Predicting Cognitive Outcomes from Stroke Using Lesion Location

About This Grant

Abstract Having a stroke is a frightening experience for the individual and their loved ones. A major source of anxiety is not knowing what stroke-related deficits will persist – will my loved one ever be able to talk again or understand me? The likelihood for recovery is estimated by the treating neurologist or rehabilitation specialist based on their personal experience. Much of the variance in outcomes depends on the location of the stroke in the brain. Currently, however, there are no tools available that can utilize this lesion location information for the purposes of improving the accuracy of prognostication. The focus of our research program is developing such a tool. Specifically, we have developed a method that uses the location of the stroke, queries it against outcome data from hundreds or thousands of other individuals, and generates a personalized, quantitative prediction of long- term stroke-related cognitive deficits. At the University of Iowa, we have one of the most comprehensive ‘lesion’ registries in the world. In patients with ischemic stroke, it includes demographics, neuroimaging, and extensive cognitive outcome information. We propose to capitalize on this unique resource in developing a tool to predict cognitive outcomes from stroke across 7 independent stroke cohorts. First, using existing data we will generate lesion-symptom maps, which identify brain regions associated with cognitive impairment. We will also use an innovative strategy that links lesion-associated deficits to specific patterns of brain network disconnection, called lesion network mapping. This method infers what networks are disrupted by brain lesions in association with specific symptoms, which improves the accuracy of cognitive outcome predictions. A major goal of this proposal is to optimize cognitive outcome predictions by evaluating novel approaches to lesion- symptom mapping and lesion-network mapping in comparison to existing methods, with the goal of identifying the approach that explains the most variance in cognitive outcomes. A second major objective is to combine different modalities of predictive information within ensemble machine learning models – both lesion and non- lesion predictors. The third major objective is to engineer a fully automated pipeline for generating these predictions, such that clinically acquired MRI data can be used as input to an algorithm that generates a report of personalized cognitive outcome information based on lesion location and lesion-associated network disruption, which will contribute to statistical models that also include non-lesion predictive information. By addressing these objectives, we will lay a strong foundation for developing a clinical tool that uses clinically acquired brain MRI data to generate personalized cognitive outcome predictions based on ischemic stroke lesion location. This will improve stroke management by informing early intervention decisions and helping to guide rehabilitation and life planning for patients with stroke.