Recently, the international journal IEEE Journal of Biomedical and Health Informatics (impact factor: 6.7) published a research paper by the team of Professor Chen Shengbo from Nanchang University's School of Artificial Intelligence, entitled "Sub-band Embedding Based EEG Spatio-Temporal Activity Representation for Emotion Recognition." Nanchang University is the first-listed affiliation. Shi Jianye is the first author, while Professor Chen Shengbo and Dr. Duan Wenying are the corresponding authors.

Emotion recognition is an important research topic in brain-computer interfaces and intelligent healthcare. Electroencephalography (EEG) signals are characterized by multiple frequency bands, multiple channels, and strong temporal dependencies. However, existing methods remain limited in their use of frequency information, modeling of higher-order spatial relationships, and extraction of long-range temporal dependencies. To address these challenges, the research team proposed a unified modeling approach integrating sub-band embedding, hypergraph modeling, and state-space temporal learning, enabling the coordinated representation of frequency, spatial, and temporal information in EEG signals.

First, the team divided EEG signals into five sub-bands and enhanced the representation of emotion-related information in different frequency bands through learnable embeddings. A hypergraph structure based on neurophysiological priors was then introduced to model higher-order relationships among multiple EEG channels at the brain-region level, better reflecting the organization of functional brain networks. For temporal modeling, a multi-head Mamba state-space model was employed to efficiently capture nonlinear features and long-range dependencies in EEG signals. Finally, graph convolution was used to fuse spatio-temporal features across frequency bands, producing a unified discriminative representation for emotion classification.

Comprehensive experiments demonstrated that the proposed method achieved excellent performance on the two public datasets, SEED and SEED-IV. On the SEED dataset, the model achieved accuracies of 85.63% in the cross-subject setting and 95.86% in the subject-dependent setting. On the SEED-IV dataset, it achieved accuracies of 84.37% and 74.77% in the subject-dependent and cross-subject settings, respectively. The model also remained stable under complex emotion-category configurations, demonstrating advantages in cross-subject generalization and complex emotion recognition.

This research improves modeling performance in EEG-based emotion recognition. Its proposed paradigm of multi-frequency collaborative modeling, higher-order spatial relationship learning, and efficient temporal modeling is also broadly applicable and can be extended to related tasks in brain-computer interfaces and complex biosignal analysis, providing useful insights for the intelligent analysis of complex temporal signals.

The study was supported by the National Natural Science Foundation of China (62566039 and 62403309), the Jiangxi Provincial Natural Science Foundation (20242BAB20066), and other research projects.

Paper link: https://doi.org/10.1109/jbhi.2026.3673733