Improved EEGNet With a Multilevel Spatial Feature Extraction Module for EEG Decoding

In this article, a multilevel spatial feature extraction module (mSEM) is presented for electroencephalography-based brain-computer interface (EEG-BCI) tasks that are conducted according to the baseline EEGNet model. Spatial EEG features are extracted in two stages: local spatial feature extraction and global feature extraction. In the local feature extraction stage, all EEG electrodes are divided into several groups according to their brain regions, and specific convolution kernels are designed for each brain region to extract local features. In the global feature extraction stage, a global convolution kernel is used to extract the spatial pattern among all the electrodes and brain regions. The mSEM can assist EEGNet and its variants in learning the latent spatial features embedded within the input EEG data. Experiments conducted on both a self-collected dataset and a public dataset show that the mSEM with various brain region division strategies can improve different backbone models in BCI tasks. In subject-dependent tasks conducted on the BCI competition IV-2A public dataset, with the mSEM, EEGNet achieves an accuracy increase of 1.38%, and the accuracy of the variant model attention temporal convolutional network (ATCNet) increases by 3.6% to 89%, which is the current state-of-the-art (SOTA) result. In addition, a validation of the brain region division strategy, an analysis of the utilization efficiency of the spatial information contained in the input data and other analysis experiments demonstrate the effectiveness of the mSEM. This study reduces the effects of limitations such as high noise and low spatial resolution in EEG measurements and provides a new solution for enhancing the usability of EEG-BCIs in practical scenarios.