EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks

TL;DR

EvoBrain introduces explicit dynamic multi-channel EEG graphs and a time-then-graph processing framework to model evolving brain networks for seizure detection and early prediction. It builds per-snapshot graphs from frequency-domain EEG features using cross-correlation, and applies a two-stream Mamba temporal model followed by a Laplacian-encoded GCN to capture spatiotemporal dynamics. The work provides a 1-WL based expressivity analysis showing explicit dynamic modeling and time-then-graph architectures strictly outperform alternatives, and demonstrates substantial AUROC and F1 improvements with far fewer parameters than large foundation models. Empirically, EvoBrain achieves state-of-the-art performance among lightweight dynamic GNNs and shows clinically relevant patterns in learned networks, suggesting potential for SOZ localization and neuromodulation planning while remaining computationally efficient.

Abstract

Dynamic GNNs, which integrate temporal and spatial features in Electroencephalography (EEG) data, have shown great potential in automating seizure detection. However, fully capturing the underlying dynamics necessary to represent brain states, such as seizure and non-seizure, remains a non-trivial task and presents two fundamental challenges. First, most existing dynamic GNN methods are built on temporally fixed static graphs, which fail to reflect the evolving nature of brain connectivity during seizure progression. Second, current efforts to jointly model temporal signals and graph structures and, more importantly, their interactions remain nascent, often resulting in inconsistent performance. To address these challenges, we present the first theoretical analysis of these two problems, demonstrating the effectiveness and necessity of explicit dynamic modeling and time-then-graph dynamic GNN method. Building on these insights, we propose EvoBrain, a novel seizure detection model that integrates a two-stream Mamba architecture with a GCN enhanced by Laplacian Positional Encoding, following neurological insights. Moreover, EvoBrain incorporates explicitly dynamic graph structures, allowing both nodes and edges to evolve over time. Our contributions include (a) a theoretical analysis proving the expressivity advantage of explicit dynamic modeling and time-then-graph over other approaches, (b) a novel and efficient model that significantly improves AUROC by 23% and F1 score by 30%, compared with the dynamic GNN baseline, and (c) broad evaluations of our method on the challenging early seizure prediction tasks.

Figures (7)

• Figure 1: ROC curve results for the 12-second seizure detection task on two datasets.
• Figure 2: Comparison of the proposed dynamic graph structure and the static structure.
• Figure 3: (a) Inference and (b) training time vs. input data length on CHB-MIT and TUSZ datasets. Our model achieves up to 17x faster training times and 14x faster inference times than its competitors, demonstrating scalability.
• Figure 4: (a) Architecture evaluation using same GNN and RNN layers. (b) Results using raw EEG instead of frequency-domain features. (c) GNN and (d) RNN layer evaluation.
• Figure 5: Learned graph structure visualizations. The yellow color of the edges indicates the strength of the connections. In (a) Normal state, the dark shows weak connections. In (b) Pre-seizure state, the connections in specific regions strengthen over time. In (c) Focal seizure, which occurs only in a specific area of the brain, strong connections are consistently present in a particular region. In (d) Generalized seizure, strong connections are observed across the entire brain.

Theorems & Definitions (19)

• Definition 1: Implicit Dynamic Graph Modeling - Static Structure
• Definition 2: Explicit Dynamic Graph Modeling - Dynamic Structure
• Definition 3: Graph-then-time
• Definition 4: Time-and-graph
• Definition 5: Time-then-graph
• Theorem 1
• proof
• Lemma 1
• proof
• Lemma 2