How many channels are suitable for independent component analysis in motor imagery brain-computer interface. (April 2019)

• Record Type: Journal Article
• Authors: Zhou, Bangyan; Wu, Xiaopei; Ruan, Jing; LV, Zhao; Zhang, Lei
• Is Part Of: Biomedical signal processing and control. Volume 50(2019)
• Page Start: 103
• Page End: 120
• Publication Date: 2019-04
• Subjects: Brain computer interface -- Independent component analysis -- Channel selection -- Motor imagery; 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.2019.01.017
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This paper proposed a channel selection method based on ICA algorithm, which can obtain high accuracy with few channels. We proposed a single-trial-based self-testing (STST) method that sequentially selected one single trial of EEG data to calculate ICA spatial filters, which were used for classification of the identical trial of EEG data. That is to say, the training data and testing data were the same single trial, so if this trial of EEG data was contaminated by artifacts, it can only affect the classification result of their own. Thus, this method can eliminate the influence of bad trials. We proposed a new method which transferred optimal channel subsets across subjects or sessions for the same subject in the ICA calculation, rather than the conventional spatial model parameters. The experiment results suggested that ICA algorithm with optimal channel subset had comparable performance with Common spatial patterns (CSP) algorithm in self-testing and run-to-run cross validation, and ICA significantly outperformed CSP in session-to-session and subject-to subject transfer. Abstract: Individual differences of classification performance remain a crucial problem in electroencephalography (EEG)-based motor imagery brain computer interface (MIBCI). Independent component analysis (ICA) is a promising spatial filtering technique in BCI system for it requires few and unlabeled training samples for calibration of the BCI system. However, both the distribution of scalp electrodes and the quality of training data are critical factors of influencing the classification performance of ICA-BCI applications. In this study, a new channel selection algorithm was proposed for automatically choosing subject-specific minimal electrode subsets that can obtain high classification accuracies of the ICA-BCI system. The algorithm consisted of two steps: selection of "main electrodes" located on the motor cortex, and subsequent searching of "subordinate electrodes", which were picked out one by one from the left electrodes until the maximum accuracy was achieved. Meanwhile, a single_trial_based_self_testing (STST) method, utilizing one single trial to train ICA spatial filters which were only applied in the identical trial for extracting motor-related independent components (MRICs), was proposed to eliminate the influence of bad trials. The channel selection algorithm was applied in 72 runs of three-class motor imagery EEG datasets for twelve BCI users. Experimental results indicated that the classification accuracies using the optimal channels were significantly higher than that of standard 8 and 9 channels. Meanwhile, ICA algorithm with optimal channel subset had comparable performance with Common spatial patterns (CSP) algorithm in self-testing and run-to-run cross validation, and ICA significantly outperformed CSP in session-to-session and subject-to-subject transfer. Although the numbers and locations of optimal channels were different between sessions and subjects, the main electrodes were basically same between different runs for long-term BCI users. Furthermore, the optimal electrodes were primarily located on the motor cortex of parietal lobe area and the frontal lobe area, few located in the occipital lobe area. Too many or too few channels were not suitable for ICA calculation, and usually, using 5–8 channels of EEG data could achieve better classification performance. These findings may offer a reference for the optimization of ICA-based BCI systems, and further improve the performance and stability of MI-BCI system.