Fine-tuning Zyphra's ZUNA1.1 EEG foundation model on the Temple University Hospital EEG Corpus — the largest public clinical EEG dataset. AGPL-3.0 code, CC-BY-4.0 data. No closed-source profiteering from this research.
Clinical EEG is underserved by AI. Most existing models are proprietary, gated behind vendor licenses, or trained only on healthy subjects rather than real clinical seizure and non-seizure recordings. ZUNA1.1 is a 380M-parameter diffusion autoencoder that can reconstruct, denoise, and upsample EEG from any montage — but it was trained on public research datasets, not clinical-grade data. This project fine-tunes it on real hospital recordings and publishes the result freely, so researchers and clinics don't have to buy access to a black box.
| Capability | Clinical relevance |
|---|---|
| Denoise channels | Removes muscle/eye-blink artifacts without losing signal |
| Reconstruct missing channels | Fixes dead electrodes, extends recordings from sparse montages |
| Upsample montages | Predicts 128-channel EEG from a standard 19-channel clinical recording |
| Channel-agnostic (3D scalp coordinates + time) | Works on any electrode layout via 4D rotary position encoding — no retraining per montage |
| Variable-length input (0.5–30s) | Handles short trial snippets and long continuous recordings with no reconfiguration |
The following evaluations are planned, not yet completed: seizure detection (ictal vs. interictal vs. normal), artifact robustness under clinical noise, montage generalization (10-20, neonatal, intraoperative ECoG), pathology-aware denoising, and cross-dataset transfer beyond TUH.
The TUH EEG Corpus requires a data access request directly from Temple University. Once access is granted: place EDF files under data/tuh_raw/, run scripts/preprocess_tuh.py, and the output lands in data/tuh_processed/ as ZUNA-compatible pairs, ready for fine-tuning.