Time-varying graph analysis comparing speech perception in healthy and aphasic brains

Aphasia affects millions of patients each year, causing communication difficulties and impeding their quality of life. While the nature of the language impairment after stroke is very patient specific, characterizing each person's impairment is useful for treatment and research. Recently, language processing frameworks as a whole are moving towards whole-brain network based analyses. Reflecting this, network based approaches to normal and impaired EEG data are also needed. We measured some aspects of language processing impairments with electroencephalography (EEG). Using EEG recordings from subjects engaged in a speech perception task, we devised a framework based on graph signal processing and information theory, to reveal features of the brain dynamics that separate the 21 healthy subjects from an aphasic patient. This patient has focal difficulty in speech perception. We performed a robust principal component analysis to jointly infer a time-invariant low rank structure and time-varying sparse matrices to characterize brain dynamics in the healthy and the aphasic subjects. With this graph theoretic approach, we demonstrate the possibility to characterize impairment at a patient specific level, that sheds insight into cognitive processes that were disrupted at the network level. This allows for a complementary measure of patient impairment to traditional ERP metrics.