Journal of Health Management & Informatics (Jan 2020)
A New method for Improvement of the Accuracy of Character Recognition in P300 Speller System: Optimization of Channel Selection by Using Recursive Channel Elimination Algorithm Based on Deep Learning
Abstract
Background and purpose:P300 speller is a kind of Brain-Computer Interface (BCI) system in which the user may type words by using the responses obtained from human focus on different characters. The high sensitivity of brain signals against noise in parallel with the similarity of responses obtained from the user focus on different characters makes it difficult to classify the characters based on their respective P300 wave. On the other hand, all areas of the brain does not carry useful P300 information. Proposed Methods: In this study, a new method is proposed to improve the performance of speller system which is based on selecting optimal P300 channels. In the proposed method, recursive elimination algorithm is presented for channel optimization, which utilizes deep learning concept (e.g. Convolutional Neural Network) as its cost function. The proposed method is examined on a data set from EEG signals recorded in a P300 speller system, including 64 different channels of responses to 29 characters. Then, its performance is compared with some existing methods. Results: The obtained results showed the ability of the proposed method in recognizing the characters in such way that it could accurately (i.e. 97.34%) detect 29 characters by using only 24 out of all 64 electrodes. Conclusion: Applying the proposed method in speller systems led to considerable improvement in classification of characters compared to its alternatives. Several experiments proved that utilizing the proposed scheme may increases the accuracy almost 12.9 percent compared to non-optimized case in parallel with reduction of the number of involved channels by approximately 1/3. Based on these results, the proposed method may be considered as an effective choice for application in P300 speller systems, thanks to reduction of the complexity of the system which is caused by the reduced number of channels and, on the other hand, due to its potential in increasing the accuracy of character recognition.
