Artificial neural network-based rapid predictor of biological nerve fiber activation for DBS applications

Objective.Computational models are powerful tools that can enable the optimization of deep brain stimulation (DBS). To enhance the clinical practicality of these models, their computational expense and required technical expertise must be minimized. An important aspect of DBS models is the prediction of neural activation in response to electrical stimulation. Existing rapid predictors of activation simplify implementation and reduce prediction runtime, but at the expense of accuracy. We sought to address this issue by leveraging the speed and generalization abilities of artificial neural networks (ANNs) to create a novel predictor of neural fiber activation in response to DBS.Approach.We developed six variations of an ANN-based predictor to predict the response of individual, myelinated axons to extracellular electrical stimulation. ANNs were trained using datasets generated from a finite-element model of an implanted DBS system together with multi-compartment cable models of axons. We evaluated the ANN-based predictors using three white matter pathways derived from group-averaged connectome data within a patient-specific tissue conductivity field, comparing both predicted stimulus activation thresholds and pathway recruitment across a clinically relevant range of stimulus amplitudes and pulse widths.Main results.The top-performing ANN could predict the thresholds of axons with a mean absolute error (MAE) of 0.037 V, and pathway recruitment with an MAE of 0.079%, across all parameters. The ANNs reduced the time required to predict the thresholds of 288 axons by four to five orders of magnitude when compared to multi-compartment cable models.Significance.We demonstrated that ANNs can be fast, accurate, and robust predictors of neural activation in response to DBS.