Explainable AI-Driven Optimization of Electrode Activation Reduces Power Consumption While Preserving Object Recognition in Retinal Prostheses

Abstract

Retinal prostheses using electrical stimulation are developed to restore partial visual function perception in blind individuals, enabling object and large-character recognition. These devices capture visual scenes via a camera, process them, and stimulate the remaining retinal neurons through an implanted microelectrode array, eventually eliciting percepts known as “phosphenes”. However, unintended current spread among electrodes remains a significant challenge, leading to blurred and/or distorted percepts. Therefore, optimizing the spatial activation patterns is critical for enhancing visual resolution and reducing power consumption. In this study, we introduce a novel application of explainable artificial intelligence (XAI) to optimize electrode activation patterns in retinal prostheses. XAI-driven phosphene generation prioritizes visually informative regions while suppressing less relevant areas, enabling more efficient stimulation, while reducing the total current usage without compromising object recognition accuracy. To replace labor-intensive psychophysical testing, deep learning models evaluate the recognition performance of different XAI-informed stimulation methods. Our findings reveal that XAI-driven activation patterns consistently achieve higher energy efficiency than conventional grayscale-based electrode mapping without compromising recognition performance. Moreover, XAI provides interpretable maps of electrode engagement, facilitating clinical understanding and personalized stimulation strategies. These results highlight the potential of XAI to enhance the efficiency, interpretability, and functional outcomes of next-generation retinal prosthetic systems.