braindecode/eegdino-medium-pretrained

EEG-DINO Medium — Self-Distillation EEG Foundation Model

EEG-DINO-Medium encoder (~33M parameters) pretrained with DINO-v2 hierarchical self-distillation (Wang et al., MICCAI 2025).

This is the eegdino-medium-pretrained checkpoint for braindecode.models.EEGDINO, curated and re-uploaded as part of the OpenEEG-Bench effort.

Which half of the DINO checkpoint

The released model_EEG_DINO_{S,M}.pt files hold both halves of the self-distillation graph: the student (module.student.*) and its EMA teacher (module.teacher.*). This repository hosts the student encoder. The paper releases the student as the foundation model ("The student model obtained from the pre-training stage will be used as the foundation model to be adapted to downstream tasks", Wang et al., MICCAI 2025), and the authors' run_finetuning.py loads only the module.student. keys. The encoder outputs of this checkpoint match the original code's student forward (max abs difference 0.0 for Small and 1.8e-6 for Medium, float32, CPU). An earlier upload of this repository held the teacher instead. The teacher differs from the student in every tensor: the median relative L2 difference per tensor is 0.13 % (Small) and 0.20 % (Medium), with a maximum below 1 %. Their patch-token outputs differ by 8 % (Small) and 13 % (Medium), measured as relative L2 on random input. convert_eegdino_checkpoints.py at the repository root reproduces the conversion (--source student, the default). The classification head (final_layer.*) is newly initialised; fine-tune it on your task.

Training data

The data that shaped these weights. The ids are the datasets: field of this card, which braindecode reads into model.card_data.datasets when you call from_pretrained (MOABB class names for MOABB datasets, OpenNeuro / NEMAR accessions, Hub ids otherwise). nm/on ids are datasets on NEMAR (on = NEMAR's copy of the OpenNeuro dataset with the same number). Rows marked private: are data the authors did not release; they are not in datasets:.

datasets: id	Dataset	Stage	Modality
TUEG	Temple University Hospital EEG Corpus (TUH EEG)	pretraining	EEG

• same pretraining corpus and recipe as EEG-DINO-Small (paper reports one pretraining dataset for all sizes)
• TUEG (the TUH EEG Corpus) contains the recordings of its sub-corpora TUAB, TUEV, TUSZ, TUAR, TUSL and TUEP; an evaluation on one of those is not held out from these weights unless the authors removed it.
• Sources: Wang et al. MICCAI 2025 (papers.miccai.org/miccai-2025/paper/3347_paper.pdf) p.5, Sec. 3.1; MICCAI 2025 paper 3347 p.5; https://huggingface.co/eegdino/EEG-DINO README; https://huggingface.co/braindecode/eegdino-medium-pretrained/blob/main/convert_eegdino_checkpoints.py.

Quick start

pip install braindecode[hub]

from braindecode.models import EEGDINO

model = EEGDINO.from_pretrained(
    "braindecode/eegdino-medium-pretrained",
    n_outputs=2,            # set to your downstream task
    n_chans=19,
    sfreq=200,
)

from_pretrained reads both the architecture configuration (config.json) and the weights (model.safetensors or pytorch_model.bin) and returns a ready-to-fine-tune nn.Module.

Model details

Architecture	braindecode.models.EEGDINO
Expected channels	19
Expected sampling frequency	200 Hz
Library	braindecode ≥ 1.5
Loaded via	huggingface_hub.PyTorchModelHubMixin (free with braindecode[hub])

For the full architecture description, parameter table, and references, see the rendered docstring at https://braindecode.org/stable/generated/braindecode.models.EEGDINO.html or in the interactive Model Explorer Space.

Training data

Temple University Hospital EEG Corpus (TUEG), 19 common 10-20 channels resampled to 200 Hz (>9000 hours), following CBraMod's preprocessing. Pretrained by hierarchical self-distillation.

Intended use

Larger EEG-DINO encoder for feature extraction or fine-tuning; the architecture is restored from config.json on load. The classification head is re-initialized on load.

Limitations

• Channel layout matters. Performance degrades when the input montage differs from the pretraining montage. Use the Interpolated* variant (where available) or resample channels with MNE before fine-tuning.
• Sampling rate matters. Resample your data to 200 Hz before inference; the positional / patch embeddings assume this rate.
• Inherited license restrictions. Downstream weights derived from this checkpoint inherit the license of the original training corpus (some braindecode pretraining corpora are CC-BY-NC). Verify the upstream dataset licence before commercial use.

Citation

If you use this checkpoint, please cite both the original architecture paper and braindecode.

@inproceedings{wang2025eegdino,
  title     = {{EEG-DINO}: Learning {EEG} Foundation Models via Hierarchical
               Self-Distillation},
  author    = {Wang, Xujia and Liu, Xuhui and Liu, Xi and Si, Qian and Xu,
               Zhaoliang and Li, Yang and Zhen, Xiantong},
  booktitle = {Medical Image Computing and Computer Assisted Intervention (MICCAI)},
  year      = {2025},
}

@article{aristimunha2025braindecode,
  title   = {Braindecode: a deep learning library for raw electrophysiological data},
  author  = {Aristimunha, Bruno and others},
  journal = {Zenodo},
  year    = {2025},
  doi     = {10.5281/zenodo.17699192},
}

License

BSD-3-Clause for the model code (matching braindecode). The pretraining data may impose additional restrictions — see Limitations.