Research map / Architectures and efficiency / RNNs and normalization
REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates
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 REST, a tiny graph-based residual state update model (about 8.4K parameters, 37KB) with binary random masking and repeated weight-shared updates for fast EEG seizure detection and classification on TUSZ and CHB-MIT.
Problem
EEG seizure detection and classification models are large, memory-hungry and slow at inference, which limits real-time use in implantable or edge clinical devices such as RNS and DBS.
Approach
Replace RNN gating with a residual state update: the state is an affine map of input and previous state plus a graph-convolution increment over the electrode graph, with binary random masking of the increment and repeated updates that reuse the same weights per time step.
Key results
- REST (RM) reaches 83.6% AUROC on TUSZ at 10-s clips with 1.292 ms inference and 8.4K parameters (37KB), about 9x faster than DCRNN.
- On CHB-MIT, REST (RM) reaches 96.7% AUROC at 4-s clips and is the only model above 90% AUROC at 10-s clips (89.2 is RM; RS reaches 93.5).
- REST uses roughly 14x less memory than the smallest competing seizure model and 38x fewer parameters than DCRNN w/SS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multiple random updates instead of a single deterministic update (ablation) | REST · TUSZ seizure detection across clip lengths | AUROC | Helped |
| use multiple random updates instead of a single random update (ablation) | REST · TUSZ 10-s seizure classification | Weighted F1-score: 0.57 -> 0.60 | Helped |
| use multiple random updates instead of a single deterministic update (ablation) | REST · TUSZ 10-s seizure classification | Weighted F1-score: 0.51 -> 0.60 | Helped |
| use MSE loss instead of BCE loss (ablation) | REST · TUSZ 10-s seizure detection validation | AUROC: 80.4 -> 83.6 | Helped |
| remove binary random mask during inference (ablation) | REST (RS) · TUSZ 10-s seizure detection | AUROC: 81.8 -> 81.5 | Hurt |
| remove binary random mask during inference (ablation) | REST (RM) · TUSZ 10-s seizure detection | AUROC: 83.6 -> 82.9 | Hurt |
| apply REST random-update strategies to RNN baselines instead of vanilla updates (ablation) | RNN · TUSZ 10-s seizure detection | AUROC: 77.3 -> 80.8 | Helped |
| apply REST random-update strategies to GRU instead of vanilla updates (ablation) | GRU · TUSZ 10-s seizure detection | AUROC: 73.5 -> 73.6 | No clear effect |
| apply REST random-update strategies to LSTM instead of vanilla updates (ablation) | LSTM · TUSZ 10-s seizure detection | AUROC: 70.4 -> 74.7 | Helped |
| include the current input in each iterative state update instead of updating from the previous state alone (ablation) | REST (RM) · Seizure analysis | validation loss | Helped |
| use REST (RM) rather than DCRNN w/SS | Seizure detection models · TUSZ 10-s | inference time: 23.25 ms vs. 1.292 ms | Helped |
| use REST (RM) rather than LSTM | Seizure detection models · TUSZ 10-s | AUROC | Helped |
Surprising findings
- Binary cross-entropy loss prevented residual updates from approaching zero, so MSE loss was used for detection instead.
- The random-mask-only update (RS) with a single update performs worse than multi-update RM, and the deterministic single update (DS) is weakest on TUSZ.
Limitations
- Training time is much longer than baselines (e.g., 70-100 min vs 5-7 min for LSTM at detection).
- Evaluated only on retrospective public datasets (TUSZ and CHB-MIT); no real-device or prospective validation.
- On TUSZ, REST (RM) is slightly below DCRNN w/SS and Transformer at some clip lengths and its classification F1 trails DCRNN w/SS.
Experimental setup
Datasets: TUH EEG Seizure Corpus (TUSZ) v2.0.0 CHB-MIT
Benchmarks: TUSZ CHB-MIT
Metrics: AUROC weighted F1-score inference time (ms) model size (MB) parameter count
Baselines: LSTM GRU ResNet-LSTM ResNet-Dilation-LSTM CNN-LSTM DCRNN DCRNN w/SS Transformer
Builds on: DCRNN GRU LSTM Dropout
Closest papers by meaning
- EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks (NeurIPS 2025 · similarity 0.88)
- Long-Term EEG Partitioning for Seizure Onset Detection (AAAI 2025 · similarity 0.88)
- DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection (NeurIPS 2024 · similarity 0.87)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.85)
- REVE: A Foundation Model for EEG - Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects (NeurIPS 2025 · similarity 0.84)
- Seizure-Semiology-Suite($S^3$): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding (ICML 2026 · similarity 0.83)
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.