GREEN: A lightweight architecture using learnable wavelets and Riemannian geometry for biomarker exploration with EEG signals

Joseph Paillard, Jörg F. Hipp, Denis A. EngemannPatterns 2025 · DOI · OpenReview

TL;DR

Researchers have developed GREEN, a novel, lightweight deep learning architecture for analyzing EEG signals. By combining learnable wavelets with Riemannian geometry, GREEN achieves state-of-the-art performance on various prediction tasks while being significantly more computationally efficient and interpretable than existing deep learning models, paving the way for more accessible biomarker discovery from brain activity.

What Problem Does This Paper Address?

Electroencephalography (EEG) is a powerful tool for understanding brain activity, with applications ranging from monitoring sleep to diagnosing neurological disorders. While traditional EEG analysis methods offer interpretability, they can miss subtle patterns. Deep learning (DL) models excel at uncovering complex patterns but often lack interpretability and can be computationally demanding, creating a gap between predictive power and neuroscientific understanding. This paper addresses the challenge of developing a DL model for EEG that is both highly performant and interpretable, bridging the gap between cutting-edge AI and established neuroscience concepts.

Methods

The core innovation is the GREEN (Gabor Riemann EEGNet) architecture. It integrates:

  • Learnable Gabor Wavelets: Instead of using fixed filter banks, GREEN employs parameterized Gabor wavelets whose frequency and standard deviation are learned during training. This allows the model to adapt its spectral analysis to the specific task, potentially capturing more relevant frequency bands.
  • Riemannian Geometry: EEG data, particularly covariance matrices derived from it, can be analyzed using Riemannian geometry. This framework provides a mathematically sound way to handle these matrices, mitigating distortions caused by volume conduction and allowing for the use of classical machine learning tools. GREEN incorporates operations like shrinkage, BiMap, and LogMap to process these geometric representations.
  • Modular Design: GREEN is built from composable blocks, allowing researchers to easily test hypotheses about the impact of different components (e.g., adding a hidden layer, incorporating phase synchrony measures) on performance.

Datasets

The study utilized three public EEG datasets:

  • Temple University Hospital EEG (TUAB): Used for age regression and pathology classification.
  • Two Decades Brainclinics Research Archive for Insights in Neurophysiology (TDBRAIN): Used for binary sex classification and eyes-open/eyes-closed classification.
  • Chung-Ang University (CAU) Hospital dataset: Used for three-way classification of diagnosis (healthy, mild cognitive impairment, dementia).

The paper also mentions reproducing benchmarks from existing studies, including those for BCI applications.

Key Insights and Ideas

  • Learnable Wavelets Outperform Fixed Filters: For many tasks, learned Gabor wavelets captured task-relevant information more effectively than predefined filter banks, leading to improved performance.
  • Riemannian Geometry Aids Interpretability and Efficiency: Integrating Riemannian geometry allows for processing of covariance matrices in a way that is both mathematically robust and more interpretable, while also contributing to the model's lightweight nature.
  • Task-Specific Spectral Signatures: The optimal number and frequencies of wavelets varied significantly across different prediction tasks, suggesting that the brain's spectral signatures for different cognitive states or conditions are not uniform.
  • Compactness is Key: GREEN's ability to achieve high performance with orders of magnitude fewer parameters than large DL models is a significant advantage for reproducibility, accessibility, and computational efficiency.
  • Neuroscience Priors are Valuable: By incorporating concepts like wavelets (representing neural oscillations) and Riemannian geometry (handling signal correlations), GREEN demonstrates the power of integrating domain knowledge into DL architectures.

Key Results

  • Performance: GREEN outperformed non-deep state-of-the-art models and performed favorably against large DL models on six prediction tasks across four datasets with over 5,000 participants.
  • Efficiency: GREEN used orders-of-magnitude fewer parameters than comparable DL models, leading to significantly faster computation times (e.g., over 400% advantage in training time compared to EEGConformer on a specific task).
  • Interpretability: The architecture naturally yields interpretable quantities like power topographies and allows for direct visualization of learned spectral features.
  • Modularity Benefits: Adding components like learned wavelets (G3) significantly improved performance for tasks like dementia diagnosis and eyes-closed/eyes-open classification. Incorporating phase-locking value (PLV) matrices also provided performance gains for certain tasks.

Limitations and Future Work

  • Non-Sinusoidal Waveforms: The current Gabor wavelets are sinusoidal. Future work could explore incorporating harmonic wavelets to capture non-sinusoidal waveforms.
  • Cross-Frequency Coupling: While GREEN can implicitly capture some cross-frequency interactions, explicitly modeling them (e.g., through power-envelope covariances) could be a future direction.
  • High-Density EEG: The study did not extensively explore GREEN's performance on very high-density EEG (128+ channels), which remains an open question.
  • Interventional Studies: The framework's potential for exploring specific medical hypotheses in interventional studies is yet to be fully realized.
  • Self-Supervised Learning: Extending GREEN with self-supervised learning could improve its ability to transfer knowledge across datasets and centers, addressing the challenge of limited patient data in clinical trials.

Summary updated

BibTeX
@article{paillard2025green,
  author = {Joseph Paillard and Jörg F. Hipp and Denis A. Engemann},
  title = {{GREEN: A lightweight architecture using learnable wavelets and Riemannian geometry for biomarker exploration with EEG signals}},
  journal = {Patterns},
  year = {2025},
  doi = {10.1016/j.patter.2025.101182},
}
Figure