Official PyTorch implementation of the paper:
Are EEG Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse BCI Tasks (ICLR 2026)
- [Jan 2026] Our paper is accepted to ICLR 2026!
- [Dec 2025] 🥇 We won 1st Place in Challenge 1 of the NeurIPS 2025 EEG Foundation Challenge!
- [Jan 2025] The original ST-EEGFormer paper was rejected from ICLR 2025
If you use our model or find it useful, please cite the following paper:
@inproceedings{
yang2026_steegformer,
title={Are {EEG} Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse {BCI} Tasks},
author={Liuyin Yang and Qiang Sun and Ang Li and Marc M. Van Hulle},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=5Xwm8e6vbh}
}Our framework provides a transparent and rigorous benchmark for EEG foundation models by evaluating them across 6 distinct decoding protocols, ranging from simple population decoding to challenging zero-shot and transfer learning scenarios.
As a baseline for this benchmark, we introduce ST-EEGFormer: a minimal, ViT-based foundation model. To ensure transparency and ease of reproduction, the model is pre-trained purely through Masked Autoencoder (MAE) reconstruction on raw EEG signals.
Our comprehensive evaluation reveals that while classic neural network decoders remain highly competitive, EEG foundation models often struggle when restricted to linear probing. However, when fully fine-tuned, ST-EEGFormer-large—achieves the best average rank (5.61) among all compared models, despite its large parameter count (>300M).
This project is licensed under the MIT License - see the LICENSE file for details.
Note: The MIT license applies to the source code provided in this repository. The associated research paper, architectural diagrams, and the name "ST-EEGFormer" are © 2026 Computational Neuroscience Group, KU Leuven. All rights reserved.
The models are implemented in PyTorch and can be used in standard Python environments.
Python version used for pre-training:
3.11.5
| Category | Package | Version | Note |
|---|---|---|---|
| Core | torch |
2.4.1 | Deep learning framework |
| Core | timm |
1.0.10 | Transformer model implementations |
| Extra | wandb |
0.22.2 | Experiment logging & monitoring |
| Extra | mat73 |
0.65 | Loading MATLAB v7.3 files |
| Extra | scikit-learn |
1.3.2 | Evaluation metrics and utilities |
If you want to run the training code for classic EEG models, you will also need:
| Package | Version | Note |
|---|---|---|
scipy |
1.16.0 | General scientific computing utilities |
numpy |
1.25.2 | Core numerical computing library |
mne |
1.9.0 | EEG preprocessing and data handling |
pyriemann |
0.6 | Riemannian geometry-based EEG classification |
scikit-learn |
1.4.2 | Machine learning toolkit |
lightgbm |
4.6.0 | Gradient boosting models for tabular features |
| Package | Version | Note |
|---|---|---|
scipy |
1.15.3 | General scientific computing utilities |
numpy |
2.2.6 | Core numerical computing library |
mne |
1.9.0 | EEG preprocessing and data handling |
scikit-learn |
1.7.0 | Machine learning toolkit |
meegkit |
0.1.9 | EEG/MEG signal processing utilities |
ST-EEGFormer is designed for 128 Hz EEG data.
- Pre-trained to reconstruct 6-second EEG segments
- Supports up to 142 EEG channels
- Recommended input: ≤ 6-second segments, sampled at 128 Hz
The list of available/pretrained channels can be found in:
pretrain/senloc_file
A Jupyter notebook containing a minimal tutorial on how to use the model can be found in:
easy_start/simple_example.ipynb
If you want to pre-train a model, use the script:
pretrain/ddp_train_eeg.py
You will need to prepare your own custom dataset that provides EEG segments and the corresponding channel indices.
If you want to run benchmark experiments on downstream BCI tasks using neural networks, use:
benchmark/neural_networks/wandb_downstream_evaluation.py
For dataset preparation and configuration details, please refer to the README file in:
benchmark/neural_networks
For the EEG 2025 Foundation Challenge, the code is located in:
eeg_foundation_2025
where the models are slightly modified (they include additional channel embeddings for the HBN dataset).
We release small, base, and large ST-EEGFormer models in the GitHub releases.
Additionally, we provide large-ST-EEGFormerV2, which has undergone further pre-training on the HBN datasets for the EEG 2025 Foundation Challenge.
ST-EEGFormer-large release-HBN.
We are now working on the following updates:
- Dataset Preprocessing Codes: Standardized scripts for cleaning and formatting benchmarked datasets.
- Step-by-Step Tutorials: More Jupyter notebooks demonstrating how to use the model.