P300 event-related potential detection using one-dimensional convolutional capsule networks

The main challenge in creating a brain-computer interface (BCI) is establishing an effective brain signal recognition model suitable for achieving direct communication between humans and computers. Recently, various deep learning-based methods have been proposed to improve the performance of P300 event-related potentials (ERPs) for BCI. However, during the detection of P300 ERP signals, even electroencephalogram (EEG) signals from the same person are inconsistent and may be significantly distorted, resulting in impaired classification accuracy in many deep learning methods. Here, we propose a machine learning model based on a one-dimensional convolutional capsule network (1D-CapsNet). This network topology can effectively detect P300 ERP signals in the time domain, thereby achieving a better detection performance than can the current convolutional neural network (CNN)-based methods. Two classifiers based on the 1D-CapsNet model are proposed, namely, 1D-CapsNet-64 and 1D-CapsNet-8, which are used for classifying EEG data with 64 and 8 electrodes, respectively. These two classifiers are tested and compared on dataset II of the third BCI competition. The results show that the 1D-CapsNet-64 classifier obtains the best character recognition rate result (96%). The proposed method is superior to both state-of-the-art CNN-based methods and various traditional machine learning methods. The experimental results reveal the feasibility of our proposed method for detecting P300 ERP signals. The proposed method is expected to expand the concept of EEG signal recognition pattern and improve BCI design and applications.