EEGSurvNet: A Deep Survival Model to Predict Time-to-Seizure After EEG

Model Description

EEGSurvNet predicts the long-term risk of seizure recurrence through time after routine EEG. The input is a 60s EEG segment with 19 channels in spectrogram format (see below for specifications). The output is the logit of the seizure recurrence hazards at 6 discrete time intervals.

Paper: Development and validation of a deep survival model to predict time to seizure from routine electroencephalography (Epilepsia, 2026)

Repository: GitLab Repository - See README for full documentation and preprocessing code

Usage - WIP

Note: See the GitLab repository README for complete preprocessing pipeline, BIDS format requirements, and dataloader examples.

Training Data

• 1,014 routine EEG recordings from 994 patients
• Tertiary care center (CHUM, Montreal)
• Temporal split: Training/validation (Jan 2018 - Sep 2019), Testing (Sep-Dec 2019)
• Median follow-up: 2.2 years (training)

Performance

• 2-year integrated AUROC: 0.69 (95% CI: 0.64–0.73)
• C-index: 0.66 (0.60-0.73)
• Maximal AUROC at 2 months (0.80)
• Outperforms spike-based predictions

Citation

@article{https://doi.org/10.1002/epi.70101,
author = {Lemoine, Émile and Xu, An Qi and Jemel, Mezen and Lesage, Frédéric and Nguyen, Dang K. and Bou Assi, Elie},
title = {Development and validation of a deep survival model to predict time to seizure from routine electroencephalography},
journal = {Epilepsia},
volume = {n/a},
number = {n/a},
pages = {},
keywords = {artificial intelligence, biomarkers, electroencephalography, prognosis, survival analysis},
doi = {https://doi.org/10.1002/epi.70101},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/epi.70101},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1002/epi.70101},
}

Authors

Émile Lemoine, An Qi Xu, Mezen Jemel, Frédéric Lesage, Dang Khoa Nguyen, Elie Bou Assi

Contact

For questions, please contact the corresponding author or submit an issue on the GitLab repository.