Graph Convolution Network Based Classification of Subjects with Prefrontal Cortex Lesion via Information-theoretic Brain Network Features

This paper investigates scalp electroencephalogram (EEG) data from 14 subjects with unilateral prefrontal cortex (pFC) lesions and 20 healthy controls during lateral visuospatial working memory (WM) tasks. The goal is to differentiate the brain networks involved in WM processing between these groups. The EEG recordings are transformed into graph signals, with proximity-weighted brain connectivity measures as edges and centrality measures as nodal features. Graph convolutional network (GCN) layers are used for feature representation, followed by a fully connected layer for classification. The GCN-based model effectively handles nine classification tasks, proving that graph-based network representation is versatile for describing brain interactions. The sparse MI-GCI-based graph model’s accuracy effectively captures the functional segregation of distinct WM tasks. The classifier using mutual information-guided Granger causality index (MI-GCI) with 20% of top edges matched prior classification performance with 67% fewer parameters and 80% less graph density, identifying the correct class of all 34 subjects in group identification using leave-one-out cross-validation and two-thirds majority voting.