A General Multi Time Scale Spatiotemporal Compound Model for EEG Classification

Recently Brain computer interfaces (BCI), a direct communication technique between machine and brain, plays an important role in the fields of brain disease diagnose, rehabilitation and robotics with the help of electroencephalography (EEG). EEG-based brain signal feature extraction and task categorization have become a popular trend. The procedure of brain signal analysis includes three steps: pre-process, feature extraction and categorization. For a given BCI paradigm, these steps are tailored to explore distinct characteristics of its expected control signals. To generalize the process, we propose a multi time scale spatiotemporal compound classification model (MTSC) based on Convolution Neural Network (CNN). The model firstly utilizes 2D convolution along time axis capturing temporal feature, then depth-wise convolution along channel axis is done for capturing spatial feature. Both of two domain features composite a facet of original EEG signal. We set up different 2D convolution kernel size according to the signal sample rate in order to generate different views, which contains various time scale information. These views are weighted summed for classification. Experiments on four different paradigm datasets have been conducted comparing with well performed deep learning and traditional methods. The results show that our model achieves better marks on all datasets in accuracy and F1-score.