Cross-Subject Emotion Recognition From Multichannel EEG Signals Using Multivariate Decomposition and Ensemble Learning

Abstract

Emotions are mental states that determine the behavior of a person in society. Automated identification of a person's emotion is vital in different applications such as brain–computer interfaces (BCIs), recommender systems (RSs), and cognitive neuroscience. This article proposes an automated approach based on multivariate fast iterative filtering (MvFIF) and an ensemble machine learning model to recognize cross-subject emotions from electroencephalogram (EEG) signals. The multichannel EEG signals are initially decomposed into multichannel intrinsic mode functions (MIMFs) using the MvFIF. The features, such as differential entropy (DE), dispersion entropy (DispEn), permutation entropy (PE), spectral entropy (SE), and distribution entropy (DistEn), are extracted from MIMFs. The binary atom search optimization (BASO) technique is employed to reduce the dimension of the feature space. The light gradient boosting machine (LGBM), extreme learning machine (ELM), and ensemble bagged tree (EBT) classifiers are used to recognize different human emotions using the features of EEG signals. The results demonstrate that the LGBM classifier has achieved the highest average accuracy of 99.50% and 98.79%, respectively, using multichannel EEG signals from the GAMEEMO and DREAMER databases for cross-subject emotion recognition (ER). Compared to other multivariate signal decomposition algorithms, the MvFIF-based method has demonstrated higher accuracy in recognizing emotions using multichannel EEG signals. The proposed (MvFIF+DE+BASO+LGBM) technique outperforms the existing state-of-the-art methods in ER using EEG signals.