A channel-mixing convolutional neural network for motor imagery EEG decoding and feature visualization

Convolutional Neural Network (CNN) has achieved great success in decoding EEG signals, decoders based on these architectures make separate feature extraction and classification into an integrated stage, however, a large number of trainable parameters introduced by the model hinder the improvement of EEG decoding performance and challenge the interpretability of decoding process used CNNs. In this paper, we propose an end-to-end shallow and lightweight CNN framework, which allows EEG-Motor Raw dataset as inputs, to boost decoding accuracy by the Channel-Mixing-ConvNet. The first block of network is designed in the way of implicitly stacking temporal-spatial convolution layers for learning temporal and spatial EEG features after EEG channels were mixed, compared to previously independently building a single temporal and spatial convolutional layer, this method combines the feature extraction capabilities of the two layers. The Mixed Channel Process block introducing a depthwise convolution layer is applied for a series of processing such as to decouple and supplement the internal and external mapping relationships existing in the mixed multi-dimensional EEG feature maps. Finally, the classification block is constructed to finish EEG decoding tasks. The lightweight architecture of Channel-Mixing-ConvNet leaves space for the model to exploit its potential performance by stacking other layers. In our experiments, the proposed Channel-Mixing-ConvNet and variants based on different hyper-parameters were evaluated on public EEG-motor datasets BCI-IV 2a and HGD respectively, Channel-Mixing-ConvNet outperformed state-of-the art (SOA) algorithms for EEG decoding. Additionally, via post-hoc interpretation techniques, the results show the learned features are consistent with the neurophysiological principle of the EEG motor imagery, meanwhile, the model also captures the remarkable features associated with channels.

Automated summary

The paper introduces an end-to-end shallow and lightweight CNN framework called Channel-Mixing-ConvNet for EEG signal decoding, which effectively combines feature extraction and classification while reducing the number of trainable parameters. Experimental results demonstrate that Channel-Mixing-ConvNet outperforms other algorithms in EEG decoding performance, with learned features consistent with neurophysiological principles of EEG motor imagery.