Classification of P300 event-related potentials using wavelet transform, MLP, and soft margin SVM

Abstract

Brain-computer interface is a communication mechanism between EEG signals and a computer, such that the system can capture the brain intention without involving motoric and muscular neurons. This study utilized the EEG recordings of four disabled subjects during repeated stimuli using a six-choice P300 paradigm. The EEG signals were processed with a Butterworth bandpass filter and Wavelet Transform, divided into two categories of the target and non-target trials. The EEG data were improved by removing the high amplitude fluctuation of the signals around the end of each file. The Wavelet Transform was implemented using Stationary Wavelet Transform (SWT) and Discrete Wavelet Transform (DWT). The target and non-target trials were averaged for every five trials, and the averaged non-target trials were reduced further by selecting one of every five consecutive data. The reduced target and non-target trial data were classified using multilayer perceptron and support vector machine. Using SWT, multilayer perceptron gave the maximum accuracy, sensitivity, and specificity of 96.4%, 96.6%, 96.2% respectively, and support vector machine obtained the maximum accuracy of 98.2%, sensitivity of 100%, and specificity of 96.4%. While using DWT, the best performance of multilayer perceptron gave the accuracy, sensitivity, and specificity of 94.5%, 100%, 89.3% respectively, and support vector machine had the maximum accuracy of 98.2%, sensitivity of 96.4%, and specificity of 100%.