ML-Based Stroke Detection Model using Different Feature Selection Algorithms

Abstract

Stroke occurs in the brain due to the blockage of blood flow carrying oxygen and nutrients or due to sudden bleeding within the brain. Delaying stroke treatment can lead to serious consequences, including death. This paper proposes a model based on classification algorithms in machine learning to detect whether a stroke has occurred. The classification algorithms used in this study are k-nearest neighbours, decision tree, random forest, naïve Bayes, multilayer perceptron and support vector machine. These algorithms were applied to the classification task using different feature selection methods, namely: all features, select K best (SelectKBest), select percentile (SelectPercentile), select false-discovery rate (SelectFdr), select false-positive rate (SelectFpr) and select family-wise error (SelectFwe). This paper compares the performance of the above algorithms using the different feature selection methods to determine which algorithm provides the best classification results in terms of accuracy, recall, precision and F1-score. The decision tree algorithm shows the highest performance in accuracy, precision and F1-score, regardless of the feature selection method used. Both decision tree and random forest yield the highest and identical recall results when the ‘all features’ selection method is applied. For the other feature selection methods, decision tree consistently provides the highest recall results. Performance evaluation was conducted by comparing the proposed model to the most relevant works using different machine learning algorithms. The results indicate that the proposed model outperforms other approaches, particularly with the decision tree algorithm. Statistical results, including means, standard deviations and 95% confidence intervals for all features and the target variable in the stroke dataset, were obtained. Trade-offs between precision and recall results for the compared algorithms are also presented.