Machine learning techniques for electroencephalogram based brain-computer interface: A systematic literature review

Abstract

Brain-computer interface systems with Electroencephalogram (EEG), especially those use motor-imagery (MI) signals, have demonstrated the ability to control electromechanical devices with promising results. EEG being easy to record and non-invasive makes it a good choice for BCI systems. MI-based BCI systems compute neuronal activity and decipher these electrical impulses into gestures or effects, aiming to enable the person to communicate with their surroundings. This study summarises techniques of EEG signal processing used in the recent decade. This research paper presents an exhaustive survey on four aspects of EEG signals in BCI systems: signal acquisition, signal pre-processing, feature extraction, and classification. The most prominent time-frequency technique, wavelet transform (WT), and its updated version, wavelet packet transform (WPT), is primarily used in EEG-BCI systems for feature extraction. The development of artificial intelligence technology motivated researchers to classify motor imagery signals for BCI systems using machine learning (ML) and deep learning (DL) techniques. This literature survey paper explores more than 220 research papers related to ML and DL approaches to classify EEG signals for BCI systems. In order to identify prospective research areas for future investigation, present challenges are carefully considered, and suggestions are also provided for appropriate feature extraction and classification techniques. The authors expect that the investigation presented in this paper will help researchers to find accurate feature extraction, ML, and DL methods and these techniques will be supportive in devising an effective EEG-BCI system.