MI-MBFT: Superior Motor Imagery Decoding of Raw EEG Data Based on a Multibranch and Fusion Transformer Framework

The potential of brain-computer interface (BCI) technology in stroke rehabilitation has garnered significant attention worldwide, particularly as motor imagery electroencephalogram (MI-EEG) signal decoding techniques have established a connection between biosignals and rehabilitation machines. However, EEG’s nonstationary nature and the susceptibility of its spatial-temporal features to brain conditions and environmental factors pose significant challenges. Most deep learning (DL) decoding networks tend to concentrate on local features, resulting in limitations when adapting to globalwise interdependencies. This article presents an end-to-end multibranch and fusion transformer (MBFT) framework to focus on and fuse full-sequence length spatial-temporal spectral information through self-attention, thereby adaptively capturing key elements in EEG signals. MI-EEG encompasses rhythmic characteristics of multiple spectra throughout the entire sequence and with intrasequence interdependencies. To learn the global patterns of local features from multiple frequency bands in parallel, MBFT employs a multibranch transformer encoder (TE) structure and further utilizes multihead self-attention (MSA) to acquire fused high-level features. Our proposed model demonstrates superior performance on various datasets. It achieves an average accuracy of 86.93% on the BCIC IV-2a dataset, 94.64% on the BCIC II and III datasets, and 93.52% on the MMIDB dataset, all achieving a new state-of-the-art (SOTA). This approach explores a novel network structure that may be more suitable for MI-EEG decoding, which helps improve the BCI system’s performance.