Comparison of different machine learning approaches to model stroke subtype classification and risk prediction

Abstract

The application of computer-aided techniques regarding stroke diagnosis is especially important in non-urban areas, because of the lack of adequate resources. The purpose of this research consists of testing the hypothesis that state-of-the-art machine learning-based modeling techniques, when integrated with non-invasive monitoring technologies, can help on the diagnosis of the type of stroke only few minutes after the crisis. This method has been also tested to predict future risks, like the eventual death of the patient. The collected dataset comprised the medical records of 119 patients with 7 predictors and two target variables: diagnosis of stroke type and prediction of death. 7 different algorithms have been employed, evaluated over 6 different metrics. 10-fold cross validation resampling method was utilized. Random Forest models produced the best performance in stroke diagnosis and death prediction compared to the other algorithms, with average values of 0.93 ± 0.03 and 0.97 ± 0.01, respectively.