Motor imagery EEG decoding based on TS-former for spinal cord injury patients

Traditional machine learning methods struggle with efficiency when processing large-scale data, while deep learning approaches, such as convolutional neural networks (CNN) and long short-term memory networks (LSTM), exhibit certain limitations when handling long-duration sequences. The choice of convolutional kernel size needs to be determined after several experiments, and LSTM has difficulty capturing effective information from long-time sequences. In this paper, we propose a transfer learning (TL) method based on Transformer, which constructs a new network architecture for feature extraction and classification of electroencephalogram (EEG) signals in the time-space domain, named TS-former. The frequency and spatial domain information of EEG signals is extracted using the Filter Bank Common Spatial Pattern (FBCSP), and the resulting features are subsequently processed by the Transformer to capture temporal patterns. The input features are processed by the Transformer using a multi-head attention mechanism, and the final classification outputs are generated through a fully connected layer. A classification model is pre-trained using fine-tuning techniques. When performing a new classification task, only some layers of the model are modified to adapt it to the new data and achieve good classification results. The experiments are conducted on a motor imagery (MI) EEG dataset from 16 spinal cord injury (SCI) patients. After training the model using a ten-time ten-fold cross-validation method, the average classification accuracy reached 95.09 %. Our experimental results confirm a new approach to build a brain-computer interface (BCI) system for rehabilitation training of SCI patients.