🧬 biology

Microstate-Guided Dynamic Graph Neural Network with Prototypical Contrastive Learning for EEG Emotion Recognition

This paper proposes a microstate-guided dynamic graph neural network enhanced with prototypical contrastive learning to model evolving brain connectivity for EEG emotion recognition, achieving state-of-the-art subject-independent accuracy of 89.64% on the SEED benchmark.

Original authors: Zheng Liu, Lei Su, Liping Wu, Di Jiang

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zheng Liu, Lei Su, Liping Wu, Di Jiang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The human brain is a vast, shifting landscape of electrical activity. When we feel a sudden surge of joy, a wave of sadness, or a moment of calm, this activity does not just sit still; it reorganizes itself in milliseconds. Scientists have long tried to read these electrical signals, known as electroencephalograms or EEG, to understand what a person is feeling. However, the brain is not a static machine. Its connections change rapidly as emotions unfold, and every person's brain is wired slightly differently. This makes it incredibly difficult to build a computer program that can accurately recognize emotions across different people. Traditional methods often rely on looking at the brain's activity as a fixed map or a simple snapshot, missing the rapid, fluid way the brain actually works.

A team of researchers at the Kunming University of Science and Technology has proposed a new way to solve this problem. Instead of treating the brain's electrical signals as a static picture, they designed a system that follows the brain's natural, rapid shifts in real time. They focused on a specific phenomenon called "microstates." Imagine the brain's electrical field as a weather pattern on a map; for a brief moment, the pattern holds a specific shape, then it suddenly snaps into a new shape. These snapshots, which last only a fraction of a second, are the microstates. The researchers found that these fleeting patterns are not random noise but are actually the building blocks of how the brain processes emotions. By teaching their computer model to pay attention to these rapid shifts, they created a system that can understand the brain's emotional language much better than previous methods.

The core of their new approach is a framework they call MG-PCL. To understand how it works, imagine trying to understand a conversation between two people who speak different dialects. A standard computer might try to translate every word literally, but it would miss the tone and the context. The researchers' system does something smarter. It first breaks down the brain signals into two types of information. One type captures the immediate frequency of the electrical waves, which is like listening to the volume and pitch of the voice. The other type captures the microstate patterns, which is like understanding the rhythm and the pauses in the conversation. The system then combines these two streams of information, using the rhythm of the microstates to guide how it interprets the voice. This allows the model to see the bigger picture of how the brain is organizing itself, rather than just getting lost in the details of the electrical noise.

Once the system has combined these signals, it builds a dynamic map of the brain's connections. In older methods, this map was often a fixed template, like a static road map that never changed. But the brain's connections are fluid; they open and close as emotions change. The new system creates a map that evolves moment by moment. It uses the microstate patterns as a guide to decide which parts of the brain are talking to each other at any given second. This means the model can track how the brain reorganizes itself as a person moves from feeling neutral to feeling happy or sad. It is not just looking at which parts are active, but how the entire network is shifting its structure to support a specific feeling.

To make this work across different people, the researchers had to solve a major hurdle: the fact that no two brains are identical. A model trained on one person often fails when tested on another. To fix this, they introduced a method that aligns the emotional patterns of different people without forcing them to look exactly the same. They used a technique that groups similar emotional experiences together, creating a shared "prototype" for each feeling. Think of it as finding the common ground between different people's experiences of happiness, rather than trying to make their brains identical. By aligning these shared prototypes, the system learns to recognize the essence of an emotion, regardless of whose brain is producing it. This allows the model to generalize well, meaning it can accurately guess the emotions of people it has never seen before.

The researchers tested their system using a well-known collection of brain data called the SEED dataset, which contains recordings from many different people watching emotional videos. They asked the system to identify emotions without having been trained on the specific person it was testing. The results were striking. The new system achieved an average accuracy of 89.64 percent, which is higher than any other method previously tested on this same data. More importantly, it was more consistent, showing less variation in its performance across different people. When the researchers removed the microstate guidance from their system, the accuracy dropped significantly, proving that these rapid brain patterns are essential for the model's success. The system also showed that it was paying attention to the right parts of the brain, specifically the frontal and temporal regions, which are known to be involved in emotional processing.

The study suggests that the key to understanding brain signals lies in respecting their dynamic nature. By treating the brain as a system that is constantly reorganizing itself in rapid, rhythmic bursts, the researchers were able to build a model that sees the brain more clearly. They did not just find a better way to classify emotions; they found a way to listen to the brain's own language of change. While the system is not yet perfect and still struggles slightly with distinguishing between neutral and negative feelings, the results point to a promising future. The researchers plan to test their approach on even larger and more diverse datasets to see if it can hold up in real-world scenarios. For now, the work demonstrates that when we stop trying to freeze the brain in time and start following its natural flow, we can unlock a deeper understanding of human emotion.

Problem Statement

Electroencephalogram (EEG)-based emotion recognition faces significant challenges due to the high dimensionality, non-stationarity, and pronounced inter-subject variability of neural signals. While deep learning has improved spatiotemporal feature extraction, existing models often rely on static connectivity templates for Graph Neural Networks (GNNs). These static topologies fail to capture the functional reorganization of brain networks as emotions unfold. Furthermore, conventional spectral features (e.g., Differential Entropy) often lack the global dynamic organization information necessary to distinguish complex affective states, and purely data-driven domain adaptation methods may entangle emotion-discriminative patterns with subject-specific noise.

Methodology: MG-PCL Framework

The authors propose MG-PCL, a microstate-guided fusion framework designed to embed EEG microstate information at both the feature and graph levels, combined with prototypical contrastive domain adaptation. The architecture consists of three primary modules:

1. Microstate-Guided Dual-Prior Learning

Feature Extraction: The model extracts standard Differential Entropy (DE) features from EEG signals. Concurrently, it computes EEG microstate parameters (coverage, duration, occurrence, and transition probabilities) derived from four archetypal classes (A-D) using a modified K-means clustering on Global Field Power (GFP) peaks.
Fusion Mechanism: A dedicated Microstate Encoder processes the 28-dimensional microstate vector into a latent embedding. These are adaptively fused with DE features via a residual gating mechanism. This allows the model to preserve the discriminative power of spectral features while incorporating global brain dynamic information without allowing the weaker modality to dominate.

2. Microstate-Constrained Dynamic Graph Evolution

Multi-Source Graph Generator: Unlike traditional GNNs that rely on fixed physical distances or purely data-driven adjacency matrices, this module constructs the graph topology from three parallel sources:
1. Feature-driven graph ( $A_{feat}$ ): Captures global dynamic patterns using a trainable pattern memory.
2. Context-memory graph ( $A_{mem}$ ): Encodes local temporal correlations within the current window.
3. Microstate prior graph ( $A_m$ ): A structural anchor derived from the cosine similarity of microstate activation profiles across electrodes.
Refinement: These sources are fused and passed through a learnable layer to produce a raw dynamic adjacency matrix. The graph is then sparsified (Top-k) and normalized.
Spatiotemporal Aggregation: The refined graph undergoes multi-hop graph convolution (order $K=2$ ) to capture long-range dependencies. A Convolutional Block Attention Module (CBAM) is applied to re-weight channels and spatial positions, emphasizing emotion-relevant regions (prefrontal and temporal areas).

3. Microstate-Aware Prototypical Contrastive Domain Adaptation

To address cross-subject variability, the framework employs a prototypical contrastive learning strategy.
Prototype Construction: Class prototypes are computed as the mean of features for labeled source samples. For the unlabeled target domain, confident predictions are clustered to generate target prototypes.
Loss Functions:
- Intra-domain Loss: Pulls samples toward their respective class prototypes within the same domain.
- Inter-domain Loss: Minimizes the conditional entropy of samples with respect to the opposite domain's prototypes to align class-conditional distributions.
- Adversarial Loss: A domain discriminator with a gradient reversal layer minimizes global domain classification loss.
The total objective integrates classification, adversarial, and prototypical contrastive losses to learn domain-invariant, semantically coherent representations.

Key Contributions

Novel Dual-Prior Mechanism: The authors introduce a method that encodes microstate temporal dynamics as semantic priors and derives a functional connectivity prior from microstate templates. These are adaptively fused with DE features, complementing conventional spectral features with global brain dynamic information.
Microstate-Constrained Dynamic Graph: The paper designs a dynamic graph evolution module where the microstate prior serves as a structural anchor. This allows the graph topology to track the dynamic reorganization of functional brain networks rather than relying on static templates or purely data-driven connections.
Microstate-Aware Prototypical Contrastive Adaptation: The study introduces a domain adaptation strategy that anchors prototypical alignment on microstate-aware representations. This provides a subject-invariant physiological reference, guiding alignment to preserve emotion discriminability while suppressing subject-specific discrepancies.

Experimental Results

The framework was evaluated on the SEED dataset using a subject-independent Leave-One-Subject-Out (LOSO) protocol.

Performance: MG-PCL achieved an average accuracy of 89.64% ± 6.74%.
Comparison: This result consistently outperformed state-of-the-art baselines, including the strong PCL-TDGCN (88.08% ± 8.05%), DGCNN (79.95%), and traditional methods like SVM (58.18%).
Ablation Studies:
- Removing the microstate feature branch reduced accuracy to 88.08%.
- Removing the microstate graph constraint reduced accuracy to 88.27%.
- Removing the microstate-aware prototypical contrastive losses caused the most significant drop, reducing accuracy to 84.62%, highlighting the critical role of prototype-based semantic alignment.
Interpretability: Analysis of the learned graph topology revealed high degree centrality in prefrontal and temporal regions, consistent with known emotional circuitry. Confusion matrices showed high accuracy for positive (94.20%) and neutral (90.80%) emotions, with lower accuracy for negative emotions (86.80%), likely due to physiological similarities between negative and neutral states.

Significance and Claims

The paper claims that microstate-guided dynamic graph modeling is an effective route for capturing task-relevant neural connectivity in EEG-based emotion recognition. By integrating neurophysiologically grounded microstate priors into both feature extraction and graph topology learning, the model overcomes the limitations of static connectivity and purely data-driven domain adaptation. The results suggest that leveraging the millisecond-scale switching of large-scale brain networks (microstates) significantly enhances generalization across subjects, offering a robust solution for cross-subject emotion recognition without requiring subject-specific calibration.

The authors note that while the framework is robust, future work will focus on extending validation to cross-dataset scenarios (e.g., SEED-IV), exploring adaptive microstate template selection, and analyzing frequency-band specific contributions to recognition performance.

Problem Statement

Methodology: MG-PCL Framework

Key Contributions

Experimental Results

Significance and Claims

More like this