Frequency Bands EEG Biomarkers for Dementia Using Graph Neural Networks

We introduce a simple model for the classification of electroencephalography (EEG) signals. Our primary focus is on employing deep learning to investigate how integrated connectivity patterns can be used to classify EEG signals and to identify the key discriminative features leveraged by the model. In this study, we utilize connectivity features across multiple frequency bands within a multi-edge Graph Neural Network (GNN) framework, demonstrating that edge features carry complementary information. We apply this model to predict Frontotemporal Dementia (FTD) using EEG data. The proposed GNN achieves an average accuracy of approximately \(76\%\) using a Leave-One-Subject-Out (LOSO) validation scheme outperforming baseline models and achieving performance comparable to state-of-the-art approaches. We further investigate the importance of connectivity edges, nodes, and frequency bands in the model’s predictions using explainable AI (XAI) methods based on saliency maps, enabling global interpretation followed by statistical analysis to understand the subjects deviations. The saliency maps reveal the critical roles of the occipital and anterior temporal regions in predicting FTD. Moreover, the alpha and theta frequency bands emerge as key contributors, consistent with previous findings in literature.