MIBE Seminar: Biophysically Accurate and Surrogate Models of Spinal Cord Electrical Stimulation

Title: Biophysically Accurate and Surrogate Models of Spinal Cord Electrical Stimulation

Speaker: Katja Frey (Cheng Group)

Date and Time: Tuesday 15 July 2025, 13:00

Abstract: Epidural electrical stimulation (EES) of the lumbosacral spinal cord has been shown to effec-
tively restore functional motor control in individuals with spinal cord injury. Computational
models offer a valuable tool for in-silico testing of electrode configurations and stimula-
tion protocols, thereby reducing reliance for time-consuming and labor-intensive patient-
specific parameter tuning performed in vivo during sessions with the patient. Typically,
these computational models follow a hybrid modeling appraoch that combines finite element
modeling (FEM) to solve the volume conduction problem with a biophysically accurate fiber
model to simulate neural activation. While accurate, these models are computationally
demanding, particularly when used for optimization tasks involving multiple parameter or
geometry variations. In the domain of peripheral intraneural stimulation, machine learning-
based surrogate models have been introduced to significantly reduce computational cost by
approximating the biophysical component. Building upon this concept, the present thesis
extends surrogate modeling to spinal cord stimulation. Specifically, we developed machine
learning-based surrogate models to approximate the output of a hybrid model by predicting
muscle recruitment in key lower-limb muscles based on stimulation parameters applied to
the lumbar spinal cord. The surrogate models were trained on a hybrid model combining
FEM with a static thresholding fiber model and demonstrated high predictive accuracy across
both monopolar and multipolar stimulation configurations. Additionally, the thesis provides
insights into the impact of feature representation as well as the distribution of training
samples on model performance. Preliminary results from a subset of stimulation sites suggest
that similar outcomes may be achievable with biophysically accurate fiber models, although
this has yet to be verified. These findings offer guidance for the future development of
efficient and generalizable surrogate models in spinal cord stimulation. The results suggest
that machine learning-based surrogate modeling is a promising approach for enabling the
personalized design and optimization of EES protocols in a computationally feasible manner.