Addressing the paradox of poor outcomes despite successful recanalization in stroke

Project Summary In recent years, endovascular stroke therapy (EVT) has emerged as a highly effective treatment for large vessel occlusions (LVO) acute ischemic stroke (AIS). However, while our ability to achieve recanalization in these patients is exceeding 90%, still over 50% of these patients are left with moderate to severe disability. In this

project we propose a novel classification for this under-studied population, which categorizes these patients based on etiology of persistent dysfunction. We hope to generate an improved understanding of this gap between successful recanalization and good clinical outcomes, to identify the cohort of patients in whom additional

measures may be necessary, and to create a machine-learning (ML) tool that will implement this improved understanding as clinical decision support tool. To do this, we propose the following Aims: Aim 1: We will create a longitudinal clinical and imaging dataset and evaluate more completely the role of final

infarct volume (FIV) on outcome. We will create a multi-center dataset including imaging and clinical data from 3 large hospital systems across the US, with pre-procedure and follow up imaging at 24-48 hour and 5-14 day time points in patients with LVO AIS treated with EVT. We will then use infarct volume and location information

from early and late follow up MRIs to predict likelihood of good outcome from FIV data. Aim 2: We propose, derive, and validate a novel classification system for patients with disability despite reperfusion and generate machine learning models to identify them based on data obtained prior to EVT. The

model will be trained using clinical and imaging data and tested in independent cohorts, which will be drawn from multiple hospitals to better represent a diverse population. Aim 3: We will create deep learning models using raw patient-level cross-sectional and angiographic imaging data to identify patients with persistent disability despite reperfusion, and their subtypes. We will leverage our

longitudinal-sensitive deep learning models to learn from the change between pre- and immediate post-EVT with imaging and clinical variables to predict clinical outcomes in patients treated with EVT. Completion of these aims will provide a means to characterize and identify patients with poor clinical outcomes

despite successful recanalization. These results will establish the foundation for future clinical trials by discovering which clinical and/or imaging features are key drivers of clinical outcomes. We will provide a positive impact by creating a simple, universally accessible method to identify the most vulnerable patients who may

need additional treatments beyond EVT for LVO AIS.