An attention-based deep convolutional recurrent model for single trial subject-independent P300 detection

Abstract

Deep learning has recently become very popular in various fields, and CNN and RNN are two very important elements in deep learning. They perform very well in the task of detecting ERPs (Event-Related Potentials) on EEG. Since EEG data contains many defects, such as huge variability among different people and low signal-to-noise ratio, it is very difficult to develop a system generalizable between subjects is still difficult. In this paper, we examined a deep learning Model (ADCRM) for EEG-based ERPs detection in subject-independent P300 detection scenario. In our model, the convolutional network module is constructed for encode the frequency domain and spatial characteristics of EEG signals and the recurrent network module is built for exploring the dynamic time domain characteristics, while a channel-wise attention module is integrated to focus on the most important part of the EEG channels. We trained the model with a relatively small number of datasets and examined its performance by comparing with the previously proposed methods. The experiment result shows that our model achieved the best performance among all the models we tested. This result indicates that our model can find the deep-seated EEG features and generalize to other subjects.