Functional Graph Image Representation applied to EEG-based Mental Workload Classification

A lot of research has been lately dedicated to develop machine learning and statistical signal processing methods exploiting graph representations to solve inference and estimation problems. This is particularly relevant in the case of functional connectivity analysis from Electroencephalography (EEG) signals. However, the widely adopted functional connectivity metrics are hand-crafted, and thus suffer from the redundant and sometimes irrelevant information due to the volume conduction problem. Besides, the actual locations of the nodes (i.e. the electrodes) in the functional graph are often overlooked. In this work, we introduce an innovative approach leveraging the image representation of functional graphs learned from EEG signals under sparsity and structural constraints, where both the locations of the electrodes and a sparse functional connectivity learned from data are explicitly encoded. The resulting images are then fed to a convolutional neural network to extract meaningful latent features prior to inference. The proposed method is applied to Mental Workload (MW) classification. Experimental results on a public dataset demonstrate promising performance compared to state-of-the-art spatial filtering techniques and those based on hand-crafted functional connectivities.