Abstract
Abstract: :
Purpose: Perimetric test algorithms using Bayesian principles for threshold estimation have enjoyed popularity in recent times both in the research literature and in commercial perimeters. All such methods have used either an initial probability density function (pdf) based on population data, or a uniform pdf that makes no assumptions about the patient. When retesting a patient, however, all of the information gleaned in the previous test can be used to alter the initial pdf in a Bayesian procedure. This study introduces an approach to choosing a pdf for re–testing a patient. It uses computer simulation to demonstrate that the re–test algorithm, RE–ZEST, is as accurate as an initial test algorithm, but much faster. Methods: 265 normal and 163 glaucomatous SAP visual fields (VF) were used as input to computer simulations, each run under three error models (no errors, 15% false positives,15% false negatives). The first simulation measured the VF using a ZEST procedure employing a population based pdf as used in our previous studies. For each patient, for each location of the measured visual field, all possible response sequences that could have generated the threshold measurement for that location were calculated. Each sequence was assigned a probability according to frequency–of–seeing curves assumed for the patient based on variability information used in the initial VF simulation. These probabilities were then combined to form a "re–test" pdf of possible thresholds at that location for that patient. A second simulation was then run using ZEST and the re–test pdf at each location (RE–ZEST). The mean number of presentations and error in threshold estimate compared to the known true threshold were calculated. An alternate technique, G–ZEST, was also trialed where the re–test pdf was set to be a Gaussian distribution with a mean equal to the threshold measured in the first test, and a standard deviation 3dB. Results:ZEST and G–ZEST required an average of about 6 presentations per location, while RE–ZEST only requires 4 for variable patients and 1–2 for perfect patients. The mean error for all methods was within 0.5 dB of each other over all patient sets, averaging 1.7dB overall. Conclusions: This work introduces a technique for deriving pdfs for use with the ZEST algorithm when re–testing a visual field. This RE–ZEST technique incorporates information from the initial test, and is as accurate as running the test again using the original population based pdf, but is 30 to 60% faster. RE–ZEST is faster than and as accurate as G–ZEST. Perhaps the presentations saved in RE–ZEST can be employed to reduce re–test variability.
Keywords: perimetry • clinical (human) or epidemiologic studies: systems/equipment/techniques • visual fields