EEGNet-MSD: A Sparse Convolutional Neural Network for Efficient EEG-Based Intent Decoding

Electroencephalography (EEG) is a noninvasive technique that can be used in brain machine interface (BMI) systems to measure and record brain electrical activity. Deep learning (DL) techniques have proved superior to conventional methods in EEG-based intent decoding. However, some DL models have overly complex structures while ensuring the accuracy of EEG recognition, resulting in reduced training and recognition speed. In this study, we proposed a compact multihead self-attention DL decoder that combined the convolutional neural network (CNN)-based EEGNet decoder with the ProbSparse multihead self-attention mechanism. Compared with traditional self-attention methods, this decoder ensures alignment dependent on both time complexity and memory usage of O(L log L) and it has been demonstrated to enhance the accuracy of EEG-based intent recognition. The test results on dataset 2a from BCI Competition IV showed that the EEGNet multihead self-attention decoding (EEGNet-MSD) decoder performed approximately 8% better than the competition-winning decoder filter bank common spatial pattern (FBCSP) and namely batch and pairwise (NBPW), and achieved better results than the latest long short-term memory (LSTM) neural decoding method. In addition, a binary classification test was performed on the Physiobank EEG motor imagery (MI) dataset, and the results showed that the accuracy of EEGNet-MSD was approximately 4% higher than EEGNet, validating the stability of the EEGNet-MSD decoder. This study provides a new solution for enhancing the performance of EEG-based intent decoding in both accuracy and decoding speed.