A novel approach of decoding four-class motor imagery tasks via wavelet transform and 1DCNN-BiLSTM

Electroencephalogram (EEG)-based human-computer interaction (HCI) has become a major research direction in the field of brain-computer interface (BCI). Although EEG research has made progress, motor imagery (MI) EEG decoding remains a challenge due to a lack of sample data, a lower signal noise ratio (SNR), and individual differences. Recently, many deep learning methods have been widely used in EEG classification tasks. Our work presents a novel approach to decoding four-class MI tasks by utilizing one-dimensional convolutional neural network (1DCNN). We use 1D multiscale CNN (1DMCNN) block, residual attention mechanism (RAM) block and bidirectional long-short-term memory (BiLSTM) networks for EEG decoding. We named it the 1DMRCNN-BiLSTM model, which can achieve good accuracy in decoding human intentions. The highlights include: (1) Based on the wavelet transform (WT), we use 1D wavelet denoising and 1D wavelet reconstruction methods not only to improve the SNR of EEG signals but also to enhance the number of EEG samples. (2) We fused the 1DMCNN block and the RAM block with dropout layers to design a new 1DMRCNN model for EEG feature extraction. (3) Based on the 1DMRCNN-BiLSTM structure, an effective end-to-end framework for MI classification is built. We trained and tested our proposed method on the BCI competition IV datasets 2a (BCICID-2a) and PhysioNet. The experimental results show that the method demonstrates excellent ability in EEG decoding.