advanced technology

Investigating Brain Connectivity with Graph Neural Networks and GNNExplainer

To distinguish people with psychiatric disorders from healthy controls, a research group led by Nico Hoffmann, Helmholtz AI young investigator group leader (HZDR), developed and successfully deployed a graph neural network approach that uses electroencephalogram data.

Maksim Zhdanov, Helmholtz-Zentrum Dresden-Rossendorf (HZDR) and Technische Universität Dresden (TU Dresden), Saskia Steinmann, co-director of the Psychiatry Neuroimaging Branch (PNB) at the University Medical Center Hamburg-Eppendorf (UKE) and Nico Hoffmann, Helmholtz AI young investigator group leader at Helmholtz-Zentrum Dresden-Rossendorf (HZDR), set out to develop a way to distinguish between schizophrenia patients with and without auditory verbal hallucinations. Using a graph neural network with state-of-the-art performance approach they succeeded by examining electroencephalogram (EEG) data from patients from the aforementioned groups and healthy controls. 

Speech processing tasks can be used to shed light on the functional and structural aspects of the brain. The precise understanding of the kinetics of brain activity can be used to derive functional biomarkers for multiple pathologies that cannot be detected anatomically through clinical imaging. One such pathology is schizophrenia which is often followed by auditory verbal hallucinations. In this work, we have made a step toward an in-depth examination of functional connectivity during a dichotic listening task for three groups of people: schizophrenia patients with and without auditory verbal hallucinations and healthy controls. First, we represent EEG data as signals on the graph domain defined by connectivity among EEG electrodes. Next, we propose a graph neural network-based EEG classification framework to 1) predict a brain mental disorder based on EEG recording; 2) differentiate the listening state from the resting state for each group; 3) recognize characteristic task-dependent connectivity. The third task is achieved via recovering interpretable explanations for the predictions of the graph neural network focusing on the most informative subgraphs of an input graph. Experimental results show that the proposed model can differentiate between the aforementioned groups with state-of-the-art performance as well as provide a researcher with meaningful information regarding the functional connectivity for each group. Our approach respects the symmetries of EEG data, namely the translational and permutation symmetries, which allows it to beat baseline methods widely used in the field.

Figure: Framework overview. 

The paper was successfully submitted to the 26th International Conference on Pattern Recognition (ICPR 2022), which will be held in Montréal, Québec, from 21 to 25 August 2022.