Which BSS method separates better the EEG Signals? A comparison of five different algorithms. (February 2022)

• Record Type: Journal Article
• Authors: Stergiadis, Christos; Kostaridou, Vasiliki-Despoina; Klados, Manousos A.
• Is Part Of: Biomedical signal processing and control. Volume 72(2022)Part A
• Publication Date: 2022-02
• Subjects: EEG -- BSS -- ICA -- Signal processing -- EEG decomposition; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103292
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An extensive comparison of the five most common BSS algorithms. There is a trade-off between performance and execution time in BSS algorithms. The results suggest AMICA outperforms RUNICA, which is currently so widely used. The number of the EEG channels does not affect the herein proposed BSS ranking. Abstract: A very common strategy for rejecting electroencephalographic (EEG) artifacts, includes the decomposition of filtered EEG signals using a Blind Source Separation (BSS) algorithm, the identification and removal of artifactual components and the reconstruction of the cleaned EEG signals. In this pipeline, the performance of the BSS algorithm, which is defined as its ability to separate properly the independent sources (like the EEG from artifactual sources), is very crucial for rejecting most of the artifacts, while maintaining the most part of EEG intact. The overwhelming majority of the published papers uses the extended INFOMAX version of Independent Component Analysis (ICA) for artifact rejection purposes. But is this the most efficient algorithm to separate EEG signals into independent components? This study comes to shed light to the aforementioned question by assessing the performance of the five most common BSS algorithms. The normalized entropy of the brain-related components, their correlation between independent components with the original sources and the amount of the overall mutual information reduction (MIR) achieved by each decomposition were computed in datasets with systematically varying numbers of electrodes (ranging from 19 tο 99), from 26 real human scalp EEG recordings. Additionally, 54 different datasets containing artificially contaminated EEG signals were also examined for the same purpose, on the basis of the Euclidean distance and the correlation, between the generated Independent Components (ICs) and the original vertical and horizontal eye signals, which were used for the contamination. The results support that the Adaptive Mixture ICA was the best performing BSS method.