A new deep structure to improve detection of P300 signals: using supervised learning as kernel of convolutional neural networks

Brain-computer interface (BCI) systems provide a safe and reliable interface between brain and outer world and detecting P300 signal plays a vital role in these systems. In recent years, convolutional neural networks (CNNs) have made a vast and rapid development in P300 signal detection. In this paper, a novel structure for CNN is proposed to enhance separability of the selected features in its convolutional layer. In proposed scheme, an artificial neural network is applied in the above layer as nonlinear filter which extracts nonlinear features which lead to improve detecting of P300 signals. The performance of the proposed structure is assessed on EPFL BCI group dataset. Then, the achieved results are compared with the basic structure for P300 detection. The obtained results demonstrate the improvement of true positive rate (TPR) of the proposed structure against its alternative by extent of 19.69%. Such improvements for false detections and accuracy are 1.97% and 10.87% which show the effectiveness of applying the proposed structure in detecting P300 signals.