Abstract
Purpose :
To compare a new perimetric thresholding algorithm (trail traced threshold test; T4) with zippy estimation by sequential testing (ZEST) in computer simulations. Conventional algorithms estimate the threshold and update the estimate with subsequent responses; prior responses are 'forgotten', thus response data are not fully utilized. This study evaluates a new algorithm (T4), which takes into account all responses as well as spatial correlation amongst test locations.
Methods :
A probability of seeing (PoS) curve at each location is modeled with a reversed cumulative normal distribution, the center of which represents the threshold. The likelihood of responses at each location is modeled as a weighted binomial distribution in which the weights of neighbor locations are defined according to anatomical distribution of retinal nerve fibers. Prior distributions of the PoS parameters were also imposed to control the 'stiffness' of the PoS curve. The parameters of PoS were optimized by maximizing the posterior distribution at individual locations. Each stimulus updates not only the PoS parameters at its own location but also those of the neighbors. The next stimulus is proposed at the location with the largest PoS standard deviation (SD) and at the level of the center. The algorithm terminates at one location if the SD of the PoS is within 0.5dB or after 10 stimulus presentations. In computer simulations, 'ground truth' VFs were the mean of 10 Humphrey Field Analyzer 24-2 VFs acquired within 8 weeks from 218 eyes of 109 glaucoma patients. Patients responses were simulated by frequency of seeing curves at various false positive (FP) and false negative (FN) rates. T4 and ZEST were compared: presentation number, accuracy (mean absolute difference; MAD) , and measurement variability.
Results :
With all simulated VFs, the number of presentations of T4 and ZEST is 151 (SD of 36) and 255 (SD of 16). T4 required 31-43% fewer presentations than ZEST in all FN and FP configurations (Figure A). T4 is more accurate than ZEST in all FN and FP configurations (Figure B): mean (SD) MAD between true and T4- and ZEST-estimated thresholds is 2.9dB (SD 0.6dB) and 4.3dB (SD 0.6dB). Measurement variability of T4 is lower than that of ZEST (Figure C).
Conclusions :
Utilizing the complete response sequence, together with spatially weighted neighbor responses, achieves better accuracy and precision than ZEST with significantly fewer stimulus presentations.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.