采用源空间套索分析和卷积神经网络方法的 高频脑电动作模式识别方法

Addressing the challenge of low classification accuracy in ipsilateral hand movements, this paper presents a novel method named source-Lasso-CNN (SLC). The approach involves analyzing electroencephalogram (LEG) signals before movement onset in the gamma band (30—100 Hz) that are associated with four specific hand movements (tip pinch, multiple tip pinch, hand close and hand open) using spatial source localization. Region of interest (ROD was selected using group Lasso, and then the selected signals were input into convolutional neural network (CNN) for multi-class hand movement pattern recognition. The specific steps are as follows. Firstly, EEG and EMG signals were simultaneously collected from 13 subjects during the execution of the four hand movements, followed by preprocessing. Next, a head model was established using a boundary element model based on magnetic resonance image, and the inverse problem of EEG imaging was solved by using the minimum norm estimation method. The EEG sequences in the source space were divided into 79 regions based on Brodmann area. Three time-domain features were extracted from each brain region and a group Lasso algorithm was employed to select the ROl. Finally, the selected ROl and its corresponding source space sequences were input into the CNN for classification. The results show that the LASSO-CNN method, utilizing high frequency (y band) source space signals, achieves a superior classification accuracy of (82. 23 + 12. 71)%, which is better than that in S (1—3 Hz), 8 (4—7 Hz), a (8—13 Hz), [3(14—30 Hz) and full frequency band (1—100 Hz). Furthermore, the results also show a significant improvement in accuracy compared to other advanced algorithms, highlighting its effectiveness in recognizing identical hand motion pattern.