Efficient, Robust, and Accurate CNN Predictor for Neuronal Activation in Directional Deep Brain Stimulation

The programming of clinical deep brain stimulation (DBS) systems involves numerous combinations of stimulation parameters, such as stimulus amplitude, pulse width, and frequency. As more complex electrode designs, such as directional electrodes, are introduced, the traditional trial-and-error approach to manual DBS programming becomes increasingly impractical. Visualization of the volume of tissue activated (VTA) can assist in selecting stimulation parameters by showing the direct effects of DBS on neural tissue. However, the standard method for VTA calculation, which involves modeling biological nerve fibers, is highly time-consuming and limits clinical applicability. In this study, we used finite element models (FEM) of implanted DBS systems to compute electric fields and obtained a large dataset of axonal responses under electrical stimulation using multicompartment cable models. We then trained a convolutional neural network (CNN) to replace the cable models. The CNN model’s performance in calculating VTA was evaluated across various electrode configurations and stimulation parameters, and compared with existing activation function (AF) methods. The CNN model achieved a mean absolute error (MAE) of 0.032V in predicting nerve fiber activation thresholds, demonstrating greater stability and accuracy in VTA prediction compared to the AF method. Additionally, the CNN reduced computation time by five orders of magnitude compared to standard axonal modeling methods. We demonstrate that the CNN-based neural fiber predictor can quickly, accurately, and robustly predict neural activation responses to DBS, thereby improving the efficiency of DBS programming.