Efficient Spatial Estimation of Perceptual Thresholds for Retinal Implants via Gaussian Process Regression.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Ophthalmology

Background:

• Retinal prostheses aim to restore vision via electrical stimulation.
• Calibration of perceptual thresholds is crucial but time-consuming for high-electrode devices.

Purpose of the Study:

• To develop and evaluate a Gaussian Process Regression (GPR) framework for efficient retinal prosthesis calibration.
• To compare GPR with different kernels and sampling strategies for threshold prediction accuracy.

Main Methods:

• Utilized perceptual threshold data from four Argus II retinal prosthesis users.
• Implemented GPR with Matern and Radial Basis Function (RBF) kernels.
• Compared spatially optimized sampling and adaptive sampling with uniform random sampling.

Main Results:

• GPR with a Matern kernel significantly outperformed RBF kernel for threshold prediction (p < .001).
• Spatially optimized sampling reduced prediction error compared to random sampling in two participants (p < .05).
• Adaptive sampling showed non-significant accuracy gains over spatial sampling, but approached significance for one participant (p = .074).

Conclusions:

• GPR combined with spatial sampling offers a scalable and efficient method for retinal prosthesis calibration.
• This approach minimizes patient burden and maintains predictive accuracy.
• The framework provides a generalizable solution for adaptive calibration in neuroprosthetic devices with spatially structured thresholds.