Multivariate lesion symptom mapping for predicting trajectories of recovery from aphasia

Abstract

Aphasia is an acquired disorder of communication that results from injury to regions of the brain that support language. While much has been learned about the neural bases of language and aphasia in the past two hundred years, it remains challenging to predict the nature and extent of recovery from aphasia after stroke. Previous studies have disproportionately investigated chronic lesion-deficit relationships, with very few incorporating time as a crucial explanatory variable. Here, support vector regression (SVR) was used to predict language outcomes based on acute clinical imaging data at multiple time points (acute, 1 month, 3 months, and 12 months) in a large cohort of individuals with left hemisphere stroke (N=359). Effects of lesion characteristics, along with other clinical, linguistic, and demographic variables, were examined for their utility in predicting aphasia outcomes across time. Intraclass correlation (ICC) between predicted and actual language scores demonstrated that outcomes in many subdomains of language could be predicted with good to excellent accuracy at multiple time points, and that information about lesion location was often crucial for making such accurate predictions, particularly at later time points post-stroke. The creation of machine learning models to predict longitudinal language outcomes from acute clinical imaging data could, in concert with clinical expertise, lead to improved allocation of resources in healthcare settings and more appropriately tailored treatments for individual patients. Such models could also provide a valuable baseline for evaluating whether different patterns of functional reorganization are associated with better or worse outcomes than might be expected based on structural damage alone.