Related Experiment Video
Updated: Jan 10, 2026
09:57
Author Spotlight: Advancing Pediatric Epilepsy Surgery in Children Through Novel Biomarkers and Enhanced Localization
Published on: September 20, 2024
3.4K
The Seizure Embedding Map: A Spatio-Temporal Transformer for Comparing Patients by Ictal Intracranial EEG Features at
Akash R Pattnaik1,2, Zhongchuan Xu1,2, William K S Ojemann1,2
1Department of Bioengineering, School of Engineering & Applied Sciences, University of Pennsylvania, Philadelphia, PA, 19104, USA.
Medrxiv : the Preprint Server for Health Sciences
|November 24, 2025
Summary
This study introduces a transformer model to analyze intracranial EEG (iEEG) seizure recordings, enabling quantitative comparison of patient data for improved epilepsy treatment planning. The model effectively categorizes seizure networks and their anatomical origins.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Medical Informatics
Background:
- Planning invasive epilepsy treatment relies on subjective interpretation of intracranial EEG (iEEG) seizure recordings.
- Current methods for mapping seizure onset and location are subjective and vary across institutions.
- Comparing new patient seizure data to historical cases is inconsistent due to differing implant strategies and electrode placements.
Purpose of the Study:
- To develop a quantitative method for analyzing iEEG seizure recordings to improve treatment planning for drug-resistant epilepsy.
- To introduce a transformer model that embeds spatial and temporal information from iEEG data.
- To categorize seizure networks and assess their relationship to patient outcomes across a large cohort.
Main Methods:
- Designed and implemented a custom spatiotemporal transformer model to extract features from iEEG seizure onset epochs.
- Utilized convolutional layers for iEEG tokenization and a spatiotemporal positional encoder to capture channel and time relationships.
- Validated seizure embeddings using unsupervised clustering and a cross-validated multi-class logistic regression model.
Main Results:
- Applied the model to 882 seizures from 102 patients, revealing 74 seizure clusters across subjects.
- Achieved a validation accuracy of 0.8159 in clustering seizure onset patterns using logistic regression.
- Found significant associations between seizure clusters and anatomical region of onset/seizure classification, but not therapy or outcome.
Conclusions:
- Developed a method to represent iEEG recordings with informative spatial and temporal embeddings for seizure analysis.
- Demonstrated that these embeddings can reveal common seizure onset patterns and are associated with anatomical origin.
- This work represents a first step toward quantitative, deep learning-based decision-making frameworks for drug-resistant epilepsy management.
