Research map / Label-efficient and robust learning / Recognition and long-tail
Long-Term EEG Partitioning for Seizure Onset Detection
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 SODor, a two-stage framework that reformulates seizure onset detection as subsequence clustering over graph-based channel-logit representations, using Toeplitz graphical-lasso model-based clustering to output onset transitions.
Problem
Classification-based EEG seizure detectors produce abrupt misclassifications and do not explicitly model seizure onset (SO) timing over long sequences; post-processing needs manual tuning.
Approach
Learn channel-wise logits with a graph (diffusion conv + GRU) classifier supervised by BCE on max-pooled channel logits, then cluster sliding-window epochs into normal/seizure subsequences via Toeplitz graphical-lasso model-based clustering with a temporal-consistency penalty; state transitions are reported as seizure onsets.
Key results
- On CHB-MIT, SODor achieves NMI 0.979, ARI 0.964, ACC 0.981, versus 0.897/0.875/0.882 for the best post-processing baseline (Burrello et al.).
- On HUH, SODor reaches NMI 0.873, ARI 0.823, ACC 0.845; on TUH, NMI 0.842, ARI 0.863, ACC 0.879.
- Replacing the Toeplitz graphical lasso clustering with MDL or DPGMM drops CHB-MIT NMI to 0.720 and 0.613 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SODor instead of the strongest reported baseline | SO detection · CHB-MIT | NMI: 0.897 -> 0.979 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · CHB-MIT | ARI: 0.875 -> 0.964 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · CHB-MIT | ACC: 0.939 -> 0.981 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · HUH | NMI: 0.778 -> 0.873 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · HUH | ARI: 0.744 -> 0.823 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · HUH | ACC: 0.810 -> 0.845 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · TUH | NMI: 0.764 -> 0.842 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · TUH | ARI: 0.743 -> 0.863 | Helped |
| use SODor instead of the strongest reported baseline | SO detection · TUH | ACC: 0.816 -> 0.879 | Helped |
| replace SODor clustering with MDL (ablation) | SODor · CHB-MIT | NMI: 0.979 ± 0.04 -> 0.720 ± 0.13 | Hurt |
| replace SODor clustering with MDL (ablation) | SODor · CHB-MIT | ARI: 0.964 ± 0.04 -> 0.694 ± 0.13 | Hurt |
| replace SODor clustering with MDL (ablation) | SODor · CHB-MIT | ACC: 0.981±0.04 -> 0.649 ± 0.14 | Hurt |
Surprising findings
- Mean and weighted pooling reach the same final accuracy (0.816) as max pooling on the TUH case, yet produce distorted channel logits, so accuracy alone does not reveal representation quality.
Limitations
- Number of clusters is fixed to two (normal/seizure) for SO detection and three for SPO, rather than learned.
- Performance on TUH is lower, attributed to diverse seizure types and rapid transitions; the method still shows mismatches versus ground truth in visualizations.
Experimental setup
Datasets: CHB-MIT HUH TUH
Benchmarks: CHB-MIT HUH TUH
Metrics: NMI ARI ACC
Baselines: Burrello et al. 2020 post-processing Boony 2021 post-processing Li et al. 2021 post-processing Siyi et al. ICLR 2022 Thi et al. AAAI 2023 Time2State E2Usd MDL clustering DPGMM
Builds on: Time2State E2Usd graphical lasso diffusion convolution Toeplitz graphical lasso
Closest papers by meaning
- REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates (ICML 2024 · similarity 0.88)
- DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection (NeurIPS 2024 · similarity 0.87)
- EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks (NeurIPS 2025 · similarity 0.86)
- Seizure-Semiology-Suite($S^3$): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding (ICML 2026 · similarity 0.86)
- REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects (NeurIPS 2025 · similarity 0.84)
- FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction (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.