Abstract
Purpose :
Controlling the dynamics of electrically induced perception in neural prostheses requires threshold estimation for each electrode, which becomes burdensome for high electrode counts. In this study, we compared four common threshold estimation methods in terms of speed of convergence and accuracy for epiretinal Argus II visual implants.
Methods :
Threshold testing methods that are compared here are 1) the standard hybrid (H) method for Argus II epiretinal implants — a combination of the method of constant stimuli (non-adaptive) and maximum likelihood estimation (MLE; adaptive); 2) stochastic approximation staircase (SC; adaptive); 3) minimizing the variance (MV; adaptive); and 4) maximizing the information (MI; adaptive). Data were collected concurrently across all electrodes for adaptive methods, but by groups of 6 electrodes for the H method.
Weibull psychometric function thresholds were estimated in a yes/no experiment for two Argus II users for 24 and 56 electrodes respectively using the four methods. For each subject, data were collected in one session per adaptive method, and two sessions for the hybrid method. Subject responses across all sessions were used offline to estimate the thresholds and confidence intervals (CI) with MLE. To test method accuracies, we computed the average normalized root mean square error (RMSE) between estimated thresholds and MLE thresholds.
Results :
Fig.1 shows that hybrid threshold testing requires more trials compared to adaptive methods and only 25% of the thresholds estimated by H were within MLE estimated CI, compared to SC 35%, MV 58%, and MI 61%. The adaptive methods were also more accurate compared to H (RMSE for H: 11%, SC: 10%, MV: 6%, and MI: 6%).
Conclusions :
Our results show that the adaptive methods provide more accurate threshold estimates within fewer trials. Adaptive methods enable threshold estimation for a greater number of electrodes concurrently in a session, which is important for neural prostheses with a high number of electrodes.
This is a 2021 ARVO Annual Meeting abstract.