EEG Valence Recognition Reaches 83.91%, but Drops for New Subjects

By Olivia Hart

DSE-TAM reached 83.91% accuracy for valence recognition on the DREAMER electroencephalography dataset under trial-wise cross-validation, a method that partitions trials within the same participant cohort. Under leave-one-subject-out (LOSO) validation, which excludes each tested participant from training, the same task fell to 65.11%.

How the model reads EEG

Emotion recognition supports analysis of cognition and behavior and may eventually inform brain-computer interfaces and mental-health monitoring. Electroencephalography (EEG), a noninvasive scalp recording of brain electrical activity, offers high temporal resolution but produces complex relationships across electrode channels and rapidly changing temporal patterns.

Existing deep-learning pipelines often divide EEG into separate sequences for individual channels. That simplification can disrupt spatial relationships and global signal structure, while a single temporal method may miss long-range dependencies or continuous transitions. DSE-TAM learns spatial relationships and temporal dynamics jointly instead of treating channels independently or relying on static representations.

The Dynamic Spatial Embedding (DSE) module uses graph convolutional networks to model relationships among EEG electrode channels. Its learnable adjacency matrix, a set of trainable channel connections, adaptively estimates inter-channel dependencies and introduces spatial information during early feature extraction.

DSE then converts spatially informed EEG segments into an ordered sequence of compact tokens. The Temporal Feature Extractor combines self-attention, which relates signals across separated time points, with a Mamba-based state-space model, which tracks dependencies and smoother changes through sequential neural activity.

The classification module uses a [CLS] summary token to aggregate the learned representations and predict an emotion category. This structure combines global temporal context with local temporal dynamics while keeping spatial and temporal learning within one framework.

Results across three EEG datasets

The researchers evaluated DSE-TAM on DEAP, DREAMER and CEED, participant-based human EEG datasets rather than animal recordings. They used trial-wise cross-validation and LOSO, making the latter the stricter test of transfer to previously unseen participants.

• DEAP: Trial-wise accuracy reached 79.75% for arousal, an activation measure, and 78.27% for valence, an emotional-positivity measure. Under LOSO, those figures fell to 61.29% and 58.80%.
• DREAMER: Trial-wise accuracy reached 80.55% for arousal and 83.91% for valence, exceeding the benchmark methods on both tasks. LOSO results were 61.96% for arousal and 65.11% for valence.
• CEED: Trial-wise three-class emotion accuracy reached 73.75%, a competitive result. LOSO accuracy fell to 59.59%.

The lower LOSO results show that strong within-participant trial scores do not guarantee transfer to new people. Those figures quantify the generalization problem across all three datasets and represent the central operational warning in these results.

Ablation analysis, which removes one model component at a time, identified self-attention as the largest contributor. Removing it produced accuracy drops reaching 18.67 percentage points on DEAP, 20.28 points on DREAMER and 29.32 points on CEED; removing the spatial embedding or Mamba modules also reduced performance.

Confusion-matrix analysis, a tabulation of correct and incorrect class assignments, showed improved balance across emotion categories. The gains were most notable for low-arousal states in DEAP and for distinguishing sadness from neutrality in CEED.

What does this change for EEG practice?

For EEG laboratories and prospective monitoring units, the immediate implication is methodological rather than operational. Model evaluation should prioritize unseen-participant performance and joint spatial-temporal learning, rather than report only within-participant trial splits.

The report does not describe an FDA-cleared device, a prospective patient study, a closed-loop brain-computer interface or a neurofeedback trial. DSE-TAM therefore remains an investigational classification framework tested on benchmark EEG datasets, not a deployable monitoring system.

The evidence comes from named research datasets rather than a defined clinical patient population, so bedside use would require separate validation. Its potential brain-computer-interface and mental-health applications remain prospective rather than demonstrated.

Limits of the current model

Four reported constraints define the current boundaries of the framework.

• EEG channels are not explicitly organized according to their spatial regions.
• The Mamba module uses unidirectional state transitions, potentially limiting bidirectional temporal dependency modeling.
• Classification of positive emotions remains challenging on the CEED dataset.
• The model assigns one emotion label to every sample within a trial, potentially overlooking emotional changes during that trial.

For monitoring units, the last assumption prevents the present framework from representing continuous emotional shifts within a recorded trial. Structured spatial modeling, bidirectional temporal architectures and methods that account for within-trial emotion variation could address the reported weaknesses, but the study did not evaluate those additions.

Publication and support

The journal Cyborg and Bionic Systems published Beyond Static Representations: Learning Electroencephalogram Spatiotemporal Dynamics for Enhanced Emotion Recognition on October 28, 2026. The paper is available at https://doi.org/10.34133/cbsystems.0701.

Ke Liu, Wenrui Zhu, Rui Xu, Wenlong Wang, Zhuliang Yu and Wei Wu authored the study. Funding came in part from:

• The STI 2030-Major Project, grant 2022ZD0211700, awarded to Xinhong Zhu.
• The National Natural Science Foundation of China, grants 62476034 and U24A20338.
• The Natural Science Foundation of Chongqing, grant CSTB2025NSCQ-JM0010.
• The Science and Technology Research Program of Chongqing Municipal Education Commission, grant KJZD-K202500607.
• The AI Program of Shanghai Municipal Education Commission, grant JWAIZD-4.

The next validation phase should test structured spatial mapping, bidirectional temporal transitions and within-trial emotion labels in independent participants before any monitoring or brain-computer-interface deployment.