Identification and Patient Benefit Evaluation of Machine Learning Models for Predicting 90-Day Mortality After Endovascular Thrombectomy Based on Routinely Ready Clinical Information.

Abstract

Endovascular thrombectomy (EVT) is regarded as the standard of care for acute ischemic stroke (AIS) patients with large vessel occlusion (LVO). However, the mortality rates for these patients remain alarmingly high. Dependable mortality prediction based on timely clinical information is of great importance. This study retrospectively reviewed 151 patients who underwent EVT at Pamela Youde Nethersole Eastern Hospital between 1 April 2017, and 31 October 2023. The primary outcome of this study was 90-day mortality after AIS. The models were developed using two feature selection approaches (model I: sequential forward feature selection, model II: sequential forward feature selection after identifying variables through univariate logistic regression) and six algorithms. Model performance was evaluated by using external validation data of 312 cases and compared with three traditional prediction scores. This study identified support vector machine (SVM) using model II as the best algorithm among the various options. Meanwhile, the Houston Intra-Arterial recanalization 2 (HIAT2) score surpassed all algorithms with an AUC of 0.717. However, most algorithms provided a greater net benefit than the traditional prediction scores. Machine learning (ML) algorithms developed with routinely available variables could offer beneficial insights for predicting mortality in AIS patients undergoing EVT.