Machine Learning Methods for Predicting Post-Stroke Aphasia and Language Recovery

Boston University (Charles River Campus), Boston MA

Investigators

Abstract

Project Summary/Abstract Aphasia is one of the most devastating consequences of stroke, yet clinicians are unable to provide a personal- ized prognosis to people with aphasia (PWA) due to the complex and multidimensional combination of factors that have the potential to influence recovery. Neuroimaging studies of aphasia largely focus on a single data modality and a restricted set of features, which contribute only a piece of the puzzle. Thus, there is a critical need for a mul- timodal exploration of aphasia recovery and for the development of new tools that can reliably forecast outcomes in PWA. Our long-term goal is to develop a robust and modular platform that can identify the most effective treatment(s) for PWA and provide personalized predictions on the course of recovery with treatment. As a crucial first step, the overall objective of this proposal is to develop new computational models that fuse multimodal neuroimaging and demographic data to parse the complex and heterogeneous etiology of aphasia. Specifically, we will leverage the complementary strengths of structural MRI, resting-state functional MRI (rs-fMRI) and diffu- sion MRI (d-MRI), which collectively capture the network organization of the brain. Our innovative strategy is to build off recent advances in artificial intelligence (AI) to automatically learn predictive patterns in the data. We will maintain interpretability of these models via biologically-informed neural network architectures tailored to each data modality and an automated feature selection strategy to isolate key biomarkers. This project will leverage multiple longitudinal datasets of chronic post-stroke aphasia (total N=301) that include behavioral scores and multimodal neuroimaging data. We will achieve the project objectives via three specific aims. In Aim 1, we will develop a predictive model that combines graph-based interactions in rs-fMRI with structural integrity in gray and white matter to predict baseline language functionality in PWA with chronic post-stroke aphasia. In Aim 2, we will use deep learning (DL) to identify neuroimaging biomarkers associated with treatment-induced changes in lan- guage functionality. Our analysis will be based on a curated dataset of PWA, who received structured language therapy for naming, spelling or sentence production. In Aim 3, we verify the extent to which the treatment-based patterns learned in Aim 2 are present in a more diverse cohort of PWA, who received variable (and unknown) en- vironmental stimulation. Specifically, we will use the similarity between learned representations to derive a “brain receptivity" index that quantifies the degree to which a PWA is expected to improve their language functionality over time. We anticipate the proposed research will have a transformative impact on aphasia research by al- lowing us to better understand brain features that predict recovery and laying the foundation to develop robust tools that forecast whether a patient will improve as a function of a given intervention.