The release uses vEpiSet, not private hospital data. Download the dataset only from its authoritative Figshare record:
- DOI: https://doi.org/10.6084/m9.figshare.28069568
- Version used for this release: v2
- Archive:
vepiset-dataset.zip - Archive size: 16,917,002,806 bytes
- MD5:
3b46ecbce87f07c438ea27be7c44b99d - License: CC BY 4.0
The source publication describes 84 subjects, 28 hours of continuous EEG, 2,516 IED epochs, and 22,933 non-IED epochs. IED epochs are divided into five spatial groups: generalized, frontal, temporal, occipital, and centro-parietal.
The training code expects NumPy windows with shape [channels, time] under:
Dataset/
├── Centro-Parietal-IED/
├── Frontal-IED/
├── Generalized-IED/
├── Non-IED/
├── Occipital-IED/
└── Temporal-IED/
Each window covers 4 seconds at 500 Hz before model-specific resampling. The first 19 channels follow the standard 10-20 order:
Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2,
F7, F8, T3, T4, T5, T6, Fz, Cz, Pz
The original archive reports 25,449 labeled epochs. The release pipeline found 25,384 usable NumPy windows after conversion and eligibility checks:
| Class | Windows |
|---|---|
| Centro-Parietal-IED | 366 |
| Frontal-IED | 458 |
| Generalized-IED | 573 |
| Non-IED | 22,870 |
| Occipital-IED | 417 |
| Temporal-IED | 700 |
Subject identity is parsed from the public vEpiSet filename prefix. A seeded 30,000-trial search assigns entire subjects to 70/15/15 target partitions while penalizing class-distribution imbalance and any missing IED subtype.
| Split | Subjects | Windows | IED | Non-IED |
|---|---|---|---|---|
| Train | 58 | 17,578 | 1,764 | 15,814 |
| Validation | 13 | 3,898 | 316 | 3,582 |
| Test | 13 | 3,908 | 434 | 3,474 |
Seed: 2026. No subject appears in more than one partition. Validation is used
for early stopping, model selection, threshold selection, and ensemble choice.
The test partition remains untouched until final evaluation.
scripts/download_vepiset.py downloads the exact archive, verifies the MD5,
and optionally extracts it. The 16.9 GB dataset is intentionally not committed
to GitHub and is not silently mirrored by this project.
@article{lin2025vepiset,
title = {An EEG dataset for interictal epileptiform discharge with spatial distribution information},
author = {Lin, Nan and Zheng, Mengxuan and Li, Lian and Hu, Peng and Gao, Weifang and others},
journal = {Scientific Data},
volume = {12},
pages = {229},
year = {2025},
doi = {10.1038/s41597-025-04523-8}
}