Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models

John F. Ingham, Frances Hutchings, Paolo Zuliani, Yujiang Wang, Sadegh Soudjani, Peter N. TaylorarXiv 2024 · DOI · arXiv

TL;DR

Researchers have developed the first large-scale computational model of Vagus Nerve Stimulation (VNS) for epilepsy. By extending a thalamocortical neural mass model with nine additional brain regions and structural connectivity data from the Human Connectome Project, they created an in silico platform that successfully replicates the seizure-reducing effects seen in clinical practice.

What Problem Does This Paper Address?

Vagus Nerve Stimulation (VNS) is a common "palliative" treatment for drug-resistant epilepsy, yet it remains something of a black box. While it reduces seizure frequency by about 50% in half of patients, clinicians still don't fully understand its mechanism of action. Consequently, choosing the right candidates for the procedure and finding the optimal stimulation parameters (current, frequency, pulse width) is currently a process of trial and error that can take months or years.

While computational modeling has been used to predict outcomes for epilepsy surgery and optogenetic stimulation, no such model existed for VNS—until now.

Methods

The authors built their model using a Neural Mass Model (NMM) approach. Rather than simulating billions of individual neurons, NMMs track the aggregate activity of neuronal populations, making it computationally feasible to simulate large-scale brain networks.

  1. Base Model: They started with an established four-population model representing a thalamocortical circuit (the "engine" of many seizures).
  2. Network Expansion: Based on a literature review, they added 18 further populations representing nine additional brain regions implicated in VNS, including the Nucleus Tractus Solitarius (NTS), Locus Coeruleus (LoC), and the Amygdala.
  3. Structural Connectivity: To determine how these regions talk to each other, they utilized diffusion tensor imaging (DTI) data to weight the connections.
  4. Simulation: They used bifurcation analysis to study the deterministic version of the model (finding the mathematical "tipping points" of seizures) and the Euler-Maruyama method for stochastic versions that include random noise to trigger spontaneous seizure-like events.
  5. VNS Modeling: VNS was simulated as a 30 Hz square wave added directly to the input of the NTS, mimicking how the physical device stimulates the vagus nerve in vivo.

Datasets

  • Human Connectome Project (HCP842): The researchers used this tractography atlas overlaid with fractional anisotropy (FA) maps to determine the relative connection strengths between brain regions.
  • Desikan-Killiany Parcellation: This was used to define the grey matter regions of interest for the connectivity mapping.

Key Insights and Ideas

  • The NTS as a Control Parameter: The model reveals that the Nucleus Tractus Solitarius (NTS) acts as a gateway. At low levels of NTS input, the brain is "bistable," meaning it can easily flip between normal activity and a seizure state. Increasing NTS input (via VNS) pushes the system into a "monostable" state where only normal activity is mathematically stable.
  • Dose-Dependent Seizure Abatement: The model showed that increasing the amplitude of the VNS signal reduced the total time spent in a seizure state in a dose-dependent manner.
  • Replicating Clinical "Magnet Mode": The model successfully captured the behavior of "Magnet Mode"—where a patient triggers a burst of stimulation to stop an ongoing seizure. In the simulation, VNS caused seizure events to terminate much earlier than they would have naturally.

Key Results

  • Seizure Reduction: In stochastic simulations representing 27.8 hours of time, the model produced seizures with a mean duration of 61.3 seconds. Applying VNS significantly reduced the proportion of time spent in these seizure states.
  • Stability Transition: The deterministic analysis showed that at NTS input values above -0.52, the system remained monostable (seizure-free), providing a clear target for stimulation efficacy.
  • Validation: The simulated local field potentials (LFP) during VNS-induced seizure termination qualitatively matched real-world clinical EEG recordings from epilepsy patients.

Limitations and Future Work

  • Timescales: The current model only captures the immediate, short-term effects of VNS (seconds to minutes). It does not account for long-term changes like neuroplasticity or anti-inflammatory effects, which likely contribute to why VNS efficacy often improves over months of treatment.
  • Regional Simplification: To keep the model manageable, all 10 brain regions were modeled using the same internal parameters as the somatosensory cortex. Future iterations could refine these parameters to reflect the unique architectures of subcortical structures.
  • Personalized Medicine: The authors envision using this framework with individual patient imaging data. By plugging a specific patient's brain connectivity into the model, clinicians might one day predict whether that patient will respond to VNS before the device is even implanted.

Summary updated

Abstract

Vagus Nerve Stimulation (VNS) is an established palliative treatment for drug resistant epilepsy. While effective for many patients, its mechanism of action is incompletely understood. Predicting individuals' response, or optimum stimulation parameters, is challenging. Computational modelling has informed other problems in epilepsy but, to our knowledge, has not been applied to VNS. We started with an established, four-population neural mass model (NMM), capable of reproducing the seizure-like dynamics of a thalamocortical circuit. We extended this to include 18 further neural populations, representing nine other brain regions relevant to VNS, with connectivity based on existing literature. We modelled stimulated afferent vagal fibres as projecting to the nucleus tractus solitarius (NTS), which receives input from the vagus nerve in vivo. Bifurcation analysis of a deterministic version of the model showed higher background NTS input made the model monostable at a fixed point (FP), representing normal activity, while lower inputs produce bistability between the FP and a limit cycle (LC), representing the seizure state. Adding noise produced transitions between seizure and normal states. This stochastic model spent decreasing time in the seizure state with increasing background NTS input, until seizures were abolished, consistent with the deterministic model. Simulated VNS stimulation, modelled as a 30 Hz square wave, was summed with the background input to the NTS and was found to reduce total seizure duration in a dose-dependent manner, similar to expectations in vivo. We have successfully produced an in silico model of VNS in epilepsy, capturing behaviour seen in vivo. This may aid understanding therapeutic mechanisms of VNS in epilepsy and provides a starting point to (i) determine which patients might respond best to VNS, and (ii) optimise individuals' treatments.
BibTeX
@misc{ingham2024vagus,
  author = {John F. Ingham and Frances Hutchings and Paolo Zuliani and Yujiang Wang and Sadegh Soudjani and Peter N. Taylor},
  title = {{Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models}},
  year = {2024},
  eprint = {2406.02729},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.02729},
}
Figure