Lightweight Graph Neural Network for Dementia Assessment from EEG Recordings

With the rapid aging observed in the global population, senile dementia is becoming a major concern. Early detection, classification, and monitoring of dementia stages are therefore considerable challenges that have attracted the interest of the research community. In this context, the use of machine learning applied to electroencephalography (EEG) recordings has emerged as a promising modality, thanks to the reduced cost and ease of use. This paper proposes an analysis of the currently largest EEG signal dataset, the Chung-Ang University Hospital EEG (CAUEEG) dataset, which reveals a bias in the data related to patients age information, and hence proposes a novel approach employing a lightweight Graph Neural Network (GNN) model to classify dementia stages from EEG recordings. The proposed GNN model inherently captures the complex interdependencies within EEG signals by modeling the spatial relationships between recording electrodes as a graph structure. Experimental results on the CAUEEG dataset demonstrate the effectiveness of the proposed lightweight GNN model in discriminating between normal, mild cognitive impairment, and dementia EEG recordings, achieving competitive performance with respect to existing state-of-the-art approaches, while being from 500 to 1000 times lighter than previously existing methods.