Incorporating spatial information into visual field testing algorithms

© 2017 Dr. Nikki Juliet RubinsteinCurrent clinical perimetric thresholding algorithms are susceptible to high test-retest variability in areas of low vision and require long periods of patient concentration. A possible approach to addressing these limitations is exploiting spatial information in the visual field to inform the choice of stimulus location and intensity. This thesis aimed develop new perimetric algorithms using this approach, in an attempt to reduce test times and test-retest variability. The first experiment described in this thesis aimed to develop an algorithm that uses a heuristic approach to incorporate spatial information (Chapter 3). The new algorithm — SWeLZ (Spatially Weighted Liklelihoods in ZEST) — is a Bayesian procedure that uses modified likelihood functions to update information at multiple locations after each stimulus presentation. Computer simulations showed that SWeLZ achieved a comparable level of error to ZEST across all levels of sensitivity. SWeLZ was able to detect both glaucomatous and non-glaucomatous localised visual field loss. The number of presentations was reduced by 25% for SWeLZ relative to ZEST for normal visual fields. This reduction in test time was not found for visual fields with areas of visual field loss. The second experiment described in this thesis aimed to develop an algorithm that uses a Markov Random Field representation of the visual field to drive stimulus placement (Chapter 4). The newly developed algorithm—BART (Bayesian Adaptive Random Test) — uses local conditional probability distributions to describe likely sensitivity values. Many variants of the procedure were explored. BART was set to terminate in the same number of presentations as ZEST. BART achieved a similar level or error to ZEST for simulations when simulated with a reliable responder, but tended to overestimate visual sensitivity values in areas of absolute visual field loss, when simulated with a typical false positive responder. As is common in algorithm development, the newly developed algorithms were tested using computer simulation, which allows many tests to be run in a relatively short period of time. The final experiment described in this thesis aimed to explore the assumptions underpinning simulated responses (Chapter 5). Frequency of seeing curves were measured at three locations in the visual field for 16 observers with glaucoma, using both a forced choice and a yes-no procedure. The relationship found between sensitivity and frequency of seeing curve upper asymptote found in yes-no experiments disappeared when observers were forced to choose. This finding suggests that observers preferentially allocate attention towards locations with higher sensitivity. These experiments confirm previous findings that areas of moderate-to-severe visual field loss are associated with high test variability and reduced maximum response rates when measured with white-on-white standard automated perimetry. Given the small amount of information gleaned from testing these locations, it may be prudent to focus testing power elsewhere, such as on spatial extent of defects rather than defect depth