Preprocessing and feature extraction techniques for brain-computer interface

Abstract

In order to improve communication and interaction between humans and computers, the signal processing and artificial intelligence (AI) are vital tools. For a seamless brain–computer interface (BCI), it integrates and analyzes the data from multiple sensors. The objective is to develop more efficient, natural, and intuitive interfaces that can comprehend and react to human input more effectively. Usually, these modalities consist of electrocorticography (ECoG), functional magnetic resonance imaging (fMRI), and electroencephalography (EEG). In this context, the preprocessing and feature extraction methods play an important role. The aim of preprocessing is noise removal and focus on the most significant frequency content of the signal. The key preprocessing approaches are the digital filtering, wavelet transform, and multiscale principal component analysis (MSPCA). The feature extraction is vital in achieving accurate representation for modeling or identifying critical elements or intentions in the human body systems using machine or deep learning techniques. Feature extraction facilitates the identification and interpretation of relevant information from input data streams. This chapter explores various feature extraction techniques employed in BCI applications, ranging from parametric model-based methods to more complex approaches. Traditional techniques encompass the signal processing methods such as digital filtering and Fourier transform. The intended parametric model-based methods are the autoregressive, Yule-Walker, covariance, and modified covariance. Further considered approaches are the subspace-based methods, eigenvector, and time–frequency analysis, such as the short-time Fourier transform and different variants of wavelet transform. Additionally, the oscillatory mode decompositions and common spatial patterns are described. These methods are effective for extracting pertinent information from the input signals and, moreover, they enable the automated decision support through machine and deep learning methodologies for the contemporary BCIs.