Emotion recognition by distinguishing appropriate EEG segments based on random matrix theory. (September 2021)

• Record Type: Journal Article
• Authors: Sarma, Parthana; Barma, Shovan
• Is Part Of: Biomedical signal processing and control. Volume 70(2021)
• Publication Date: 2021-09
• Subjects: EEG segment -- Emotion recognition -- Entropy of EEG in emotion -- Random matrix theory; 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.2021.102991
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Emotion recognition by distinguishing EEG segments with high affective content. Selected segment extraction from EEG signals and its associated EEG sub-bands. Classification by k -NN, SVM & RF with PSD & CWT features from SEED & DEAP dataset. The α, β and γ bands from frontal, temporal and occipital regions are suitable. Abstract: This work proposes an emotion recognition technique by distinguishing appropriate electroencephalogram (EEG) segments from acquired signal for target emotions. Generally, EEG based emotion recognition system works considering entire acquired signal from different channels, which is improper as emotional states do not remain steady for entire duration of signal acquisition. Therefore, in this work, appropriate EEG signal segments ( L ) from diverse channels have been distinguished using random matrix theory (RMT). Next, classification tasks have been performed considering two features of the selected EEG segments — power spectral density (PSD) and continuous wavelet transform (CWT), in three classifiers — k -nearest neighbour ( k -NN), support vector machine (SVM) and random forest (RF). Experimental validations have been performed on two EEG datasets for emotion analysis — SEED (distinct emotions) and DEAP (dimensional emotions). The proposed method achieved accuracies upto 95% ( L = 60 s) and 86% ( L = 2 s) for SEED (positive vs negative) and DEAP (high vs low arousal) respectively, when considering segments from entire EEG signals. Meanwhile, the EEG sub-bands, α, β and γ, displayed the highest accuracy upto 98% ( L = 60–80 s) and 94% ( L = 2–15 s) for SEED and DEAP respectively. For both cases, k -NN classifier with CWT feature and channels from frontal, temporal and occipital brain regions have been found very suitable. Additionally, the selected EEG segments having low coefficient-of-variation of entropy achieves the highest classification accuracy. In a comparative study, this method reveals its superiority by achieving the highest classification performance by considering automatically selected short length EEG signal segments from distinct channels.