Classification of P300 based brain computer interface systems using long short-term memory (LSTM) neural networks with feature fusion

Abstract

Enabling to obtain brain activation signs, electroencephalography is currently used in many applications as a medical diagnostic method. Brain-computer interface (BCI) applications are developed to facilitate the lives of individuals who have not lost their brain functions yet have lost their motor and communication abilities. In this study, a BCI system is proposed to make classification using Bi-directional long short term memory (Bi-LSTM) neural networks. In the designed system, spectral entropy method including instantaneous frequency change of signal is used as feature fusion. In the study, electroencephalography (EEG) data of 10 participants are collected with Emotiv EPOC+ device using 2x2 visual stimulus matrix prepared on Unity. Each symbol of the 2x2 matrix includes stimulus such as doctor, police, fireman and family. These stimuli are demonstrated to participants with a fixed order. As data collection protocol, 200 ms stimulus time and 300 ms interstimulus interval are used. As the performance success of classification, the average accuracy rates are obtained to be 98.6% for training set and 96.9% for the test set. In addition, in classification of P300 EEG signals, the results obtained via Bi-LSTM are compared with the results obtained using 1 dimensional convolutional neural networks (1DCNN) and support vector machines (SVM) classification methods. Moreover, in the study, information transfer rate (ITR) is provided as 40.39 at an acceptable level.