Abstract
Deep brain stimulation (DBS) is a critical surgical intervention for Parkinson's Disease. While traditional modeling often relies on quasi-static approximations (e.g., Laplace or Poisson equations), these simplifications may fail to capture the nuances of rapid, high-frequency stimulation pulses. This paper proposes a novel hybrid approach: a Three Spatial Dimension Wave Latent Force Model (LFM). By integrating the non-homogeneous wave equation with Gaussian processes, we create a probabilistic framework capable of describing the dynamic behavior of DBS and solving inverse problems to recover excitation sources from observed potential distributions. Our results demonstrate high fidelity to traditional Finite Element Method (FEM) solutions while providing a more realistic description of time-varying electromagnetic fields in brain tissue.
Introduction
The therapeutic efficacy of Deep Brain Stimulation (DBS) depends heavily on the precise adjustment of stimulus parameters, including frequency, pulse width, and electrode configuration. Standard engineering models typically assume that the electric field propagates instantaneously, ignoring time derivatives. However, the high-frequency power contained in DBS pulses suggests that a dynamic approach is necessary to fully understand the interaction between the stimulus and neural tissue.
We address this by developing a Wave Latent Force Model (LFM) that moves beyond steady-state approximations. This model is physically grounded in the wave equation yet flexible enough to handle the uncertainties inherent in biological systems through a probabilistic Gaussian Process (GP) framework.
Methodology
The Wave Latent Force Model (LFM)
The core of our approach is the non-homogeneous wave equation in three spatial dimensions. Unlike the static Poisson equation, the wave equation accounts for propagation velocity and the time-varying nature of the excitation source. We treat the excitation source—the "latent force"—as a Gaussian Process. Because the wave equation is a linear operator, the resulting electric potential also follows a GP distribution.
To bridge the gap between the differential equation and the statistical model, we derive specific covariance and cross-covariance functions using Green’s functions. This ensures the physics of wave propagation is "baked" into the kernel of the Gaussian Process.
Electrode Configurations and Domain
Our simulations utilized a $10\text{cm} \times 10\text{cm} \times 10\text{cm}$ cubic domain of brain tissue. We modeled various clinical electrode configurations, including monopolar and bipolar setups, to test the forward and inverse capabilities of the LFM.
Computational Efficiency
Gaussian Processes in 3D typically suffer from $O(N^3)$ complexity. To mitigate this, we employed Tensor Product Kernels (TPK), which allow for significant computational savings when data is structured on a grid. We found that a Green's function expansion of approximately seven terms provided an optimal balance between computational speed and approximation accuracy.
Results and Discussion
Forward Modeling and Fidelity to FEM
In static benchmarks, the electric potentials predicted by the Wave LFM were compared against solutions obtained via the Finite Element Method (FEM) using the FEniCS Project toolbox. The LFM's predictions were nearly identical in shape and magnitude to the traditional FEM results, validating the model's accuracy in the steady-state limit while maintaining its capacity for dynamic simulation.
Solving the Inverse Problem
A significant advantage of the LFM is its ability to solve the inverse problem—recovering the spatial configuration of the excitation source given an observed electric potential. In 2D experiments, the model successfully recovered four distinct source configurations with low Mean Squared Error (MSE), ranging from $1.2 \times 10^{-3}$ to $2.5 \times 10^{-3}$. This capability suggests that the LFM could eventually serve as a clinical tool for tuning electrode parameters based on desired field distributions.
Dynamic Capabilities
Unlike Fourier-based FEM methods that focus on steady-state solutions, the LFM can model transients—the complex changes occurring at the onset and offset of a stimulation pulse. We successfully visualized how time-varying sources, such as sine-modulated pulses, propagate through the 3D domain over time.
Conclusion and Future Work
The 3D Wave Latent Force Model provides a robust, physics-informed framework for modeling DBS. It successfully bridges the gap between mechanistic differential equations and flexible probabilistic modeling. While the current model assumes a homogeneous and isotropic medium, future iterations will incorporate non-stationary LFMs to account for the heterogeneity and anisotropy of real brain tissue (e.g., grey vs. white matter). This will allow for even more realistic simulations of complex neuroanatomy and improved clinical outcomes.
Summary updated