Evaluation on the Effect of EEG Pre-processing and Hyper parameters Tuning to the Performance of Convolutional Neural Network Motor Execution Classification

Abstract

Electroencephalogram (EEG) based classification has achieved a promising performance using deep learning models like Convolutional Neural Network. Various pre-processing strategies such as smoothing the EEG data or filtering are commonly used to pre-process the captured EEG signal before the subsequent feature extraction and classification while hyperparameters tuning might help to improve the classification performance. As well, the number of layers used in the CNN can affect the performance of the classification. In this paper, the number of layers needed for the CNN to classify the EEG data correctly, the effect of apply smoothing to pre-process the EEG signal for modern end-to-end CNN and the effect of enabling hyperparameters tuning during the training phase of CNN is investigated and analyzed. Two CNN models, namely Deep CNN with 5 layers and Shallow CNN with 1 layer, with convincing classification accuracy on motor execution classification as reported in the literature were chosen for this study. Both the CNN models are trained on EEG motor execution dataset with different training strategies and dataset pre-processing. Based on the obtained training and test classification accuracy, Shallow CNN trained with enabling hyper parameters tuning and without smoothing the EEG data achieved the best classification accuracy with average training accuracy of 99.9% and test accuracy of 96.87%. This indicates that CNN does not need to have many layers to correctly classify the motor execution data and the EEG data does not require smoothing.