Image-based Predictive Modeling of Clinical Outcome in Acute Ischemic Stroke using Machine Learning

Abstract

Acute ischemic stroke is caused by a blockage in the arteries causing hypo-perfusion and subsequent death of brain regions. With a loss rate of approximately 1.9 million neurons per minute in a stroke patient, there is a need for administering the appropriate treatment as fast as possible. After the patient is discharged from the intensive care unit, patient-specific rehabilitation management is necessary to ensure optimal outcome. Within this context, a detailed understanding of the importance of different brain regions for different clinical outcomes is required for acute treatment decision-making and optimal rehabilitation therapy. Therefore, the primary objectives of this work are: (i) to improve the current understanding of brain-behaviour relationships by advancing existing lesion-symptom mapping techniques; and (ii) to explore the utility of lesion-symptom mapping for predictive modelling of long-term stroke outcomes. Statistical learning, medical image analysis, and machine learning techniques are employed to develop improvements in defining structure-function relationships, quantifying burden of lesion across brain regions, and integrating these image-based features with clinical descriptors of a patient’s medical condition into a predictive model. Specifically, a multi-modal stroke-specific brain atlas for the elderly is developed, the relative importance of covariates in a lesion-symptom mapping analysis are investigated, lesion-symptom mapping techniques are employed with sub-score information to reveal hidden category-specific structure-function relationships, a novel metric to quantify the structural integrity of white matter tract is proposed, and finally, the predictive modelling of long-term stroke outcomes with and without the use of lesion-symptom mapping is explored in a nested regression framework. This research presents a general framework for predictive modelling in patients with brain lesions that includes neuroimaging and clinical information. It can be adapted for various time points along the standard stroke care protocol (treatment support in-hospital or rehabilitation support post-discharge). Generally, the methods proposed in this work could also be applied to many other problems in the neuroimaging and precision medicine communities. Overall, the proposed methods have the potential to improve patient care by enabling more precise patient-specific treatment and disease management for stroke.