Machine-learning predictor of nerve fiber firing rate allows the automatic optimization of electrical stimulation protocols.

Computational models have been widely employed to study the electrical stimulation of the nervous system. Still, most applications either study fundamental mechanisms underlying stimulation, or address qualitative scientific questions. When quantitative questions are posed, they are mostly evaluated on a small, regular grid of parameter values, thus greatly reducing the wealth of admissible possibilities. The main obstacle to the use of computational models is their very high computational complexity, which prevents testing a large number of parameter values. Here, we show that it is possible to train a regressor that predicts the firing rate of nerve fibers stimulated according to a given multipolar electrical stimulation protocol, and show its possible application to a simplified model of optic nerve stimulation. Our results show that it is possible to build a very accurate surrogate model of nerve fiber stimulation, and that its reduced computational cost allows to perform automatic optimization of multipolar electrical stimulation protocols via evolutionary heuristics.