Variance reduction for visual field testing algorithms through optimization of initial estimation

Abstract

Purpose : Variance of the probability mass function (PMF) for sensitivity estimation is an important measure for visual field (VF) testing algorithms, as a smaller variance indicates that the algorithm is more confident that the PMF represents the true sensitivity. We developed and tested a variance reduction technique to optimize the initial sensitivity estimation, thereby improving the overall performance of visual field testing algorithms.

Methods : The Zippy Estimation by Sequential Testing (ZEST) algorithm was simulated on our visual field dataset with 504 normal and 784 glaucomatous visual fields with a 24-2 testing pattern. Visual fields were divided into multiple VF regions and fixed the testing locations of the first batch of light stimuli to be inside the nasal and arcuate region. After obtaining the sensitivity estimations from these testing locations, a gradient boosting decision tree classifier was used to determine whether the visual field is from a normal or abnormal group. Then, the initial PMFs of the remaining locations were set to be the empirical distribution obtained from a normal or abnormal group based on prediction outcomes. Finally, these locations were tested with the adjusted initial PMFs.

Results : By adopting the proposed variance reduction approach, ZEST becomes 4.57% faster and 4.17% more accurate (p<0.01). Meanwhile, the gradient boosting decision tree classifier achieves a 97% average classification accuracy.

Conclusions : Our results indicate that using a carefully designed testing sequence can effectively reduce the variance in sensitivity estimation and greatly improve the quality of visual field tests, leading to shortened testing time, higher testing accuracy, and improved patient’s satisfaction.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.