Rapid prediction of peripheral axonal activation under multipolar stimulation using artificial intelligence methods

Objective: Computational models of nerve-stimulating electrodes enable the study of far more stimulation parameter sets than are feasible in situ, but analyzing nerve responses to thousands of stimulation sets requires rapid calculation of axonal responses to external voltage fields. Current methods involve a tradeoff between accuracy and computational efficiency: existing rapid predictors rely on activating-function thresholds across nodes of Ranvier, which cannot fully address nonlinear axonal dynamics such as anodal block or anodic break. Additionally, present rapid estimators do not accurately predict axonal responses to multipolar stimulation in which many voltage fields are applied to the same axon simultaneously. We aimed to develop a rapid predictor that captures these nonlinear behaviors and axonal responses to complex multipolar fields. Approach: We trained an artificial neural network (ANN) to predict axonal activation in peripheral nerve axons in response to multipolar stimuli from dense voltage sources around each nerve, as predicted by the MRG double-cable model. Main results: The implemented ANN achieves 98.59% ± 0.04% (95% C.I.) accuracy in axonal activation prediction over pulse widths from 10 to 200 μs, myelinated axon diameters from 5 to 20 μm, and stimulation amplitudes between −3 mA and 3 mA per contact in multipolar stimulation configurations. It demonstrates an increase of more than three orders of magnitude in computational speed compared to the MRG double-cable model, with less than a 2% loss in accuracy with nonlinear behaviors included. Significance: By capturing nonlinear axonal behavior in response to multipolar stimulation that existing linear rapid estimators miss — while retaining a speed advantage of more than three orders of magnitude over biophysical cable models — the ANN enables accurate, high-throughput exploration of large multipolar stimulation parameter spaces that would otherwise be computationally prohibitive.