Localising the Seizure Onset Zone from Routine Interictal and Ictal Scalp EEG Via a Factorised Spatiotemporal Ensemble Deep Learning Framework
- 1University College London
Abstract
Background: Drug-resistant focal epilepsy requires precise localisation of the epileptogenic zone (EZ) for surgical resection. Current practice depends on invasive recordings or spontaneous seizure capture, a costly and risky bottleneck. Interictal scalp EEG can exhibit localising features like spikes, but scalp EEG suffers from volume conduction (the smearing of brain signals across multiple electrodes) and low signal-to-noise ratio (SNR), which obscure subtle spatial dynamics. Neurophysiologist review of long, often multi-day recordings is time-consuming and expertise-dependent, and while computational methods exist, they tend to rely heavily on human-designed features. Deep learning for EEG localisation could therefore help to improve clinical efficiency and sensitivity. Objectives: To develop and validate a non-invasive deep learning approach for lobar-level localisation of the seizure onset zone (SOZ), the primary electrophysiological proxy for the epileptogenic zone (EZ), using interictal and ictal segments from routine scalp EEG. The model is trained on expert-annotated clinical labels and is intended to augment, not supplant, clinical judgement in EZ delineation. Methods: We curated 262 sessions (6,000+ hours) from 100 patients with unilateral focal epilepsy, tested with an extra 28-patient internal and a 9-patient external dataset. The framework integrates: (1) an online augmentation pipeline (Gaussian noise, amplitude scaling, temporal cropping, and hemispheric channel swap) to mitigate overfitting. (2) a multi-scale temporal block for 20-200 ms discharge morphologies; (3) a factorised spatiotemporal Transformer decoupling temporal and spatial attention to counter volume conduction; and (4) a hemispheric mutual exclusivity penalty encoding unilateral priors. A five-model heterogeneous ensemble, fused via model soup, was evaluated on held-out, seizure-free, and cross-institutional cohorts. Results: The ensemble achieved 88.46% session-level exact match on internal test data, outperforming EEGNet (50.00%). Performance generalised to the seizure-free holdout set (82.76%) and to cross-institutional data (84.62%). Rare occipital class (RO) recall rose from 0.500 to 1.000, with zero false positives on healthy controls (n=108). On unlabelled random segments, accuracy was 66.67% (vs. 14% chance). A full 24-hour recording was processed in 4.26 minutes, with inference under 8 seconds, highlighting localising windows for targeted expert clinical review Conclusion: Clinically actionable SOZ spatial information is encoded in routine interictal scalp EEG. Our factorised ensemble provides a safe, interpretable, and efficient tool to expedite presurgical evaluation, potentially reducing reliance on prolonged monitoring and invasive implantation.
Related
- SeverYX: An ensemble framework for robust variant severity classification
- An Evidence-Tiered Developmental Framework for Congenital Vertebral Malformations: Spatiotemporal Thresholds in Genotype–Phenotype Interpretation
- An ensemble framework for robust drug response prediction and large-scale sensitivity profiling across cancers
- HBOCpred: An Ensemble Framework for Prioritisation of Germline Missense Variants in Hereditary Breast and Ovarian Cancer
- An operational spatiotemporal prediction framework for sector congestion in air traffic management
BibTeX
@misc{shi2026localising,
title = {{Localising the Seizure Onset Zone from Routine Interictal and Ictal Scalp EEG Via a Factorised Spatiotemporal Ensemble Deep Learning Framework}},
author = {S. Shi},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.16.759501},
url = {https://www.biorxiv.org/content/10.64898/2026.10.16.759501v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Localising the Seizure Onset Zone from Routine Interictal and Ictal Scalp EEG Via a Factorised Spatiotemporal Ensemble Deep Learning Framework
AU - S. Shi
PY - 2026
DA - 2026/10/16
PB - bioRxiv
DO - 10.64898/2026.10.16.759501
UR - https://www.biorxiv.org/content/10.64898/2026.10.16.759501v1
AB - Background: Drug-resistant focal epilepsy requires precise localisation of the epileptogenic zone (EZ) for surgical resection. Current practice depends on invasive recordings or spontaneous seizure capture, a costly and risky bottleneck. Interictal scalp EEG can exhibit localising features like spikes, but scalp EEG suffers from volume conduction (the smearing of brain signals across multiple electrodes) and low signal-to-noise ratio (SNR), which obscure subtle spatial dynamics. Neurophysiologist review of long, often multi-day recordings is time-consuming and expertise-dependent, and while computational methods exist, they tend to rely heavily on human-designed features. Deep learning for EEG localisation could therefore help to improve clinical efficiency and sensitivity. Objectives: To develop and validate a non-invasive deep learning approach for lobar-level localisation of the seizure onset zone (SOZ), the primary electrophysiological proxy for the epileptogenic zone (EZ), using interictal and ictal segments from routine scalp EEG. The model is trained on expert-annotated clinical labels and is intended to augment, not supplant, clinical judgement in EZ delineation. Methods: We curated 262 sessions (6,000+ hours) from 100 patients with unilateral focal epilepsy, tested with an extra 28-patient internal and a 9-patient external dataset. The framework integrates: (1) an online augmentation pipeline (Gaussian noise, amplitude scaling, temporal cropping, and hemispheric channel swap) to mitigate overfitting. (2) a multi-scale temporal block for 20-200 ms discharge morphologies; (3) a factorised spatiotemporal Transformer decoupling temporal and spatial attention to counter volume conduction; and (4) a hemispheric mutual exclusivity penalty encoding unilateral priors. A five-model heterogeneous ensemble, fused via model soup, was evaluated on held-out, seizure-free, and cross-institutional cohorts. Results: The ensemble achieved 88.46% session-level exact match on internal test data, outperforming EEGNet (50.00%). Performance generalised to the seizure-free holdout set (82.76%) and to cross-institutional data (84.62%). Rare occipital class (RO) recall rose from 0.500 to 1.000, with zero false positives on healthy controls (n=108). On unlabelled random segments, accuracy was 66.67% (vs. 14% chance). A full 24-hour recording was processed in 4.26 minutes, with inference under 8 seconds, highlighting localising windows for targeted expert clinical review Conclusion: Clinically actionable SOZ spatial information is encoded in routine interictal scalp EEG. Our factorised ensemble provides a safe, interpretable, and efficient tool to expedite presurgical evaluation, potentially reducing reliance on prolonged monitoring and invasive implantation.
ER -