Fig. 5: Optimization of current amplitudes (1200 degrees of freedom) for spatially selective activation of 5.7 µm fibers using arbitrary waveforms.
From: Highly efficient modeling and optimization of neural fiber responses to electrical stimulation
a–b Example selectivity performance (Supplementary Note 1 P4) using gradient-free optimization (differential evolution; DE) and NEURON (a), and gradient descent (GD) with the surrogate model (S-MF) (b). c Final optimized stimulation waveforms corresponding to panels a, b. The total pulse duration for every contact for all optimizations was constrained to 1 ms (between 0.2 and 1.2 ms). Frame colors and numbers correspond to contact colors and numbers in a, b. d Summary of optimization performance using arbitrary waveforms across all 12 nerves (6 pig vagus nerves in red and 6 human vagus nerves in blue) for both optimization methods: % target fascicles activated, % off target fascicles activated, and weighted binary cross entropy (Eq. (18)). We assumed that all fibers of a specific diameter within a given fascicle had the same threshold78; we thus converted percent fibers activated to percent fascicular area activated. For all optimized stimulus parameters generated by algorithms using S-MF (GD + S-MF), we evaluated the ‘true’ fiber activations using the NEURON model and report performance metrics for those activations (Supplementary Note 7). e Compute time for each optimization method on different hardware, per nerve (left, n = 12) and total (right). Compute time for 4 CPUs was estimated from 350 CPUs assuming linear scaling. d, e Boxplots are displayed using the Tukey method (center line, median; box limits, upper and lower quartiles; whiskers, maximum / minimum point within 1.5× interquartile range of nearest hinge). Source data are provided as a Source Data file.
