Decision tree and artificial immune systems for stroke prediction in imbalanced data

Although cerebral stroke is a important public worldwide health problem with more than 43 million global cases reported recently, more than 90% of metabolic risk factors are controllable. Therefore, early treatment can take advantage of a fast and low-cost diagnosis to minimize the disease’s sequels. The use Machine Learning (ML) techniques can provide an early and low-cost diagnosis. However, the performance of these techniques is reduced in problems of prediction of rare events and with class imbalance. We proposed Machine learning approach to cerebral stroke prediction based on Artificial Immune Systems (AIS) and Decision Trees (DT) induced via Genetic Programming (GP). In general, the approaches for stroke prediction presented in the literature do not allow the development of models considered interpretable; our approach, on the other hand, uses a simplification operator that reduces the complexity of the induced trees to increase their interpretability. We evaluated our approach on a highly imbalanced data set with only 1.89% stroke cases and used AIS combined with One Sided Selection (OSS) to create a new balanced data set. This new data set is used by the GP to evolve a population of DTs, and, at the end of this process, the best tree is used to classify new instances. Two experiments are used to test the proposed approach. In the first experiment, our approach achieved, in terms of sensitivity and specificity, are 70% and 78%, respectively, indicating its competitiveness with the state-of-the-art technique. The second experiment evaluates the proposed simplification mechanism in creating rules that can be interpreted by humans. The proposed approach can effectively increase sensitivity and specificity while maintaining accurate prediction using interpretable models, indicating its potential to be clinically used in stroke diagnosis.