Introduction
Electroencephalography (EEG) serves as a critical non-invasive modality for monitoring brain activity, yet it presents a significant "data nightmare" for researchers. EEG signals are high-dimensional, inherently noisy, and exhibit extreme variability across different individuals. This survey provides a comprehensive roadmap of the rapidly evolving landscape of spatio-temporal EEG analysis, tracing the transition from traditional machine learning to cutting-edge self-supervised learning (SSL), Graph Neural Networks (GNNs), and Large Language Models (LLMs). These advanced architectures aim to bridge the gap between raw electrical signals and high-level semantic interpretations.
The Core Challenges in EEG Analysis
The authors identify three primary hurdles that modern research must overcome:
- Representation: Extracting meaningful features from EEG data in the absence of large-scale labeled datasets.
- Pattern Recognition: Identifying subtle, complex patterns indicative of specific brain states (e.g., seizures or emotional shifts) across diverse subjects.
- Interpretation: Transforming abstract electrical signals into human-readable formats, such as visual images or textual descriptions.
Taxonomy of Modern EEG Analysis
The field is currently defined by three main pillars of research: Representation Learning, Discriminative Analysis, and Generative Analysis.
1. Representation Learning via Self-Supervised Learning (SSL)
To mitigate the scarcity of labeled data, researchers are increasingly turning to SSL. This approach includes Contrastive Learning, which teaches models to distinguish between similar and dissimilar signal segments, and Masked Autoencoders (MAE), which force models to reconstruct missing portions of an EEG signal to learn its underlying temporal and frequency structures.
2. Discriminative Analysis: GNNs and Foundation Models
- Graph Neural Networks (GNNs): By treating EEG electrodes as nodes in a graph, GNNs model the physical and functional connectivity of the brain. This allows for a more biologically relevant analysis than treating signals as simple one-dimensional strings.
- Foundation Models & LLMs: A significant trend involves "reprogramming" models like GPT-2 or LLaMA for EEG data. By pre-training on massive unlabeled datasets, these models develop a general understanding of brain signal dynamics before being fine-tuned for specific clinical or cognitive tasks.
3. Generative EEG Analysis
The survey explores the frontier of Generative AI (AIGC), utilizing Diffusion Models and GANs to perform "brain-to-image" or "brain-to-text" translation. This effectively allows for the visualization of what a subject is seeing or thinking, moving the field from simple classification to complex synthesis.
Key Insights and Results
- The Graph Advantage: GNNs can adaptively learn the "brain graph," focusing on how regions communicate rather than just their physical proximity.
- LLMs as Sequential Processors: LLMs are surprisingly effective at EEG analysis because both language and EEG are complex, multi-dimensional sequences with long-range dependencies.
- Performance Benchmarks: Models like BENDR and LaBraM have shown that SSL pre-training significantly outperforms supervised models in data-scarce environments. In clinical applications, GNN-based models like GraphSleepNet have set new benchmarks in sleep staging, while generative frameworks like DreamDiffusion have achieved high-fidelity image reconstruction from brain signals.
Limitations and Future Directions
Despite these advancements, several challenges remain. EEG datasets, while growing (e.g., the TUH EEG Corpus), are still small compared to the trillions of tokens used in natural language processing. Furthermore, cross-subject generalization remains a "holy grail," as models often fail when applied to new individuals due to physiological differences. Future research is expected to focus on more efficient architectures like Mamba or KAN for real-time monitoring and the expansion into multimodal models that integrate EEG with other sensory data.
Summary updated