Regression, classification and ensemble machine learning approaches to forecasting clinical outcomes in ischemic stroke

Abstract

We applied different machine learning approaches to predict (forecast) the clinical outcome, measured by the modified Rankin Scale (mRS) score, of ischemic stroke patients 90 days after stroke. Regression, multinomial classification, and ordinal regression tasks were considered. M5 model trees followed by bootstrap aggregating as a meta-learning technique produced the best regression results. The same regression technique when used for classification after discretization of the target attribute also performed better than regular multinomial classification. For the ordinal regression task, the logit link function (ordinal logistic regression) outperformed the alternatives. We discuss the methodology used, and compare the results with other standard predictive techniques. We also analyze the results to provide insights into the factors that affect stroke outcomes.