[Paper Review] BrainNet: Epileptic Wave Detection from SEEG with Hierarchical Graph Diffusion Learning

Junru Chen, Yang Yang+4 more|arXiv (Cornell University)|Jun 15, 2023

EEG and Brain-Computer InterfacesNeuroscience3 citations

TL;DR

This paper proposes BrainNet, a hierarchical graph diffusion learning model that jointly detects epileptic waves and infers dynamic epileptogenic networks from stereo-EEG (SEEG) data. By modeling brain wave propagation across channels, brain regions, and patients using self-supervised learning, BrainNet achieves state-of-the-art performance on a large-scale real-world SEEG dataset, outperforming existing time-series-based baselines in detecting epileptic waves under severe label imbalance and noise.

ABSTRACT

Epilepsy is one of the most serious neurological diseases, affecting 1-2% of the world's population. The diagnosis of epilepsy depends heavily on the recognition of epileptic waves, i.e., disordered electrical brainwave activity in the patient's brain. Existing works have begun to employ machine learning models to detect epileptic waves via cortical electroencephalogram (EEG). However, the recently developed stereoelectrocorticography (SEEG) method provides information in stereo that is more precise than conventional EEG, and has been broadly applied in clinical practice. Therefore, we propose the first data-driven study to detect epileptic waves in a real-world SEEG dataset. While offering new opportunities, SEEG also poses several challenges. In clinical practice, epileptic wave activities are considered to propagate between different regions in the brain. These propagation paths, also known as the epileptogenic network, are deemed to be a key factor in the context of epilepsy surgery. However, the question of how to extract an exact epileptogenic network for each patient remains an open problem in the field of neuroscience. To address these challenges, we propose a novel model (BrainNet) that jointly learns the dynamic diffusion graphs and models the brain wave diffusion patterns. In addition, our model effectively aids in resisting label imbalance and severe noise by employing several self-supervised learning tasks and a hierarchical framework. By experimenting with the extensive real SEEG dataset obtained from multiple patients, we find that BrainNet outperforms several latest state-of-the-art baselines derived from time-series analysis.

Motivation & Objective

• To address the challenge of detecting epileptic waves in high-dimensional, noisy, and imbalanced stereo-EEG (SEEG) data from individual patients.
• To model the dynamic propagation of epileptic activity across brain regions by learning patient-specific epileptogenic networks.
• To overcome label imbalance and noise in SEEG data through a self-supervised learning framework within a hierarchical architecture.
• To develop an end-to-end, data-driven method that jointly learns diffusion patterns and detects epileptic wave onset in real-time clinical settings.
• To enable clinical deployment via an online system that supports physician review and interpretation of model predictions.

Proposed method

• BrainNet employs a hierarchical framework that models epileptic wave diffusion at three levels: channel-level, brain region-level, and patient-level.
• It learns dynamic diffusion graphs using a graph neural network (GNN) that captures time-evolving interactions between SEEG channels and brain regions.
• The model integrates multiple self-supervised pre-training tasks—such as masked signal reconstruction and contrastive learning—to improve robustness against label scarcity and noise.
• A multi-level attention mechanism aggregates features across channels and brain regions to enhance detection sensitivity and spatial localization.
• The architecture combines temporal modeling via 1D convolutions and graph convolutional layers to capture both local signal patterns and long-range propagation dynamics.
• The framework is trained end-to-end on real SEEG data, with loss functions optimized for both epileptic wave detection and graph structure learning.

Experimental results

Research questions

• RQ1 How can we effectively detect epileptic waves in noisy, imbalanced, and patient-specific SEEG data?
• RQ2 Can we jointly learn the dynamic diffusion graph of epileptic wave propagation and improve detection performance?
• RQ3 To what extent can self-supervised learning improve model robustness in low-resource, high-noise SEEG settings?
• RQ4 Can a hierarchical graph diffusion model better capture the epileptogenic network structure than conventional time-series models?
• RQ5 How interpretable and clinically useful are the learned diffusion graphs for identifying seizure onset zones?

Key findings

• BrainNet outperforms multiple state-of-the-art time-series models on epileptic wave detection in SEEG data, achieving higher F1-scores and better precision-recall balance.
• The model demonstrates superior robustness to label imbalance and noise, with significant improvements in detection performance when self-supervised pre-training is applied.
• The learned dynamic diffusion graphs show strong alignment with clinical knowledge, revealing plausible propagation patterns across brain regions in individual patients.
• The online system deployment enables real-time visualization of predictions, with doctors able to validate model outputs against ground-truth annotations.
• The hierarchical architecture enables effective feature representation at multiple scales, improving detection accuracy across diverse patient-specific SEEG signal patterns.
• The model's performance is validated on a large-scale real-world SEEG dataset with 53-hour average recordings per patient, demonstrating scalability and clinical relevance.