AUC = 0.98, an F1-score = 0.91, precision = 0.92, recall = 0.90, and accuracy of 0.93. Furthermore, the XAI analysis utilizing SHAP showed that the CatBoost model not only delivers highly accurate predictions but also successfully identifies the most relevant features leading to stroke risk, namely age, body mass index (BMI), and mean level of glucose. Finally, a comparative examination with various different machine learning models demonstrates that the CatBoost model obtains the best performance and is extremely useful in predicting stroke risk.