You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is provided for academic research and NeuroMM-2026 challenge participation only. By requesting access, your team confirms that all submitted information is accurate and complete. The dataset, annotations, and any derived files must not be redistributed, mirrored, modified, or used for commercial purposes without prior written permission. Access will be granted only after manual review by the NeuroMM-2026 organizers.
Log in or Sign Up to review the conditions and access this dataset content.
NeuroMM-2026 — Train + Val Open Release
This is the train + val partition of the NeuroMM-2026 multimodal seizure detection challenge dataset. It contains 25,426 EEG samples (20,298 train / 5,128 val) and matching pre-extracted visual features from 7 vision backbones. Test data is not included and will be evaluated by the organizers on a private leaderboard.
Challenge website: https://2026.neuromm.org
Dataset Access Form
Please follow this format before submitting the gated form. Many requests are rejected because the team information does not match the expected format.
Example Application
| Field | Example |
|---|---|
| Team Name | GML-MM-Lab |
| Team Leader Name | Alice Chen |
| Team Leader Email | alice.chen@university.edu |
| Team Members (comma-separated) | Alice Chen, Bob Li, Carol Wang |
| Organization / University / Company | Guangdong Laboratory of Artificial Intelligence and Digital Economy (Shenzhen) |
| Country / Region | China |
- Submit one request per team, not one request per member.
- The request should be submitted by the team leader or the main contact person.
- Make sure the team name and member list are consistent with your challenge registration.
Tasks
- Task 1 — Binary spike-vs-non-spike classification (EEG only)
- Task 2 — Binary spike classification (EEG + Video)
- Task 3 — 5-class seizure subtype classification (positives only, EEG + Video)
Layout
NeuroMM-2026/
├── README.md ← this file
├── annotations/
│ └── neuromm2026_train_val.csv ← 25,426 rows (20,298 train / 5,128 val)
├── splits/
│ └── split.md ← patient-level partition documentation
└── archives/
├── eeg.tar ← 25,426 raw EEG .npy
├── video_clip-base.tar ← OpenAI CLIP ViT-B/32 features (8, 512)
├── video_videomae-base.tar ← VideoMAE-base features (1, 768)
├── video_videomae-large.tar ← VideoMAE-large features (1, 1024)
├── video_dinov2-base.tar ← DINOv2-base features (8, 768)
├── video_dinov2-large.tar ← DINOv2-large features (8, 1024)
├── video_siglip-base.tar ← SigLIP-base features (8, 768)
└── video_timesformer-k400.tar ← TimeSformer (Kinetics-400) (1, 768)
After extraction:
tar -xf archives/eeg.tar
# -> processed/features/eeg/{sample_id}.npy
tar -xf archives/video_clip-base.tar
# -> processed/features/video/clip-base/{sample_id}.npy
Manifest Columns
| Column | Description |
|---|---|
sample_id |
Unique window identifier (matches the .npy filename stem) |
split |
train (20,298) or val (5,128); patient-disjoint |
label |
Binary label: 1 = spike / seizure positive, 0 = negative |
label_type |
Multi-class label: 0 = negative, 1–5 = seizure subtype |
subject_id |
Patient identifier; use to enforce patient-disjoint cross-validation |
EEG Feature Format
Each .npy is a NumPy array shape (29, 2000) (mixed float16 / float32):
- 29 raw EEG channels at 500 Hz
- 2000 timesteps = 4-second window
The reference baseline derives 23 + 3 ECG/EMG = 26 channels via differential pairs. See the official baseline repository for the loader.
Loading Example
import numpy as np, pandas as pd
df = pd.read_csv("annotations/neuromm2026_train_val.csv")
print(df["label_type"].value_counts()) # multi-class distribution
sid = df.iloc[0]["sample_id"]
x = np.load(f"processed/features/eeg/{sid}.npy")
print(x.shape, x.dtype) # (29, 2000)
# Video feature for the same sample (CLIP-base)
v = np.load(f"processed/features/video/clip-base/{sid}.npy")
print(v.shape, v.dtype) # (8, 512)
Important: Patient-Level Splitting
The provided train/val split is patient-disjoint (no patient appears in both). If you do additional CV folds, also enforce patient-level splitting via subject_id to avoid label leakage.
License
CC BY-NC 4.0 — academic research and NeuroMM-2026 challenge participation only. No redistribution, no commercial use.
Citation
If you use this dataset, please cite:
@dataset{neuromm2026,
title = {NeuroMM-2026: Multimodal Seizure Detection Dataset},
year = {2026},
url = {https://2026.neuromm.org}
}
- Downloads last month
- 122