What Do EEG Foundation Models Capture from Human Brain Signals?

Ling Tang^1,2*, Qian Chen^1,4*, Jilin Mei^1,3, Houshi Xu², Quanshi Zhang²
Jing Shao¹, Na Zou^1,5, Xia Hu¹, Dongrui Liu^1†
1Shanghai Artificial Intelligence Laboratory   ²Shanghai Jiao Tong University   ³Fudan University
⁴Tongji University   ⁵University of Houston
^*Equal contribution   ^†Corresponding author

     

This is the official repository for BrainPEC, an audit framework for understanding what EEG foundation models learn, use, and explain from human brain signals. BrainPEC evaluates three EEG foundation models, CSBrain, CBraMod, and LaBraM, across five clinical EEG tasks using a six-family, 63-feature EEG lexicon. The pipeline combines layer-wise probing, concept erasure, and transparent classifier closure to separate linear readability from model reliance.

Overview

BrainPEC asks three questions:

1. Probe: Does the model encode classical EEG features?
2. Erase: Does the model use those features for downstream prediction?
3. Closure: How much of the foundation model's performance can be explained by confirmed features?

The audit covers 945 model-task-feature units. In the reported run, 648 units are representation-causal and 199 are encoded-only. Confirmed features recover 79.3% of the foundation-model advantage over a same-dimension random-feature baseline on average.

Main Results

Probing Atlas

Layer-wise ridge probes show where different EEG feature families become linearly readable across model depth, tasks, and architectures.

Feature Taxonomy

Across tasks, confirmed features form broad universal candidates rather than a small task-specific list. Frequency-domain features dominate, but time-domain, time-frequency, complexity, cross-frequency, and cross-channel families all contribute causal mass.

Usage

Setup

Install the required dependencies:

pip install -r requirements.txt

Prepare datasets and fine-tuned checkpoints following the instructions in inference/README.md.

BrainPEC Audit

Run the full factor-atlas pipeline:

scripts/factor_atlas/run_all.sh

Run individual stages:

scripts/factor_atlas/run_features.sh
scripts/factor_atlas/run_probe.sh
scripts/factor_atlas/run_erasure.sh
scripts/factor_atlas/run_aggregate.sh

The scripts call the released package entrypoint:

python -m brainpec.experiments.factor_atlas

Citation

@article{tang2026eeg,
  title={What Do EEG Foundation Models Capture from Human Brain Signals?},
  author={Tang, Ling and Chen, Qian and Mei, Jilin and Xu, Houshi and Zhang, Quanshi and Shao, Jing and Zou, Na and Hu, Xia and Liu, Dongrui},
  journal={arXiv preprint arXiv:2605.11410},
  year={2026}
}

License

This project is licensed under the MIT License.