Versatile Neural Activation Predictor With Axon Structure Tailoring Capability Enabling Personalized Neuromodulation Computation

Abstract

Neuromodulation therapies are evolving to be more and more intelligent and personalized, driving the need for more precise and efficient stimulation strategies. Biophysically detailed computational models integrated with anatomically accurate neural structures could offer critical insights into neural activation patterns under various stimulation conditions, which are essential to optimize the treatment. However, solving these models containing a large number of nerve fibers is computationally intensive, especially when the neural targets comprise heterogenous axons, e.g., with varying geometries. Also, current methods lack generalizability across various neuromodulation scenarios, limiting the scalability and clinical utility of such models. In this study, we present a convolutional neural network (CNN)-based framework as a universal, rapid, and accurate alternative to conventional case-by-case brutal force computation methods. Our approach achieves a mean absolute error (MAE) of $\boldsymbol {\textbf {6}.\textbf {91}\times \textbf {10}^{-\textbf {3}}}$ mV and over 95% prediction accuracy under diverse extracellular stimulation scenarios, facilitating personalized simulations and tailored neuromodulation treatments.