Key result
The proposed GCSFormer model achieved an average accuracy of 85.36%, precision of 85.23%, recall of 86.35%, and F1 score of 84.52% in cross-subject emotion recognition tasks.
Why the study?
Most EEG-based emotion recognition datasets are unimodal or bimodal, which may overlook emotional information reflected by other physiological signals.
The GCSFormer model demonstrates high accuracy in cross-subject emotion recognition using multimodal physiological signals.
Multimodal fNIRS-EEG signals may boost cross-subject emotion recognition; leaves open clinical translation and cardiovascular applications.
Cross-subject emotion recognition is an important research direction in the fields of affective computing and brain-computer interfaces, aiming to identify the emotional states of different individuals through physiological signals such as functional near-infrared spectroscopy (fNIRS) and electroencephalogram (EEG). Currently, most EEG-based emotion recognition datasets are unimodal or bimodal, which may overlook the emotional information reflected by other physiological signals of the subjects. In this paper, a multimodal dataset named Multimodal Emotion Four Category Dataset (MEFD) is constructed, which includes EEG, Heart Rate Variability (HRV), Electrooculogram (EOG), and Electrodermal Activity (EDA) data from 34 participants in four emotional states: sadness, happiness, fear, and calm. This will contribute to the development of multimodal emotion recognition research. To address the recognition difficulty caused by individual differences in cross-subject emotion recognition tasks, a classification model named Global Convolution Shifted Window Transformer (GCSFormer) composed of an EEG-Swin Convolution module and an improved Global Adaptive Transformer (GAT) module is proposed. By using a parallel network, the feature discrimination ability and generalization ability are enhanced. The model is applied to classify the EEG data in the self-built MEFD dataset, and the results are compared with those of mainstream methods. The experimental results show that the proposed EEG classification method achieves the best average accuracy of 85.36%, precision of 85.23%, recall of 86.35%, and F1 score of 84.52% in the cross-subject emotion recognition task. The excellent performance of GCSFormer in cross-subject emotion recognition task was verified.
No takes yet. Share an insight, caveat, or question.
Xiangyu Deng (2025) studied Emotion recognition (n=34). GCSFormer (Global Convolution Shifted Window Transformer) model vs. Mainstream methods was evaluated on Cross-subject emotion recognition average accuracy. The proposed GCSFormer model achieved an average accuracy of 85.36%, precision of 85.23%, recall of 86.35%, and F1 score of 84.52% in cross-subject emotion recognition tasks.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: