A Comparison of Feature Extraction and Dimensionality Reduction Techniques for EEG-Based BCI System

Abstract

Brain Computer Interface (BCI) plays a pivotal role in transforming the lives of physically disabled people. BCI also provides a new mode of communication to healthy people. It uses signals derived from brain to establish a connection between a user’s state of mind and a computer. The Electroencephalography (EEG) based BCI measures the scalp-projected electrical activity of the brain with millisecond resolution up to over 200 electrode locations. It results in a high dimensional dataset which is hard to visualize, analyze and model. For analyzing these signals, a subset of features often leads to better classification than the full set of features. The present investigation compares various feature extraction and dimensionality reduction techniques, viz., Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Factor Analysis (FA), Multidimensional Scaling (MDS) and Isometric feature mapping (ISOMAP). The techniques have been tested on Motor Imagery (MI) EEG data obtained during left hand and foot motor imagination. Dimensionality reduction ability and discrimination power of all the techniques have been accessed for comparison. It is found that LDA outperforms all other tested methods. It results in effective dimensionality reduction and high discrimination power.

Keywords: BCI, Feature extraction, Dimensionality reduction techniques