MRMHNet: a new convolutional neural network approach for decoding electroencephalogram motor imagery signals

The decoding of electroencephalogram (EEG) signals plays an extremely important role in brain-computer interfaces (BCI). However, the processing of physiological signals, particularly the decoding of multi-channel EEG signals, still poses significant challenges. Past deep learning methods often relied on subject-dependent settings, which resulted in new users needing to perform complex calibration procedures before they could use BCI devices. Therefore, we proposed a novel end-to-end deep learning model, MRMHNet, for motor imagery (MI) classification. Firstly, we utilized a feature extraction block based on a Multi-Resolution convolutional neural network (MRCNN) to extract features in both frequency and spatial domains. Secondly, we utilized a block based on the Multi-Head Attention (MHA) to extract global temporal information of the features. Finally, we validated the classification performance of our method using OpenBMI datasets, and the results showed that our method achieved the highest accuracy in both subject-dependent and subject-independent settings. Specifically, in the subject-independent setting, our method achieved the highest accuracy and F1-score, with values of 73.74±13.35% and 73.33±14.87%, respectively. This indicates that our method has good classification performance and high practical value in the field of BCI.