An 'EEG ruler' that transcends time: cross-time stable decoding in affective brain-computer interfaces
Science China Press
image:
Multi-level Dynamic Integrated Perception Network for cross-time EEG emotion recognition.
view moreCredit: ©Science China Press
A brain-computer interface that identifies emotion accurately today may not perform as reliably tomorrow. For affective brain-computer interfaces, that is the central obstacle to sustained use.
Affective brain-computer interfaces (aBCIs) use neurophysiological signals such as EEG to sense and recognize human emotional states. Potential applications include monitoring mental states and enabling human-machine interaction. However, EEG is highly non-stationary: even in the same person, neural states shift and recording conditions vary over time, so signals collected on different days or at different stages of an experiment rarely look the same. A model that performs well on today's data may therefore falter when applied to a session recorded weeks or months later.
The team led by Shuang Liu from the Medical School of Tianjin University/Haihe Laboratory of Brain-Computer Interaction and Human-Machine Integration has now developed a multi-level dynamic integrated perception network, or MDIN, to address temporal variability in cross-time EEG emotion recognition (Figure 1). The study, published in National Science Review, models and structurally calibrates temporal variability at three levels: emotion-related neural representations, dynamic functional connectivity, and cross-time discriminative structure.
Finding stable emotional signals across multiple EEG dimensions
Emotion-related EEG activity is not determined by a single moment, brain region, or frequency band. It emerges from interactions across time, space, and frequency. MDIN therefore begins with a temporal-spatial-spectral joint perception module. A dual-branch temporal-attention mechanism identifies time segments carrying important emotional information, while spatial attention captures local and global dependencies among brain regions. Weighting these jointly highlights task-relevant time points and channels, yielding more discriminative representations for dynamic brain-network modeling.
Explicitly modeling sample-specific changes in brain-network connectivity
Conventional graph-convolution models commonly use a fixed adjacency matrix to describe connections among brain regions. This static structure cannot fully capture changes in functional connectivity across emotional states and recording times. MDIN addresses this limitation through an adaptive dynamic perception module, or ADP, which builds brain-network connections tailored to each sample and then prunes away weaker connections to keep only the most discriminative edges.
MDIN doesn't throw away the stable structural cues a static brain network provides. Instead, a gated fusion mechanism combines static and dynamic connections at the edge level. This allows MDIN to preserve a relatively stable network backbone while capturing local and selective functional reconfiguration associated with emotion.
Selectively calibrating temporal variability
Once time-related changes have shifted the EEG distribution, the model must still preserve the relationships that separate one emotion from another. MDIN introduces a task-driven discriminative alignment framework, or TDAF, for this purpose. A diffusion-based process separates relatively stable emotional patterns from disturbances associated with temporal variability. Because the target session is unlabeled, TDAF also screens and corrects pseudo-labels using confidence, predictive uncertainty, and class-center structure. Cross-time alignment therefore considers not only the overall feature distribution but also the decision boundaries among emotion categories.
The connection-level analyses showed that TDAF did not indiscriminately strengthen the entire network. Connections shared across emotions remained comparatively stable, whereas emotion-specific connections underwent clearer dynamic reorganization. Time-varying connections with strong discriminative value were selectively enhanced, while unstable or weakly discriminative connections were suppressed or corrected.
The evolution of the sample distribution told a consistent story. Introducing dual-threshold filtering, multi-indicator consistency constraints, and sample rebalancing steadily improved pseudo-label reliability and corrected the early category-mapping errors caused by temporal variability. Confusion-matrix analyses also showed fewer errors between commonly confused emotion categories, along with clearer separation in the discriminative space.
The researchers evaluated MDIN on the laboratory-developed ECPL dataset and the public SEED dataset, each containing recordings from three days. MDIN achieved cross-time emotion-recognition accuracies of 96.97% on ECPL and 94.21% on SEED, the highest results among the methods evaluated in the study.
Toward affective brain-computer interfaces that remain reliable over time
The findings suggest that cross-time EEG emotion recognition requires more than removing statistical differences between recording periods. A model must explicitly characterize how emotion-related neural activity changes over time and determine which changes should be preserved, strengthened, or corrected. MDIN provides such a framework and may lay the methodological groundwork for long-term applications in mental-state monitoring, human-machine interaction, and personalized emotional intervention.
Feifan Yan is the first author of the paper. Shuang Liu and Dong Ming of Tianjin University are the co-corresponding authors. The work was supported by the Science Fund for Distinguished Young Scholars of Tianjin, the National Natural Science Foundation of China, and the Tianjin Science and Technology Major Special Project and Engineering.
Journal
National Science Review
No comments:
Post a Comment