EpiNT — Epilepsy Neurophysiological Transformer

Pretrained weights for braindecode.models.EpiNT, a braindecode port of EpiNT (Zhang et al., 2025), a channel-independent transformer pretrained on 2,700 hours of scalp EEG and intracranial EEG (ECoG, sEEG) from 1,199 patients.

• Paper: Cross-Modal Epileptic Signal Harmonization: Frequency Domain Mapping Quantization for Pre-training a Unified Neurophysiological Transformer (arXiv:2506.17068)
• Original code & weights: RunKZhang/EpiNT
• braindecode docs: https://braindecode.org

Provenance & license

These weights are the authors' pretrained backbone, weights/representations.bin at commit 781e1d61 (sha256 b4449974058fd6d314debeef0a09c05730a70e822f17244e88d1ee4e7d9a8b1d), converted into the braindecode EpiNT module through its mapping (all 102 backbone tensors; the pretraining mask token embed.mask_encoding is dropped). The classification head is not part of the release and is randomly initialized.

The original release is under the MIT license, which this repository preserves:

MIT License

Copyright (c) 2024 Auton Lab, Carnegie Mellon University

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Model configuration

This checkpoint uses the paper's configuration (~19.2M parameters):

Each encoder LayerNorm spans tokens and features, so keep n_times=3072: 12 s at 256 Hz for scalp EEG, or 3 s at 1024 Hz for iEEG. Signals are expected in microvolts. Every channel is encoded alone and the class tokens are averaged, so n_chans is free; n_outputs resets the task head.

Usage

from braindecode.models import EpiNT

model = EpiNT.from_pretrained("braindecode/epint-pretrained", n_chans=17, n_outputs=2)

Citation

@article{zhang2025epint,
  title   = {Cross-Modal Epileptic Signal Harmonization: Frequency Domain Mapping Quantization for Pre-training a Unified Neurophysiological Transformer},
  author  = {Zhang, Runkai and Yu, Hua and Gan, John Q. and Wang, Haixian},
  journal = {arXiv preprint arXiv:2506.17068},
  year    = {2025}
}