Predicting Stroke Risk Using SMOTE and ROS Machine Learning Techniques

The World Stroke Organization has reported that globally 12.2 million people have a stroke for the first time, and 90% of them were related to risk factors that could have been modified. Despite the abundance of machine learning models developed to predict stroke, their performances were unsatisfactory due to low recall values. The aim of this study is to implement a machine learning pipeline to predetermine the likelihood of a person getting a stroke with a high recall and AUROe. An electronic health dataset with 5110 patients, including 249 stroke cases and 4861 non-stroke cases, was preprocessed by deleting unwanted data, imputing missing data, and encoding. Then, the dataset was balanced using the synthetic minority oversampling technique (SMOTE) and random oversampling (ROS) technique. Machine learning algorithms were developed using Python and Scikit-Iearn libraries. Multilayer perceptron, decision tree, random forest, extreme gradient boost, and naive Bayes algorithms were trained and tested. The random forest (ROS) model outperformed the other models with a recall of 0.67 and AUROe of 0.84. The extreme gradient boost (ROS) model performed the second best with a recall of 0.57 and AUROC of 0.84. The key features to predict stroke are age, average glucose level, body mass index, hypertension, and heart disease. These results exhibit the potential of utilizing machine learning algorithms for stroke prediction and highlight the importance of feature engineering for improved performance.