Unsupervised EEG channel selection based on nonnegative matrix factorization. (July 2022)

• Record Type: Journal Article
• Authors: Xu, Lingfeng; Chavez-Echeagaray, Maria Elena; Berisha, Visar
• Is Part Of: Biomedical signal processing and control. Volume 76(2022)
• Publication Date: 2022-07
• Subjects: EEG signal -- Emotion recognition -- Nonnegative matrix factorization -- Channel selection -- Feature extraction; 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.2022.103700
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Semi-nonnegative matrix factorization is used for channel selection. Critical brain regions are identified to reduce the number of EEG channels. The method requires no feature engineering or labeling information. The method outperforms unsupervised methods and is on par with supervised methods. Abstract: High-density Electroencephalogram (EEG) systems have proven to be useful in enhancing the performance of emotion recognition algorithms. However, the high-dimensional nature of this data modality may also result in irrelevant information being captured, causing overfitting problems and increasing the computational cost of downstream algorithms. To perform efficient and accurate emotion recognition, an unsupervised channel selection framework based on semi-nonnegative matrix factorization (semi-NMF) is proposed. The algorithm excels in analyzing signals with complex internal correlations and produces results that are easy to interpret. Semi-NMF was used to decompose the high-density EEG signal matrices into several activation patterns. The strongest activation pattern was considered as most related to emotion recognition and channels with large weights in that activation pattern were selected for valence-based emotion recognition. It was found that the proposed framework can effectively detect brain regions that were active during emotional activities, and, using only this reduced set of channels, achieve better recognition performance than using all channels. Compared to existing methods, the framework selects channels in a physiologically explainable way and requires no supervised feature engineering or class labels. It results in higher accuracy compared to other unsupervised energy-based methods, and on par with the supervised ReliefF method. In all, the proposed framework not only serves as a valid channel selection tool for practical emotion recognition, but also has the possibility to be transferred to other non-classification tasks, potentially contributing to a variety of EEG applications, such as brain state monitoring, pathological brain activation analysis and brain disease diagnosis.