Statistical modeling to improve the detection of glaucoma progression

• Main Author: Kummet, Colleen
• Other Authors: Burns, Trudy L.
• Format: Others
• Language: English
• Published: University of Iowa 2013
• Subjects: Glaucoma; Perimetry; Progression; Statistics; Visual Field Testing; Clinical Epidemiology
• Online Access: https://ir.uiowa.edu/etd/5008; https://ir.uiowa.edu/cgi/viewcontent.cgi?article=5008&context=etd

Glaucoma is the second leading cause of blindness affecting over 60 million people worldwide. The objectives of this study were to expand the existing methods of trend analysis in visual field time series data testing to aid in the early and accurate detection of glaucoma progression. Visual field data including 54 locations for each of 140 eyes (one per participant among 96 cases and 44 controls) were evaluated using the Humphrey Field Analyzer II program 24-2 Swedish interactive thresholding algorithm (SITA) standard test strategy and Goldmann size III stimuli. One eye was randomly selected for the study and data were collected between 2003 and 2009. Two visual field examinations were conducted at baseline and at eight additional time points of visual field exams taken every six months for four years. Demographic, clinical, structural and other health data in the VIP study were collected from the electronic medical record and health questionnaires. A variety of pointwise linear regression (PLR) criteria have been proposed for determining glaucomatous visual field progression. However, alternative PLR criteria have only been assessed on a limited basis. The first aim of this glaucoma progression detection study thoroughly examined PLR cut-point criteria to maximize the sensitivity and specificity of this standard tool in visual field analysis. The pointwise linear regression A2 (PLRA2) method was used to analyze the data, and Ocular Hypertension Treatment Study (OHTS) data were used to validate the decision rule. Results showed that visual field trend analysis using PLR can be refined by adjusting the standard slope-based and significance level-based criteria. By considering more restrictive declines in visual field data (e.g., < -1.2dB/y, which is approximately 12 times the normal rate of age-related decay) and relaxing the significance level criterion of the PLR slope to p < 0.04 a high specificity can be maintained, while increasing the hit rate, i.e., the proportion of glaucoma cases in which progression was detected by PLR. This work serves to improve a familiar and commonly used method of time series visual field trend analysis that can be implemented quickly to improve early detection of glaucoma progression. The second aim of this project was to investigate the performance of the nonlinear exponential and tobit regression models relative to the normal regression model in the analysis of visual field decay. The goodness-of-fit, as measured by Akaike Information Criteria (AIC), and rates of progression obtained by fitting three alternative regression models to longitudinal visual field data were compared at the location level. The results showed that visual field trend analysis using the tobit regression model results in a better model fit to visual field data, increased precision in the estimation of the rate of progression, and provides a specific advantage in modeling data from cases with advanced glaucoma. The third, and final, aim of this glaucoma progression research project sought to determine if demographic, clinical and health factors, including intraocular pressure, retinal nerve fiber layer thickness, hypertension and diabetes, differ in participants whose visual field data are best fit overall by one statistical model compared to another. This was the first study to examine person-level factors that may affect the fit of proposed analysis models for visual field data, and to utilize bivariable and multivariable methods to understand patient-level predictors of visual field model fit. In the majority of eyes, the tobit model provided either a significantly better fit or there was no difference among models. Significant differences in patient characteristics included baseline MD and previous ocular surgery. This indicates that the tobit model may fit visual field time series data at least as well as the normal and nonlinear exponential models in all cases and controls; and in some advanced cases, it may provide a significantly improved fit. This research overcame critical barriers in visual field trend analysis by increasing the sensitivity of PLR methods and further developing methods using alternative distributions to determine significant loss of function within each visual field test location. Furthermore, results of this study will contribute to the ongoing improvement of visual field trend analysis and the early detection and treatment of glaucoma progression.