Year: 2026
Type: article
Abstract: While most cross-subject EEG emotion recognition studies focus on average accuracy, the high inter-subject variability of EEG signals causes severe performance inconsistency across individuals. This inconsistency poses a critical barrier to real-world deployment but has received insufficient attention. To address this issue, this study proposes a hybrid emotion recognition model integrating moment channel attention and Transformer, termed the Moment Attention Convolution–Emotion Transformer Network (MAC-ETN). The proposed hybrid network consists of three components: a local temporal feature extractor, a Moment Channel Attention Network (MCA), and a global temporal feature extractor. Specifically, depthwise and separable convolutions are employed to extract local temporal features. The moment channel attention mechanism incorporates the first-order moment (mean), second-order moment (variance), and third-order moment (skewness) for channel importance calibration; and global dependency modeling is achieved through a Transformer encoder. Performance evaluation is conducted on two public datasets, DEAP and MAHNOB-HCI, using leave-one-subject-out cross-validation (LOSO-CV). The experimental results show that the main advantage of MAC-ETN lies in markedly reduced cross-subject performance variability rather than higher mean accuracy. On the DEAP dataset, the mean accuracy of MAC-ETN (61.26% for arousal and 60.47% for valence is lower than that of the strongest baselines reported under the same LOSO protocol, while its accuracy standard deviation (7.88% and 5.31%) is the lowest among all compared methods. On the MAHNOB-HCI dataset, the F1 standard deviation of MAC-ETN is 1.48% for arousal and 0.30% for valence, far below the values reported for the baselines in the literature, although the latter were obtained under a subject-dependent protocol and with a different F1 definition, so this comparison is indicative only. These results indicate that MAC-ETN reduces cross-subject performance variability at a moderate level of accuracy, which is relevant for practical EEG-based emotion recognition applications.
Source: Frontiers in Psychiatry
Authors: Shuran Yang, Mushan Han
Institutions: Northeastern University, Northeastern University
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,801
OpenAlex ID: W7225830218