Detection of alcoholism by combining EEG local activations with brain connectivity features and Graph Neural Network

Of late, Machine Learning (ML) and Deep Learning (DL) based techniques have become popular for automated screening of long-term alcoholism using Electroencephalogram (EEG) signals. However, most of the ML and DL-based methods for alcoholism detection rely upon the features extracted from individual EEG electrodes' signals. In fact, the existing methods do not fully exploit the inherent topological structure of brain activity. On the other hand, Brain Connectivity Analysis (BCA) being an advanced approach provides an efficient way to express the brain topology and more significantly has the capability of synchronizing the co-activation between different brain regions in the form of a brain network. In the present study, synergistic integration of individual EEG electrodes' features relevant to alcoholism and knowledge of inherent connectivity patterns between spatially distributed electrodes were performed. This work combined both the information in the form of a graph, where the individual electrodes' features were embedded as node features and the edges represent the connectivity information. After that, the generated alcoholic and non-alcoholic graphs were classified using Graph Neural Network (GNN). A publicly available alcoholism dataset was used to validate the proposed framework. Based on the Phase Lag Index (PLI) connectivity estimator and Graph Convolution Neural Network (GCNN) classifier, the 10-fold cross-validation substantiated the highest classification accuracy of 93.28%. Further, the effects of alcoholism in different EEG sub-bands were also investigated, where the Beta band exhibited the highest classification accuracy of 81.76% among the sub-bands. Lastly, the different aspects and design considerations of the proposed framework were analyzed thoroughly by conducting multiple experiments.