EEG Signal Classification for Predicting the Movement Intent of Lower Limb

Abstract

Brain-computer interface (BCI) is a system, based on hardware and software communication mechanism that enables brain activity to control external devices or computer. The main objective of developing a BCI system is to provide communication capabilities to people who are totally paralyzed or severely disabled by neuromuscular disorders. A standard BCI system comprise of signal acquisition, preprocessing, feature extraction, classification and control interface. In this study, we investigated the performance of independent component analysis (ICA), average band power, mean value, peak value, variance, kurtosis and skewness as feature vector and Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Naïve Bayes, Support Vector Machine (SVM) and k-Nearest Neighbor (kNN) as classifier for the classification of flexion/extension movements of lower limb knee and ankle joints. An optimal feature-classifier combination to develop EEG based four-class BCI system for lower limb has also been determined. It is shown that QDA classifier trained with ICA as feature vector has achieved the maximum classification accuracy among all other features and classifiers. However, determination of optimal feature-classifier combination is achieved in such a novel manner that twenty sets, containing 3-combinations of six feature vectors are used to evaluate the performance of each classifier. Feature combination set containing average band power, mean value and peak value has outperformed all other possible combinations by achieving maximum average classification accuracy. However, in classifiers LDA has shown quite promising results among all other classifiers used. Apart from enhancing the classification accuracy, this study will eventually contribute towards developing better controllers for neuro-prosthetic devices particularly design for lower limbs. The study has been performed experimentally with Emotiv headsets equipped with fourteen electrodes to acquire EEG data from four human test subjects in synchronous mode. Classification and data analysis has been performed offline however in future the study will be extended to an online BCI system.