The Impact of Enriching Electroencephalogram in Spatial Metadata on Interpretability and Generalization Ability of Graph Neural Networks

This paper proposes a formalization of current approaches to processing electroencephalograms, as well as a generalization of the original problem of processing multidimensional time series in terms of graph neural networks. The proposed architecture exploits information about the recording device by encoding information about the relative position of the recording electrodes as a graph. The effectiveness of the method has been demonstrated in tasks of emotion recognition, as well as P300 impulse recognition. The model achieved outstanding results on the considered datasets using the same set of hyperparameters. The experiments conducted demonstrate the usefulness of graphs composed of metainformation, as well as the dependence of the structure of the optimal graph on the problem being solved. Potentially, this technique can be extended to other subject areas where additional information about the connectivity of multivariate time series channels is available.