Abstract WP530: A Machine Learning Approach in Prediction of Recurrent Stroke

Abstract

Introduction: Although recurrent stroke is clinically important, prediction of recurrent stroke is less well known. Machine learning encompasses a body of statistical approaches that can detect complex interaction patterns from multi-dimensional data. ML is gradually being adopted in medical science. The purpose of this study was to illustrate the possibility of using machine learning as an aid in prediction of recurrent stroke. Methods: We obtained clinical data from a consecutively registered, hospital-based acute stroke registry. Using a dataset of 4,766 acute ischemic stroke patients, predictive models were built using 1) linear 2) non-linear/non-parametric machine learning algorithms. The accuracy obtained was compared to that of the baseline model built without the clinical data. Results: There were 810 patients with recurrent stroke after ischemic stroke or transient ischemic attack (median follow-up, 365 days). The performances of the various machine learning models were compared. The accuracy of the baseline model was 31.4%, but the best non-linear model showed an accuracy of 64.6%, which was significantly higher than that of non-informative random guessing (p=0.002). Overall, the models using the nonlinear algorithms showed better accuracy than the linear ones. Conclusions: The high performance of the developed models demonstrated the predictive capacity of recurrent stroke and justified the application of machine learning algorithms. However, the practical application of machine learning models may need refinement and larger-scale data collection.