Research map / Probabilistic and scientific modeling / Dynamical systems
NeuroBOLT: Resting-state EEG-to-fMRI Synthesis with Multi-dimensional Feature Mapping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuroBOLT, a transformer framework that fuses LaBraM-based spatiotemporal EEG features with multi-scale spectral features to regress resting-state fMRI ROI signals from 26-channel EEG.
Problem
Translating EEG to fMRI BOLD signals is limited to few brain regions, task conditions, and subject-specific models; resting-state eyes-closed EEG-to-fMRI synthesis across regions and subjects is largely unexplored.
Approach
Combines a fine-tuned EEG foundation model (LaBraM) for spatiotemporal representation with a multi-scale STFT-style spectral encoder (windows 0.5–4 s) using a linear transformer, summed and fed to a regression head that predicts a parcel's fMRI value from a 16-second EEG window.
Key results
- Intra-subject average R of 0.531 vs 0.484 for the next best (LaBraM), a 12.26% relative improvement over the best EEG encoder.
- Inter-subject (unseen subject, full ~20-min scans) average R of 0.473 vs 0.435 for BIOT, the best baseline.
- Ablation: adding the spectral module raised average R from 0.151 to 0.436 and the four-scale setting reached 0.473.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the multi-scale spectral module to the temporal-spatial module (ablation) | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.151 -> 0.473 | Helped |
| Add the temporal-spatial module to the multi-scale spectral module (ablation) | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.432 -> 0.473 | Helped |
| Increase spectral scales from one scale (l0) to four scales (l3) (ablation) | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.436 -> 0.473 | Helped |
| Increase spectral scales from l1 to l2 (ablation) | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.442 -> 0.461 | Helped |
| Increase spectral scales from l3 to l4 (ablation) | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.473 -> 0.460 | Hurt |
| Pretrain on resting-state data and fine-tune on auditory task data instead of training on task data alone | NeuroBOLT · Auditory task inter-subject | Average Pearson R: 0.373 -> 0.418 | Helped |
| Jointly train on resting-state and auditory task data instead of training on task data alone | NeuroBOLT · Auditory task inter-subject | Average Pearson R: 0.373 -> 0.420 | Helped |
| Jointly train on resting-state and auditory task data instead of resting-state data alone | NeuroBOLT · Resting-state inter-subject | Average Pearson R: 0.473 -> 0.469 | Hurt |
Surprising findings
- Zero-shot transfer from resting-state to auditory-task fMRI performed comparably to training directly on task data (average R 0.380 vs 0.373).
- Joint training on resting-state and task data did not improve resting-state prediction (0.469 vs 0.473).
Limitations
- Model is trained separately for each brain region, which is inefficient for high-resolution whole-brain fMRI reconstruction.
- Small sample size (22 resting-state subjects, 10 task subjects) may introduce bias or overfitting.
- Large performance variability across methods and ROIs; results reported without multiple random seeds.
Experimental setup
Datasets: in-house resting-state EEG-fMRI (29 scans, 22 subjects) in-house auditory task EEG-fMRI (16 scans, 10 subjects) DiFuMo 64-region parcellation
Benchmarks: in-house resting-state EEG-fMRI in-house auditory task EEG-fMRI
Metrics: Pearson correlation R MSE
Baselines: BIOT LaBraM BEIRA Li et al. FFCL CNN Transformer STT Transformer
Builds on: LaBraM BIOT vision transformer DiFuMo
Closest papers by meaning
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.87)
- Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion (NeurIPS 2024 · similarity 0.87)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.86)
- BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals (NeurIPS 2025 · similarity 0.86)
- CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding (NeurIPS 2025 · similarity 0.86)
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.