SenuaLab/EEG-IED-Detection

SenuaLab EEG IED Detection

Five patient-independent binary IED checkpoints, aggregate benchmark artifacts, and an app-ready model pack for SenuaLab EEG Annotation Tool. Training and evaluation use only the public vEpiSet dataset.

Research use only. These models are not medical devices and must not be used as the sole basis for diagnosis or treatment. Every candidate requires review by a qualified EEG professional against the original recording.

Privacy. No private hospital recording, patient identifier, physician annotation, local path, or hospital-derived patient result is included. The model parameters, public configuration metadata, and aggregate vEpiSet results were reviewed before upload.

Source, methodology, and reproducibility documentation: github.com/SenuaLab/EEG-IED-Detection

Models

Folder	Model	Params	Rate	Window	Test AUROC	Test AUPRC	Test F1
iednet_lite	IEDNet-Lite	0.25M	250 Hz	4 s	0.8793	0.6848	0.6372
resnet_attention	ResNet-Attention	4.99M	250 Hz	4 s	0.8998	0.7018	0.6667
eegpt	EEGPT + temporal head	25.59M	250 Hz	4 s	0.8641	0.6532	0.6139
eegpt_linear_probe	EEGPT linear probe	25.32M	250 Hz	4 s	0.7519	0.3722	0.3227
eegdino_medium	EEG-DINO Medium	34.45M	200 Hz	4 s	0.9363	0.8020	0.7415

All test thresholds were selected on validation subjects and then frozen. The test split contains 3,908 windows from 13 held-out vEpiSet subjects. Results are single-dataset, four-second window metrics, not prospective clinical claims.

Files

models/<name>/
├── model.safetensors     # tensor-only weights (preferred)
└── config.json           # reviewed input/provenance/metric contract

pytorch/<name>/
├── model.pth             # sanitized checkpoint for the Python runner
└── config.json

eegannotationtool/senua_<name>/
├── model.py              # executable architecture adapter
├── preprocessing.py      # released preprocessing contract
├── model.pth             # app-compatible sanitized checkpoint
└── manifest.json

ensembles/                # validation-selected ensemble definitions
results/                  # aggregate benchmark table and figure
MANIFEST.json             # SHA-256 and byte size for every release file

PyTorch .pth is included because the current EEG Annotation Tool model discovery contract supports .pth/.pt/.ckpt. Prefer safetensors elsewhere and load executable/model files only from this official repository.

Input contract

• 19 standard 10-20 channels in this order: Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz, Pz.
• Four-second full windows.
• Fourth-order zero-phase 1-45 Hz Butterworth band-pass.
• Common-average reference across the 19 channels.
• Resampling to the rate listed in the table.
• One global z-score over the complete channel-time window, clipped to [-8,8].
• Binary output order: [Non-IED, IED].

Legacy temporal aliases T7/T8/P7/P8 map to T3/T4/T5/T6.

Download

hf download SenuaLab/EEG-IED-Detection \
  pytorch/iednet_lite/model.pth \
  --local-dir ./weights

Or from Python:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="SenuaLab/EEG-IED-Detection",
    filename="pytorch/iednet_lite/model.pth",
)

The source repository runner accepts model short names and handles the download:

python -m eeg_ied_detector.robust_predict \
  --data-dir /path/to/brainvision-recordings \
  --model-paths iednet_lite resnet_attention \
  --output-dir ./predictions \
  --no-label-comparison

EEGPT and EEG-DINO require braindecode[hub]==1.6.1.

EEG Annotation Tool

Use the installer from the source repository:

python scripts/install_eeg_annotation_models.py \
  --models iednet_lite resnet_attention

IED Finder discovers the installed folders, validates channels and the declared input contract, and uses the validation-selected threshold as its starting point. Foundation-model adapters are included but need Braindecode in the application's Python environment.

Evaluation limitations

• One public dataset and one fixed internal held-out split.
• No prospective, external multi-center, or medical-device validation.
• Four-second window metrics do not equal continuous event-level sensitivity.
• Performance can shift with montage, hardware, preprocessing, age, disease mix, artifacts, prevalence, and annotation policy.
• Threshold changes require a prospectively defined validation protocol and must not use the final evaluation cohort.
• Automated output is a review candidate, never a diagnosis.

Model and data provenance

• vEpiSet: Lin et al., Scientific Data 12, 229 (2025), https://doi.org/10.1038/s41597-025-04523-8, CC BY 4.0.
• EEGPT: NeurIPS 2024, upstream code BINE022/EEGPT, encoder checkpoint braindecode/eegpt-pretrained.
• EEG-DINO: MICCAI 2025, paper, encoder checkpoint braindecode/eegdino-medium-pretrained at revision 191f73eb50d68a184b4cabb623938df914235d3c.

See MODEL_LICENSE.md for the per-artifact licensing boundary and upstream conditions.