Abstract
Purpose :
The promise of visual health monitoring and personalized medicine depends on vision metrics that can precisely track an individual’s vision over time. Common proxies for test precision are based on repeatability, such as the coefficient of repeatability (CoR). However, precision and repeatability are not the same. A test with coarse resolution may be repeatable, but changes in vision within or between individuals are obscured by large steps between test scores. To address this confound, we developed a new Fractional Rank Precision (FRP) metric to evaluate the precision of visual testing, based on concepts of machine learning: how well can an individual be identified in the population distribution of retest measures, based on their initial test measure? We assessed 3 vision tests using FRP: ETDRS visual acuity (VA), Pelli-Robson (PR) contrast sensitivity (CS), and quick Contrast Sensitivity Function (qCSF) testing.
Methods :
From healthy observers (20-85 years), we obtained 164 monocular and 100 binocular test-retest pairs of qCSF (one week apart). For a broad, scalar summary statistic, we computed the Area Under the Log CSF (AULCSF) from 1.5 to 18 cycles per degree. We also collected 189/180 test-retest pairs from PR CS and ETDRS VA testing. For each test, we computed CoR and FRP: the rank of the retest of a subject when all subjects’ retests are sorted by their similarity to a subject’s initial test, averaged across all subjects. FRP ranges from .5 (chance) to 1.0 (perfect identification of test from retest for each subject). We also recomputed FRP for increasing quantization, i.e. rounding of values to coarse step sizes.
Results :
CoR and FRP were .214 and .844 (AULCSF), .243 and .721 (PR CS), and .149 and .718 (ETDRS VA), respectively. As expected, increasing quantization reduced FRP. The precision of AULCSF was reduced to that of unmodified (non-quantized) PR CS and ETDRS VA, when strong quantization collapsed the AULCSF population distribution to only 5 step-sizes.
Conclusions :
The FRP metric is sensitive to a test's resolution (step-size), variability (CoR), and dynamic range. Despite apparently better repeatability (lower CoR), the precision of ETDRS VA was similar to that of PR CS. The AULCSF provides highest FRP despite intermediate CoR, due to small step-sizes and low variability relative to its range. These features may be useful to detect visual changes in clinical trials and clinical practice.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.