Unsupervised feature extraction with autoencoders for EEG based multiclass motor imagery BCI

Abstract

Decoding of motor imagery (MI) from Electroencephalogram (EEG) is an important component of BCI system that helps motor-disabled people interact with the outside world via external devices. One of the main issues associated with the multiclass classification of MI based EEG is the informative confusion due to non-stationary characteristics of EEG data. In this work, an innovative idea of transforming EEG signal into a new domain, weight vector of autoencoder, unsupervised neural network, is proposed for the first time to solve that confusion. These weight vectors are optimized according to that particular EEG signal. The features: autoregressive coefficients (ARs), Shannon entropy (SE) and wavelet leader were extracted from the weight vector. A rectangular windowing-based feature extraction technique is implemented to capture the local features of the EEG data. Finally, extracted features were used in the support vector machine (SVM) as a classifier network. The proposed method is implemented on two openly available EEG dataset (BCI competition-III and Competition-IV) to validate the effectiveness and superiority of the proposed methodology over the newly reported methods. For four-class EEG based MI classification, the proposed technique has achieved an average test accuracy of 95.33% and 97% for dataset-IIIa from BCI-III and dataset-IIa from BCI-IV respectively. The experimental results reveal that, the proposed technique is a promising solution to improve the decoding performance of BCIs.