A Compact Convolutional Neural Network for Decoding EEG Functional Connectivity: Application to Motor Imagery

The design of neural decoders exploiting brain functional connectivity is growing interest in the neuroscience community, mainly for brain-computer interface applications. By exploiting also/exclusively the interactions among different brain regions as input features, neural decoding improved. However, the adoption of connectivity estimates for neural decoding is still in its infancy. Indeed, it is mainly adopted with non-directed connectivity measures, and it is mainly exploited in classic machine learning pipelines, limiting the analysis only on few interactions or frequency ranges extracted from the connectivity estimate, without fully exploiting the totality of the information contained in the measured functional connectivity (i.e., spatial and frequency domains characterizing connectivity matrices). To overcome this limitation, we designed a convolutional neural network for handling directed connectivity measures estimated via spectral Granger causality. The network automatically learned features in the frequency and spatial domains, separately resuming connectivity inflow and outflow. Our approach was applied to motor imagery decoding, and achieved state-of-the-art performance compared to existing decoders. Moreover, the features learned by the network matched the directional interaction known occurring when imagining movements, confirming that the network feature learning was able to capture the most relevant brain network characteristics.