FBATCNet: A Temporal Convolutional Network With Frequency Band Attention for Decoding Motor Imagery EEG

• Published in: IEEE access Vol. 13; pp. 11265 - 11279
• Main Authors: Ma, Shuaishuai, Lv, Jidong, Li, Wenjie, Liu, Yan, Zou, Ling, Dai, Yakang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Brain modeling; Brain-computer interface; Convolution; Convolutional neural networks; Datasets; Decoding; EEG; Electrodes; Electroencephalography; Feature extraction; fine motor imagery; Frequencies; Hands; Human-computer interface; Imagery; Motors; Rehabilitation; Stroke (medical condition); temporal convolutional network; upper limb rehabilitation
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2025.3525528

Motor imagery-based brain-computer interfaces (MI-BCIs) hold significant promise for upper limb rehabilitation in stroke patients. However, traditional MI paradigm primarily involves various limbs and fails to effectively address unilateral upper limb rehabilitation needs. In addition, compared to decoding MI-EEG signals from different limbs, decoding MI-EEG signals from same limb faces more challenges. We introduced a novel tri-class fine motor imagery (FMI) paradigm and collected electroencephalogram (EEG) data from 20 healthy subjects for decoding research. Furthermore, we proposed a frequency band attention-based temporal convolutional network (FBATCNet) for MI-EEG decoding. First, an innovative use of the channel attention mechanism adaptively assigned weights to segmented EEG frequency bands, improving the frequency resolution of MI-EEG signals. Subsequently, convolutional block further integrated frequency-domain features and extracted spatial features. Finally, a temporal convolution block was utilized to capture advanced temporal features. The proposed model achieved accuracy of 84.73% on BCI Competition IV-2a (Dataset 1) and 66.06% on the private FMI dataset (Dataset 2). In the classification of subject dependent, the FBATCNet is better than the baseline methods mentioned in this paper. These results confirm that the FBATCNet is feasible and offer fresh insights for designing and applying FMI-BCI.