Research map / Graph learning / Knowledge and temporal graphs
EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EvoBrain, a time-then-graph dynamic GNN with explicit per-snapshot EEG graphs (two-stream Mamba plus LapPE-GCN) for seizure detection and early prediction, with theoretical expressivity analysis.
Problem
Seizure detection and early prediction from EEG require capturing evolving brain connectivity, but existing dynamic GNNs use temporally fixed static graphs and handle temporal and graph interactions poorly.
Approach
Builds per-snapshot dynamic channel-correlation graphs, models node and edge temporal evolution with two separate Mamba streams (time-then-graph), then applies a GCN with Laplacian positional encoding; theoretically shows explicit dynamic modeling and time-then-graph are more expressive under 1-WL.
Key results
- On TUSZ 60s seizure detection, EvoBrain reaches AUROC 0.865 and F1 0.483, versus EvolveGCN's 0.670 and 0.340.
- On TUSZ 12s detection, EvoBrain achieves AUROC 0.877 and F1 0.539, above LaBraM (0.825, 0.472).
- EvoBrain uses about 30x fewer parameters than LaBraM (114,794 trainable parameters), and trains up to 17x faster than DCRNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EvoBrain instead of LaBraM | TUSZ seizure detection 12s · TUSZ seizure detection 12s | AUROC: 14% | Helped |
| Use EvoBrain instead of LaBraM | TUSZ seizure detection 12s · TUSZ seizure detection 12s | F1: 12% | Helped |
| Use EvoBrain instead of EvolveGCN | TUSZ seizure detection 60s · TUSZ seizure detection 60s | AUROC: 23% | Helped |
| Use EvoBrain instead of EvolveGCN | TUSZ seizure detection 60s · TUSZ seizure detection 60s | F1: 0.340 -> 0.483 | Helped |
| Use time-then-graph instead of graph-then-time or time-and-graph (ablation) | EvoBrain architecture with GRU and GCN layers · TUSZ seizure detection 12s | performance | Helped |
| Replace GCN with GIN (ablation) | EvoBrain · TUSZ seizure detection | performance | No clear effect |
| Remove the GNN and use only the RNN (ablation) | EvoBrain · TUSZ seizure detection | performance | Hurt |
| Replace Mamba with GRU layers (ablation) | EvoBrain · TUSZ seizure detection 60s | performance | Hurt |
Surprising findings
- On early seizure prediction, LaBraM outperforms EvoBrain on the 60s TUSZ setting in AUROC (0.669 vs 0.651) and F1 (0.413 vs 0.401).
Limitations
- Performance on seizure prediction is weaker than LaBraM, which benefits from more parameters and large-scale pretraining.
- Evaluated only on seizure tasks; generalization to other EEG tasks is untested.
- Potential bias across demographic groups and false alarms in early prediction are noted as risks.
Experimental setup
Datasets: TUSZ CHB-MIT
Benchmarks: TUSZ CHB-MIT
Metrics: AUROC F1
Baselines: SVM Random Forest LSTM CNN-LSTM BIOT LaBraM EEGPT EvolveGCN DCRNN GRAPHS4MER GRU-GCN
Builds on: Mamba GCN Laplacian positional encoding 1-WL test GRAPHS4MER
Closest papers by meaning
- REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates (ICML 2024 · similarity 0.88)
- Long-Term EEG Partitioning for Seizure Onset Detection (AAAI 2025 · similarity 0.86)
- DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection (NeurIPS 2024 · similarity 0.86)
- B[FM]$^2$: Brain Foundation Model via Flow Matching with SplitUNet (NeurIPS 2026 · similarity 0.85)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.84)
- THD-BAR: Topology Hierarchical Derived Brain Autoregressive Modeling for EEG Generic Representations (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.