Cross-subject EEG emotion recognition combined with connectivity features and meta-transfer learning

• Published in: Computers in biology and medicine Vol. 145; p. 105519
• Main Authors: Li, Jinyu, Hua, Haoqiang, Xu, Zhihui, Shu, Lin, Xu, Xiangmin, Kuang, Feng, Wu, Shibin
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.06.2022; Elsevier Limited
• Subjects: Accuracy; Adaptation; Arousal; Artificial neural networks; Asymmetry; Brain research; Connectivity feature; Cross subject; Datasets; EEG; Electroencephalography; Emotion recognition; Emotional factors; Emotions; Humans; Internal Medicine; Machine Learning; Meta-transfer learning; Neural networks; Neural Networks, Computer; Other; Physiology; Transfer learning; emotion recognition; Meta-transfer learning; EEG; Connectivity feature; Cross subject
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2022.105519

In recent years, with the rapid development of machine learning, automatic emotion recognition based on electroencephalogram (EEG) signals has received increasing attention. However, owing to the great variance of EEG signals sampled from different subjects, EEG-based emotion recognition experiences the individual difference problem across subjects, which significantly hinders recognition performance. In this study, we presented a method for EEG-based emotion recognition using a combination of a multi-scale residual network (MSRN) and meta-transfer learning (MTL) strategy. The MSRN was used to represent connectivity features of EEG signals in a multi-scale manner, which utilized different receptive fields of convolution neural networks to capture the interactions of different brain regions. The MTL strategy fully used the merits of meta-learning and transfer learning to significantly reduce the gap in individual differences between various subjects. The proposed method can not only further explore the relationship between connectivity features and emotional states but also alleviate the problem of individual differences across subjects. The average cross-subject accuracies of the proposed method were 71.29% and 71.92% for the valence and arousal tasks on the DEAP dataset, respectively. It achieved an accuracy of 87.05% for the binary classification task on the SEED dataset. The results show that the framework has a positive effect on the cross-subject EEG emotion recognition task. •The work proposed a novel method for Cross-subject EEG emotion recognition. The highlights are as follows:•We combined the multi-scale residual network, meta-transfer learning and connectivity features for superior performance.•The proposed method has good performance in cross-subject EEG emotion recognition tasks based on DEAP and SEED datasets.•MSRN was adopted to capture interactions of different brain regions in the manner of multi-scale.•MTL made full use of the merits of meta learning and transfer learning to shallow the gap of individual difference.