Combining biophysical models and machine learning to optimize implant geometry and stimulation protocol for intraneural electrodes

Abstract

Objective: Peripheral nerve interfaces have the potential to restore sensory, motor, and visceral functions. In particular, intraneural interfaces allow targeting deep neural structures with high selectivity, even if their performance strongly depends upon the implantation procedure and the subject's anatomy. Currently, few alternatives exist for the determination of the target subject structural and functional anatomy, and statistical characterizations from cadaveric samples are limited because of their high cost. We propose an optimization workflow that can guide both the pre-surgical planning and the determination of maximally selective multisite stimulation protocols for implants consisting of several intraneural electrodes, and we characterize its performance in silico. We show that the availability of structural and functional information leads to very high performances and allows taking informed decisions on neuroprosthetic design. Approach: We employ hybrid models (HMs) of neuromodulation in conjunction with a machine learning-based surrogate model to determine fiber activation under electrical stimulation, and two steps of optimization through particle swarm optimization (PSO) to optimize in silico implant geometry, implantation and stimulation protocols using morphological data from the human median nerve at a reduced computational cost. Main results: Our method allows establishing the optimal geometry of multi-electrode transverse intra-fascicular multichannel electrode (TIME) implants, the optimal number of electrodes to implant, their optimal insertion, and a set of multipolar stimulation protocols that lead in silico to selective activation of all the muscles innervated by the human median nerve. Significance: We show how to use effectively HMs for optimizing personalized neuroprostheses for motor function restoration. We provide in-silico evidences about the potential of multipolar stimulation to increase greatly selectivity. We also show that the knowledge of structural and functional anatomies of the target subject leads to very high selectivity and motivate the development of methods for their in vivo characterization.

A computational model to design wide field-of-view optic nerve neuroprostheses

“…Here, we have manually set the stimulation parameters to obtain visual perceptions similar to the target scene. This process should be automatized using optimization techniques like genetic algorithms or particle swarm optimization, like shown in [46], [49]. To reduce the computational cost of evaluating the effect of the applied stimulation on the target nerve fibers, surrogate models capturing the behavior of biophysical models without including all their computational details should be used.…”

“…While fiber activation can be predicted at a low computational cost using model-based surrogate models like the activating function formalism [50], [51], these methods are based on linear cable theory and do not generalize to the prediction of firing rates, which depend upon inherently nonlinear phenomena. Machine learning-based surrogate models like the multilayer perceptron approach shown in [49] should be explored to this aim.…”