Balancing and metaheuristic techniques for improving machine learning models in brain stroke prediction

Abstract

A brain stroke, medically referred to as a stroke, represents a critical condition triggered by the disruption of blood flow to a region of the brain. Early detection of stroke is crucial to prevent fatal complications. In this study, we worked with an unbalanced dataset of 4981 entries on stroke, which we balanced using the K-means synthetic minority over-sampling technique (KMeansSMOTE) algorithm. We then employed five machine learning algorithms: decision tree, random forest, support vector machine, K-nearest neighbors, and gradient boosting. We compared the hyperparameter optimization of these algorithms using four metaheuristic techniques: gray wolf optimization, particle swarm optimization, genetic algorithm, and artificial bee colony. The models' effectiveness was evaluated using multiple metrics, such as accuracy, recall, precision, F1-score, and area under the receiver operating characteristic curve. Our findings indicate that the random forest optimized by the genetic algorithm achieved the best performance, with an accuracy of 97.39% and an F1-score of 97.35%. This study highlights the effectiveness of balancing and metaheuristics techniques in optimizing machine learning models for stroke forecasting.