| name | prism-cross-subject-eeg-emotion |
| description | PRISM framework for cross-subject EEG emotion recognition using prioritized channel importance and semi-supervised domain adaptation. Differentiable channel weighting via lightweight expert ensemble plus confidence-filtered pseudo-labels for label-efficient generalization across subjects. Activation: EEG emotion recognition, cross-subject BCI, channel selection, domain adaptation, PRISM, semi-supervised EEG |
| metadata | {"arxiv_id":"2607.00358","published":"2026-07-01","authors":["Xin Zhou","Xiang Zhang","Hao Deng","Lijun Yin"],"tags":["eeg","emotion-recognition","cross-subject","domain-adaptation","channel-selection","bci","semi-supervised"]} |
| license | Complete terms in LICENSE.txt |
PRISM: Prioritized Channel Importance with Semi-supervised Domain Adaptation for Cross-Subject EEG Emotion Recognition
Core Contribution
A novel framework combining differentiable channel importance weighting with semi-supervised domain adaptation to solve two key obstacles in EEG emotion recognition: channel redundancy and inter-subject variability. Achieves state-of-the-art on DEAP, DREAMER, and SEED datasets with limited annotations.
Architecture: Two-Pillar Design
Pillar 1: Prioritized Channel Importance (Channel Side)
Problem: Standard EEG uses fixed channel sets with equal weighting. Many channels are redundant or subject-specific noise sources.
Solution: PRISM assigns differentiable, data-dependent channel weights via a lightweight expert ensemble:
- Multiple lightweight "expert" modules each specialize on different channel subsets
- Channel weights learned end-to-end via gradient descent
- Amplifies reliable electrodes while suppressing distractors
- Differentiable: weights adapt during training, not a fixed selection
Key advantage over prior channel selection methods: Most methods use discrete selection (keep/drop channels). PRISM uses continuous, differentiable weighting, allowing fine-grained importance scoring.
Pillar 2: Semi-supervised Domain Adaptation (Domain Side)
Problem: Inter-subject variability means models trained on one subject perform poorly on others. Labeling EEG data is expensive.
Solution: Leverages unlabeled data through confidence-filtered pseudo-labels:
- Generate pseudo-labels on unlabeled target-subject data using current model
- Confidence filtering: Only retain pseudo-labels with prediction confidence above threshold
- Use retained pseudo-labels for consistency regularization
- Domain alignment via adversarial or MMD-based loss between source and target distributions
- Iteratively refine as model improves
Methodology Workflow
- Preprocessing: Standard EEG preprocessing (bandpass filter, artifact removal, epoching)
- Feature extraction: Time-frequency decomposition or raw signal input
- Channel weighting: Forward pass through expert ensemble to get per-channel importance scores
- Supervised loss: Cross-entropy on labeled source-subject data
- Pseudo-label generation: Forward pass on unlabeled target data
- Confidence filtering: Retain predictions with confidence > threshold tau
- Consistency regularization: Enforce consistent predictions under input perturbations
- Domain alignment: Minimize distribution discrepancy between source and target
- Joint optimization: Total loss = L_sup + lambda1 * L_consistency + lambda2 * L_domain
Datasets and Results
| Dataset | Subjects | Channels | Emotions | Improvement over SOTA |
|---|
| DEAP | 32 | 32 | Valence/Arousal | Significant cross-subject gains |
| DREAMER | 23 | 14 | Valence/Arousal/Dominance | Robust with limited labels |
| SEED | 15 | 62 | 3 classes | State-of-the-art generalization |
Key result: PRISM achieves robust cross-subject generalization given limited annotations (as few as 10-20% labeled target data).
Pitfalls
Confidence threshold sensitivity
The pseudo-label confidence threshold tau is critical: too low introduces noisy labels, too high wastes unlabeled data. Recommend warmup schedule: start high (0.9) and gradually decrease.
Expert ensemble overhead
Multiple expert modules increase parameter count. Use lightweight experts (1-2 layer MLPs) to keep overhead minimal. Number of experts should match expected number of distinct channel importance patterns.
Channel importance not transferable across emotions
Optimal channel subsets differ between valence vs arousal decoding. Train separate channel weights per emotion dimension or use multi-task learning.
Pseudo-label confirmation bias
Early in training, model may generate systematically biased pseudo-labels that reinforce errors. Use exponential moving average (EMA) teacher model for more stable pseudo-labels.
Related Skills
- [[eeg-channel-adaptation-benchmark]] - Channel adaptation methods for EEG
- [[cross-subject-eeg-decoding]] - Comprehensive survey of cross-subject EEG decoding
- [[eeg-test-time-adaptation-benchmark]] - Test-time adaptation for EEG models
- [[eeg-biomarker-robustness-cross-population]] - Cross-population robustness
