Enhancement of egg signals classification by linear discriminant analysis for brain computer interface

Abstract

Motor imagery (MI) based electroencephalogram (EEG) signals classification is under research for the last few decades to develop a robust and user-friendly brain-computer interface (BCI) system without compromising its simplicity and efficiency. The number of channel selections is still the most challenging task to extract features and classify them for MI movement detection. Hence, an advanced but required simple computation with minimal channels selection, Linear Discriminant Analysis (LDA) based algorithm has been developed. BCI competition IV dataset-I has been utilized in this research that was collected by the renowned BCI group from the Berlin Institute of Technology. Initially, the signal is preprocessed in a few steps by applying a sliding window and utilizing a finite impulse response (FIR) filter to obtain a cutoff frequency ranging from 8-30 Hz. The power spectral density (PSD) technique has been adopted to extract the power spectrum of µ and β features over frequency components. A common spatial pattern (CSP) filter is also applied to optimize feature extraction and feature selection from the signal. Then, classification has been done in two stages, training, and evaluation phase. Comparatively lower classification error has been recorded by the LDA classifier for left and right-hand MI classification. The classification accuracy is measured at 91.14% and 81.4% in the training and evaluation phase respectively. Cohen's kappa coefficient is calculated at 0.822 in the training phase and 0.629 in the evaluation phase which proves the research's viability. Therefore, to aid persons such as with spinal cord injuries, the suggested approach can be applied to real BCI devices.