Learning Optimization Using Genetic Algorithm in Post-Stroke EEG Signal Classification

Abstract

A stroke is an attack that often requires long-term rehabilitation. One result of this condition can be seen from abnormal electrical signals in the brain, recorded by an electroencephalogram (EEG). Therefore, EEG can be used for monitoring and evaluation of post-stroke rehabilitation. Neurologists usually observe EEG signals based on their density, amplitude, waveform, and comparison of the channel pairs, but this analysis is not easy. Besides, using machine learning, such as Backpropagation, is sometimes constrained by random initial weights. This state can lead to a long convergence. This paper proposes the selection of initial weights in Backpropagation training using Genetic Algorithms. The use of Genetic Algorithms can optimize the initial weight selection in Backpropagation. The EEG signal used has been extracted into Alpha, Theta, Delta, and Mu waves. The experimental results show that using the Genetic Algorithm can increase non-training data accuracy to 75%, compared to only 65% without the genetic algorithm. Genetic Algorithms can overcome overfitting and local maximums. The results also show that the use of Wavelet transform for feature extraction can increase the accuracy from 60% to 75%. The optimization of training parameters also determines the accuracy.