Source-linked AI summary
Dynamic network drivers of seizure generation, propagation and termination in human epilepsy
Ankit Khambhati, Brian Litt, Danielle S. Bassett
TL;DR
Drug-resistant epilepsy requires understanding how seizure foci interact dynamically with surrounding epileptic networks. Using intracranial ECoG recordings and a network-reconfiguration method, the study identifies distinct seizure states and finds that onset-zone isolation diminishes at initiation before strong focal subnetworks re-emerge during termination.
Problem
Drug-resistant epilepsy is often treated by targeting seizure-generating tissue, but the dynamic connectivity linking seizure foci with surrounding epileptic networks remains difficult to characterize.
Method
The study analyzes pre-seizure and seizure ECoG recordings by constructing weighted functional networks and clustering time windows with similar connectivity configurations.
Results
Seizures most often progressed through three network configurations: start, middle, and end, with weak connectivity during initiation and propagation and strong connectivity during termination.
Takeaways & Limitations
Seizure generation coincided with breakdown of onset-zone isolation, while later re-isolation and focal strong connections may support seizure termination and guide intervention targets.
Takeaways & Limitations
Intracranial electrode implantation incompletely samples the epileptic network, so onset zones or seizure-spread regions may not be fully represented.
Abstract
from arXiv · showhide
Drug-resistant epilepsy is traditionally characterized by pathologic cortical tissue comprised of seizure-initiating `foci'. These `foci' are thought to be embedded within an epileptic network whose functional architecture dynamically reorganizes during seizures through synchronous and asynchronous neurophysiologic processes. Critical to understanding these dynamics is identifying the synchronous connections that link foci to surrounding tissue and investigating how these connections facilitate seizure generation and termination. We use intracranial recordings from neocortical epilepsy patients undergoing pre-surgical evaluation to analyze functional connectivity before and during seizures. We develop and apply a novel technique to track network reconfiguration in time and to parse these reconfiguration dynamics into distinct seizure states, each characterized by unique patterns of network connections that differ in their strength and topography. Our approach suggests that seizures are generated when the synchronous relationships that isolate seizure `foci' from the surrounding epileptic network are broken down. As seizures progress, foci reappear as isolated subnetworks, marking a shift in network state that may aid seizure termination. Collectively, our observations have important theoretical implications for understanding the spatial involvement of distributed cortical structures in the dynamics of seizure generation, propagation and termination, and have practical significance in determining which circuits to modulate with implantable devices.
Significance Statement.
The study identifies three seizure-related network states with distinct temporal, strength-based, and spatial connectivity patterns. These reconfigurations suggest that seizure initiation involves loss of onset-zone isolation, while later strong, focal connectivity accompanies termination.
- Epileptic Network Reconfiguration: Start and end communities were less similar to each other than either was to the middle community, supporting an intermediary role for the middle state.
- Network Geometry: Mean connectivity increased across seizure states, indicating progressively greater gross network synchronization.
- Network Geometry: Seizure initiation was dominated by weak connections, whereas termination was dominated by strong connections; the middle state varied substantially across seizures.Strong and weak connections were defined as the highest and lowest 10% of connection weights within each epoch.
- Network Tightening During Seizures: Anatomical connectivity patterns changed across seizure states, with strong connections becoming shorter and more focal at seizure onset before tightening into a focal subnetwork during seizures.The analysis found significant effects of both connection location and community on the balance of strong and weak connections.
- Epileptic Network Reconfiguration: Community detection revealed three stereotyped network configurations corresponding to seizure start, middle, and end states.The middle state appears transitional rather than a simple propagation phase.
- Epileptic Network Reconfiguration: The weighted-network method distinguishes strong synchronous from weak asynchronous connections, enabling analysis of network geometry rather than connectivity alone.
- Functional Incorporation of Seizure Onset Sub-Network: The onset zone loses isolation at seizure initiation, while strong within-zone connections reappear later as surrounding epileptogenic cortex is recruited.This pattern suggests a network-level shift from seizure initiation toward termination.
- Limitations: Incomplete intracranial sampling can leave seizure onset zones or spread regions insufficiently represented, motivating validation in larger and more varied patient cohorts.
1 Patient Demographics
The patient-demographics table identifies the study cohort, seizure classifications, and imaging status for implanted patients.
- The table records patient identifiers, sex, age at onset and surgery, etiology, seizure type, and imaging status.
- I002_P023 was female, aged 18 at onset and 25 at surgery, with left temporal lobe CPS+Gen seizures and non-lesional imaging.
- Seizure types were classified as complex partial or complex partial with secondary generalization.
2 The Configuration Vector
The configuration vector represents each dynamic network state through its node-to-node connections, enabling quantitative comparison of network connectivity across time.
- Dynamic networks lack established approaches for quantifying how gross connection patterns vary over time.
- For N fixed nodes across T windows, each time window is represented by the network’s N(N−1)/2 possible connections.
- Changes in configuration vector v_t between time windows indicate dynamical changes in network connectivity.
- Pearson correlation between configuration vectors produces a symmetric T × T similarity matrix for comparing connectivity patterns across windows.
3 Modularity Optimization for Community Detection
The method applies modularity optimization to configuration-similarity networks, using a resolution parameter and null model to identify communities at appropriate scales.
- Community detection clusters densely interconnected nodes into modules representing meaningful group structure in complex networks.
- The modularity formulation assigns nodes to communities, uses γ as a structural resolution parameter, and defines P_ij as expected edge weight under a null model.
- The Newman-Girvan null model is a commonly used choice for expected connection weights in modularity optimization.
- Modularity Q compares observed within-community connection weight with a null model and partitions the network by maximizing this relative within-community weight.
- Varying γ examines network structure across topological or geometric scales, including finer-grained smaller communities.
- Choosing γ involves a trade-off between grouping all nodes together with high modularity and separating nodes into individual communities with low modularity.
4 Network Configuration Communities
Configuration communities group similar network states over time and reveal how seizure epochs progress through temporally organized connectivity patterns.
- Flexibility F measures temporal reconfiguration as the fraction of possible community-assignment changes occurring within an epoch.
- Configuration communities are re-assigned across seizures by ranking them according to the fraction of epoch time windows assigned to each community.