Datasets:

NeurdyLab
/

NeuroBOLT

This dataset provides synchronized resting-state EEG and BOLD fMRI ROI time series collected simultaneously from healthy adults. It is intended for research on NeuroAI, cross-modal modeling (EEG-fMRI), multimodal fusion, and hemodynamic modeling.

We release:

Preprocessed fMRI ROI time series (DiFuMo parcellation, n=64).
Preprocessed and resampled EEG time series aligned to fMRI.

For task-condition data, higher-resolution DiFuMo ROIs (>64), or any inquiries, please contact yamin.li@vanderbilt.edu.

📖 Data Collection Summary

Participants: 22 healthy volunteers.
Sessions: Two 20-minute resting sessions per participant (eyes closed); final dataset contains 29 scans after artifact exclusion.
Ethics: Written informed consent obtained. Procedures approved by the Institutional Review Board (IRB).

fMRI

Scanner: 3T
Sequence: Multi-echo gradient-echo EPI
TR: 2100 ms
Condition: Rest (eyes closed)

EEG

System: MR-compatible, 32-channel (10–20), FCz reference (BrainAmps MR, Brain Products GmbH)
Sampling rate: 5 kHz, synchronized to the scanner’s 10 MHz clock (facilitates MR gradient artifact reduction)

🔩 Preprocessing Summary

fMRI → ROI time series

Parcellation: DiFuMo (n = 64) with 2 additional global signals
- global signal clean (cleaned whole-brain average signal, with confounds regressed)
- global signal raw (unprocessed whole-brain average signal)
Confound regression: Motion regressors removed
Temporal filtering: Low-pass filter applied at < 0.15 Hz
Normalization: Demeaned and scaled by the 95th percentile of the absolute amplitude (per ROI)

In the paper we evaluated 7 representative ROIs spanning diverse spatial and functional domains:

Primary sensory: Cuneus, Heschl’s gyrus
High-level cognitive: Precuneus anterior, Middle frontal gyrus anterior
Subcortical: Putamen, Thalamus
Global: global signal clean

EEG

Channels: ECG/EOG/EMG removed → remaining 26 scalp channels
Resampling: 5 kHz → 200 Hz (retains <100 Hz content, improves efficiency)
Alignment: synchronized to fMRI acquisition for one-to-one pairing with BOLD time points
Model input convention (if used): 16-second EEG windows preceding each fMRI TR (covers HRF peak and variance)

Detailed acquisition and artifact-reduction procedures are documented in the NeuroBOLT paper’s Appendix D.

⛳ Model Checkpoints

Model checkpoints for selected brain regions are available for reproducing our results:
👉 Hugging Face Repository
A step-by-step tutorial can be found here:
📘 Google Colab Notebook

📚 Dataset Citation

If you use our dataset in your research or publications, please cite the following paper:

@inproceedings{
  li2024neurobolt,
  title={Neuro{BOLT}: Resting-state {EEG}-to-f{MRI} Synthesis with Multi-dimensional Feature Mapping},
  author={Yamin Li and Ange Lou and Ziyuan Xu and Shengchao Zhang and Shiyu Wang and Dario J. Englot and Soheil Kolouri and Daniel Moyer and Roza G Bayrak and Catie Chang},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://openreview.net/forum?id=y6qhVtFG77}
}

---
license: Apache-2.0
---

NeuroBOLT EEG-fMRI ROI Dataset

Short name: neurobolt-restModalities: EEG (simultaneous with fMRI), fMRI ROI time seriesSetting: Resting-state (eyes closed)Subjects / Scans: 22 participants, 29 fMRI scans (7 participants with two scans)

🧠 Overview