Classification and Analysis of Minimally-Processed Data from a Large Magnetoencephalography Dataset using Convolutional Neural Networks

Convolutional neural networks were used to classify and analyse a large magnetoencephalography (MEG) dataset. Networks were trained to classify between active and baseline intervals of minimally-processed data recorded during cued button pressing. There were two primary objectives for this study: (1) develop networks that can effectively classify MEG data, and (2) identify the important data features that inform classification. Networks with a simple architecture were trained using sensor and source-localised data. Networks trained with sensor data were also trained using varying amounts of data. The important features within the data were identified via saliency and occlusion mapping. An ensemble of networks trained using sensor data performed best (average test accuracy 0.974 {+/-} 0.001). A dataset containing on the order of hundreds of participants was required for optimal performance of this network with these data. Visualisation maps highlighted features known to occur during neuromagnetic recordings of cued button pressing.