Comparing the Performance of Regression Models, Random Forests and Neural Networks for Stroke Patients’ outcome Prediction

The paper aimed to explore the performance of three classification methods which were used for the prediction of the clinical outcome of patients after an acute ischemic stroke. We performed the analysis with the logistic regression, the random forests algorithm and the multilayer neural networks. Each method can be evaluated with respect to its flexibility, model interpretability or prediction accuracy. In this paper, we focused on the investigation of properties of the methods with respect to our major problem, i.e. prediction of stroke patients' outcome.For several years, the University Hospital Ostrava in the Czech Republic has been systematically collecting various information about patients who were transported there after suffering from an acute ischemic stroke. Therefore, efficient statistical analysis of the collected data is a vital part of the research. Naturally, the primary interest lies in the analysis of factors affecting the clinical status of patients three months after stroke which was evaluated with widely-used modified Rankin Scale. For the analysis, dichotomized clinical status was used as a target variable. By meaning, the values correspond with the values 0-3 and 4-6 of the modified Rankin Scale. All three models provided satisfactory results. The final binary logistic regression model reached the overall classification accuracy 87%, the random forests models had classification accuracy almost 90% and the multilayer neural networks almost 80% accuracy. The paper also compared the performance with other measures, e.g. well-known sensitivity, specificity or precision.