Unsupervised mapping of causal relations between brain and behavior

Abstract

AbstractStudies of patients with focal brain lesions provided the historical foundation for cognitive neuroscience, but how to identify a precise mapping between specific brain regions and the cognitive variables affected remains unclear. The challenge lies both in identifying anatomical regions wherein lesions have a shared causal effect, as well as in the precise delineation of the behavioral outcome. Currently, either the relevant brain region or the dimensionality of the behavior being mapped are pre-specified by the investigators rather than both being informed by optimal brain-behavior relationships. Here we apply a novel data-driven causal aggregation algorithm, Causal Feature Learning (CFL) to tackle this challenge in 520 individuals with focal brain lesions. CFL simultaneously constructs macro-level summaries of the spatial distribution of brain lesions and the itemized responses on psychometric tests to optimally characterize the causal brain-behavior relationships. Focusing on the domains of language, visuospatial ability, and depression, we recapitulate established findings, provide new and more precise anatomical results, and present an aggregation of item-wise data that provides an empirical test of extant behavioral scores and can be used to identify novel, psychologically meaningful factors. Future work could use our approach to construct entirely new psychometric variables that might cut across established categories.