A novel toolkit based on Artificial Intelligence and Automated Machine Learning to aid outcome prediction and decision-making in the selection of patients undergoing Mechanical Thrombectomy in Acute Ischaemic Stroke.

Abstract

Aim: To develop an Artificial Intelligence (AI) based Automated Machine Learning (AutoML) toolkit to aid decision-making for mechanical thrombectomy (MT) based on readily available patient variables that could predict functional outcome following MT. Methods: Datasets of 1097 patients from Systematic Evaluation of Patients Treated With Stroke Devices for Acute Ischemic Stroke (STRATIS) Registry and SWIFT PRIME Trial were retrospectively evaluated. Linear and non-linear models were built using an automated ML platform, DataRobot. We developed two stage models for predicting the outcome of the patient: Model 1 predicted survival, defined as an mRS score of 0-5 (alive) or 6 (dead). Model 2 predicted good/bad survivor, defined as an mRS score of 0-2 (good) or 3-5 (poor). Results: The primary outcome was the modified Rankin Scale (mRS) score at 90 days after stroke. Prediction of survival was 83% accurate (area under the curve [AUC] 0.7780). Prediction of good/poor survivor was 61% accurate (AUC 0.7061). A two-stage machine learning model has an improved 80% overall accuracy of prediction. Conclusion: The proposed AI-based AutoML toolkit evaluates various baseline clinical and radiological characteristics and predicts significant variations in treatment benefit between patients. With its improved prediction accuracy, the toolkit is clinically useful as it helps in distinguishing between individual patients who may experience benefit from Mechanical Thrombectomy treatment for acute ischaemic stroke from those who may not.