Post-Stroke Recognition Based on EEG Using PCA and Recurrent Neural Networks

Abstract

One instrument for stroke identification is Electroencephalogram (EEG). Previous studies often used the wave variables Delta, Theta, Mu, Alpha, and amplitude in stroke analysis. For this purpose, they are often using Wavelet and Fast Fourier Transform (FFT). Although the first is more appropriate for non-stationary signals such as EEG. Likewise, in this study. However, processing EEG signals also give complexity to the use of many channels. Therefore, in addition to wave extraction, it is necessary to reduce the information from multi-channel. This paper proposed using Principle Component Analysis (PCA) for extracted signals of multichannel, which are then identified against three classes using Recurrent Neural Networks (RNN). The experimental results showed that the use of PCA produced greater accuracy of 86% compared to without PCA, which only provides an accuracy of 60%. The choice of the number of components is also an essential configuration in PCA channel reduction. Experiments using six components of PCA, Delta-Theta-Alpha-Mu waves, and amplitude as features gave the best performance. The research showed that both Adam and SGD models carried the same accuracy. Nevertheless, Adam model faster and more stable compares to SGD Model.