gms | German Medical Science

Objective: To improve the perceptual experience of patients using visual neuroprostheses by optimizing patient-specific stimulation parameters through a deep neural network encoder combined with human-in-the-loop Bayesian optimization.

Materials and Methods: We trained a deep neural network encoder to invert a forward model of the visual system, enabling the generation of optimal stimuli tailored to individual patients. This was integrated with a human-in-the-loop (HILO) preferential Bayesian optimization (PBO) strategy, where a blind volunteer with a Utah array implanted in their visual cortex participated in a two-alternative forced choice brightness discrimination task. The task involved varying temporal modulation patterns of microstimulation while maintaining constant delivered current. The volunteer's feedback continually updated a Gaussian Prior and an acquisition function, guiding the optimization of temporal stimulation parameters.

Results: Our approach produced high-fidelity stimuli that outperformed conventional strategies. The HILO-PBO method significantly improved the detectability and thresholds of phosphenes across multiple electrodes and sessions. The optimized patterns required lower charge thresholds than baseline patterns, demonstrating robust and substantial improvements in perceptual quality.

Discussion: The integration of deep learning and Bayesian optimization addresses the challenge of high-dimensional stimulus optimization and variability in patient-specific responses. Our results suggest that this HILO approach could pave the way for more natural and effective sensory experiences in visual neuroprostheses, advancing the development of intelligent neuroprosthetic systems.