Partial least squares multimodal analysis of brain network correlates of language deficits in aphasia.

Lesion-symptom mapping techniques are essential to determine brain regions critical for language functions. However, high collinearity in neuroimaging and behavioural data remains a challenge for distinguishing neural substrates supporting multiple language domains (shared variance) and those subserving specific language functions (unique variance). Here, we employed a novel approach to multimodal lesion-symptom mapping using multivariate partial least squares regression to delineate the latent structure of lesion-behavioural mapping in aphasia and decompose the shared and unique neural determinants of language impairments. A total of 86 participants with chronic (>12-month post-stroke) aphasia after left hemisphere strokes were examined. Language impairment was assessed with the Western Aphasia Battery-Revised, and brain damage was defined by multimodal neuroimaging (including lesion characteristics, structural and functional connectivity, volumetric measures and functional activity). Neuroimaging modality-specific models were constructed to evaluate the shared versus unique lesion anatomy associated with performance across Western Aphasia Battery-Revised subtests: auditory comprehension, naming, repetition and spontaneous speech. Model accuracy was validated using leave-one-out cross-validation. Latent decomposition revealed that 50% of the covariance between neuroimaging data and language performance was explained by two to six latent variables across models. The spontaneous speech subtest emerged as the most influential language measure across all models, with damage to regions surrounding the perisylvian fissure accounting for the largest amount of shared variance across Western Aphasia Battery-Revised subtests. Critically, the highest-ranking features represented in the latent variable models yielded moderately accurate simultaneous prediction for all language measures (highest