Related Experiment Video
Updated: Jun 13, 2026
09:32
Network Analysis of Foramen Ovale Electrode Recordings in Drug-resistant Temporal Lobe Epilepsy Patients
Published on: December 18, 2016
Towards Interpretable Seizure Detection: An Excitation/Inhibition Dynamic Polynomial Network Framework for
Xihan Sun1, Ying Yan2,3, Na Liu4
1Reading Academy, Nanjing University of Information Science and Technology, Nanjing 210044, China.
Sensors (Basel, Switzerland)
|June 12, 2026
Summary
A new deep learning model, the Excitation/Inhibition Dynamic Polynomial Network (E/I-DynPolyNet), offers interpretable seizure detection from EEG signals. It accurately identifies seizures and quantifies excitation/inhibition imbalance, aligning with clinical observations of epilepsy.
Area of Science:
- Neuroscience
- Computational Neuroscience
- Machine Learning
Background:
- Epilepsy is a neurological disorder causing recurrent seizures, detectable via electroencephalogram (EEG).
- Current deep learning models for EEG seizure detection lack physiological interpretability.
- Understanding the brain's excitation/inhibition (E/I) balance is crucial for epilepsy research.
Purpose of the Study:
- To develop a biologically grounded deep learning framework for interpretable seizure detection using EEG signals.
- To create a model that reflects latent E/I representations and neurophysiological principles.
- To bridge the gap between data-driven seizure detection and mechanistic understanding of ictogenesis.
Main Methods:
- Proposed the Excitation/Inhibition Dynamic Polynomial Network (E/I-DynPolyNet) with dual E/I pathways and sign-constrained weights.
- Embedded a differentiable Wilson-Cowan module to model E/I temporal dynamics.
- Employed a physics-informed optimization strategy integrating supervised learning with dynamical residual constraints and E/I balance regularization.
Main Results:
- Achieved high detection accuracies of 95.81% (CHB-MIT) and 98.5% (Bonn) datasets.
- E/I-DynPolyNet quantitatively estimated E/I imbalance, showing an increase from 1.01 (pre-ictal) to 1.38 (ictal).
- Demonstrated findings consistent with clinical observations of ictogenesis and E/I imbalance during seizures.
Conclusions:
- E/I-DynPolyNet provides accurate and interpretable seizure detection from EEG data.
- The model offers a mechanistic insight into seizure dynamics by quantifying E/I imbalance.
- This framework enhances understanding of epilepsy pathophysiology and improves diagnostic capabilities.
