Evaluating Machine Learning Models for Heart Stroke Prediction: a Hybrid Feature Selection Approach

According to recent studies, heart diseases are proven to have a major share of death globally. Predicting the disease in early stage plays a main role in reducing mortality rate. Recent advancements in the field of Machine learning(ML) have shown better results in early detection of the disease. Our paper aims to build a predictive ML model for predicting the heart stroke using a benchmark dataset related to heart disease. The dataset we used includes various health parameters that influence stroke risk. We applied a hybrid feature selection which is a combination of low variance filter and PCA to get the dominant features and applied several classification algorithms namely Logistic Regression, Nearest Neighbors, Decision Tree, XGBoost and the Random Forest algorithm Among the five models experimented in this study, Random Forest gives the highest accuracy of 96.68% but has good predictive capability. In our analysis, distance-based models such as KNN have high sensitivity to the reduction of dimensions, while ensemble models such as Random Forest and XGBoost remain strong even with decreased features. Our results indicate that using ensemble methods along with SMOTE and proper feature selection is very suitable for stroke prediction. This is part of future work and research into deep models for further improvement in performance.