Jonathan Skaza¹, Shravan Murlidaran¹, Apurv Varshney¹, Ziqi Wen¹, Miguel P Eckstein¹, Michael Beyeler¹; ¹University of California, Santa Barbara

Introduction: Visual prostheses aim to restore sight, but current devices generate only rudimentary phosphenes with limited visual capabilities. Predicting patient perception before implantation is crucial for evaluating device performance, optimizing design, and setting realistic expectations. We present a novel Computational Virtual Patient (CVP) pipeline to predict perceptual performance. Methods: The CVP pipeline simulates prosthetic perceptual experiences using two models: a traditional “scoreboard” approach and an “axon map” model that leverages the spatial layout of retinal ganglion cell axons (Beyeler et al., 2019). These simulations, constrained by psychophysical and neurophysiological data, allow customization of electrode configurations, including tests with 6x10 (Argus II), 6x15, and 12x20 arrays. Sighted participants (n = 18) completed shape and facial emotion classification tasks using simulated phosphenes. Additionally, a ResNet-18 deep neural network (DNN) was fine-tuned through transfer learning, with its final classification layer retrained on a simulated phosphene dataset to predict perceptual performance. Results: The DNN accurately predicted the hierarchical ordering of task difficulty, with binary emotion detection (64.2% with the “axon map” 12×20 configuration) proving more challenging than three-class shape classification (84.3% with the same configuration). Human and DNN results showed similar effects of electrode configuration on perception. For shape classification tasks, increasing electrode resolution from 6×15 to 12×20 improved accuracy in humans (64.2% ± 1.7% to 85.7% ± 1.3%) and the DNN (62.6% ± 3.6% to 84.3% ± 2.6%). Fully trained ResNet-18 and in-house DNNs were not consistent with human performance, suggesting that humans may rely on pre-existing visual processing rather than developing new perceptual mappings. Conclusion: Our computational framework demonstrates the potential to predict prosthetic vision capabilities across multiple tasks and devices, offering a novel approach for evaluating visual prostheses pre-implantation. This tool could accelerate device development and provide more accurate expectations for implant recipients.