Methodology
Standardized evaluation protocols for EEG seizure-detection benchmarks
Working paper · 2025
AI benchmarking for EEG seizure detection
Test, compare, and evaluate seizure-detection models with confidence
SeizureBench is an open research platform that helps universities and labs benchmark EEG seizure-detection algorithms on standardized datasets, shared metrics, and fully reproducible pipelines.
For research and educational purposes only. Not a medical diagnostic device.
- Detection models
- 18+
- Public EEG datasets
- 6
- Evaluation metrics
- 24
- Reproducible runs
- 100%
Detection models
Public EEG datasets
Evaluation metrics
Reproducible runs
About
A common ground for evaluating EEG seizure detection
Seizure-detection research moves fast, but comparing results is hard when every group uses different datasets, preprocessing, and metrics. SeizureBench provides a neutral, transparent environment where models are measured under identical conditions.
The platform is intended for academic and educational use — helping students, engineers, and clinical researchers understand how detection algorithms behave, where they excel, and where they fall short. SeizureBench does not provide clinical decisions and is not intended for patient care.
Standardized evaluation
Every model runs against the same datasets, splits, and metrics so results are directly comparable across labs and publications.
Reproducible by design
Pipelines, preprocessing, and random seeds are versioned and logged, making it straightforward to re-run and verify any benchmark.
Built for rigor
Report sensitivity, false-alarm rates, latency, and more with clinically-informed metrics that reflect real research needs.
For the community
Designed with academic researchers in mind, encouraging open comparison, shared baselines, and transparent methodology.
Features
Everything a lab needs to benchmark detection models
A complete evaluation workflow — from raw EEG to publication-ready comparison — built around transparency and reproducibility.
Bring your own model
Wrap a model with a lightweight interface and submit it for evaluation against curated EEG benchmarks.
Side-by-side comparison
Compare detection performance across models with sortable leaderboards and shared axes.
Signal preprocessing
Configurable filtering, montage selection, and windowing applied uniformly to every run.
Clinically-informed metrics
Sensitivity, specificity, false-alarm rate per hour, detection latency, and event-based scoring.
Dataset management
Standardized splits across public EEG corpora so experiments stay comparable and honest.
Reproducible reports
Every benchmark generates a versioned report with configuration, seeds, and environment details.
Scalable execution
Queue evaluation jobs and run large model sweeps without managing infrastructure by hand.
Exportable results
Download raw predictions and metrics as CSV or JSON for your own analysis and papers.
Models
Sample benchmark leaderboard
An illustrative comparison of detection models evaluated on a shared EEG benchmark. Metrics shown are sample values for demonstration and do not reflect validated clinical performance.
| Rank | Model | Architecture | Sensitivity | Specificity | FA / 24h | Latency |
|---|---|---|---|---|---|---|
| 1 | ChronoNet-XL | Recurrent CNN | 94.2% | 91.8% | 1.4 | 3.1s |
| 2 | EEG-Transformer | Attention | 92.7% | 93.1% | 1.1 | 4.6s |
| 3 | SpectroNet | Spectrogram CNN | 90.5% | 90.2% | 2 | 2.4s |
| 4 | TCN-Seizure | Temporal Conv | 89.1% | 88.7% | 2.6 | 1.9s |
| 5 | BiLSTM-Base | Recurrent | 86.4% | 87% | 3.3 | 2.8s |
FA / 24h = false alarms per 24 hours. Values are illustrative sample data for the research platform preview.
Research
Grounded in open, reproducible science
SeizureBench builds on publicly available EEG corpora and transparent evaluation methods so findings can be independently verified and extended.
Benchmark
A reproducible baseline suite across public scalp-EEG datasets
Preprint · 2025
Analysis
False-alarm characterization in deep seizure-detection models
Working paper · 2024
Supported datasets
Benchmarks are defined over widely used, openly documented EEG datasets. Access to the underlying data is subject to each source's own license and terms.
- CHB-MIT
- TUH EEG Seizure
- Siena Scalp EEG
- Bonn
- Freiburg
- Helsinki
“Shared benchmarks are how a field learns which ideas truly work. Reproducibility is not optional — it is the science.”
Demo — Coming Soon
An interactive demo is on the way
We're building a hands-on environment where you can load sample EEG recordings, run detection models, and explore results in real time. Join the early-access list to be notified when it launches.
Launch Demo (unavailable)Notify me at launch
Interactive EEG viewer
In progress
Live model inference playground
Coming soon
Public leaderboard submissions
Planned
Contact
Collaborate with the SeizureBench team
Whether you're a researcher, educator, or student, we'd love to hear how you plan to use the platform. Reach out about early access, dataset contributions, or academic collaborations.
Email
research@seizurebench.org
Institution
Neural Signals & Machine Learning Research Group
Location
Remote-first · Global collaboration