New Multiview Graph Learning Boosts Brain Disease Detection from EEG

Detecting cognitive diseases such as Alzheimer's and dementia using electroencephalogram (EEG) signals presents significant challenges due to the inherent noise, non-stationary nature, and inter-subject variability of these brain measurements. Traditional methods often struggle to extract robust discriminative information from long EEG recordings. A new framework, M-LINKX, addresses these issues by employing a multi-view graph learning approach. For each segment of an EEG recording, the system extracts channel-level features and constructs several functional connectivity graphs, each defined by different metrics, frequency bands, and topological filters. These diverse graph representations are then fused using trainable weights, and segment-level predictions are averaged to achieve a subject-level diagnosis. Experimental results on two distinct EEG datasets, CAUEEG and AHEAP, demonstrate that M-LINKX outperforms existing methods in subject-level classification of various dementia-related conditions. This suggests that integrating multiple perspectives of functional connectivity within an appropriate graph-learning architecture can substantially enhance the accuracy of EEG-based dementia classification.