Purchase this article with an account.
Linda C. McCarthy, Paul J. Newcombe, John C. Whittaker, John I. Wurzelmann, Michael A. Fries, Sandra W. Stinnett, Trupti M. Trivedi, Chun-Fang Xu; Predictive models of Geographic Atrophy incidence within short timelines (<3 years). Invest. Ophthalmol. Vis. Sci. 2012;53(14):3323.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Comprehensive predictive models were developed for Geographic Atrophy (GA) incidence within 3 years, motivated by possible application to clinical trial design. Clinical, environmental, demographic and genetic risk factors were assessed, and the contribution of genetic markers to improvement in predictive performance was evaluated.
The predictive performance of GA risk factors was explored in regression models using data from 2,003 subjects from the Age-Related Eye Disease Study (AREDS) study. 10-fold cross-validated receiver operating characteristic (ROC) curves were used to compare the performance of predictive models. We compared the performance of our models to a well documented simple clinical model: the AREDS Simplified Severity Scale.
Logistic regression models which included clinical, demographic and environmental factors, had better predictive performance for 3 year GA incidence, than the AREDS Simplified Severity Scale, as measured by the ROC area under the curve (AUC). Regression models for three year GA incidence had a ROC AUC of 0.89 (0.03), compared to a ROC AUC of 0.76 (0.05) for an AREDS Simplified Severity Scale score of 4, and a ROC AUC of 0.84 (0.07) for an AREDS Simplified Severity Scale score of ≥3. Although some genetic factors are significantly associated with both 2 and 3 year GA incidence after adjustment for other GA risk factors (CFH Y402H, p=0.03; rs1410996, p=0.05), their addition to the regression model did not meaningfully improve predictive performance (ROC AUC: 0.89).
Although simple clinical models such as the AREDS Simplified Severity Scale combine good predictive performance with ease of application in the clinic, our regression models can be more effective predictors of GA incidence within 3 years. For application to recruitment of patients into GA prevention clinical trials, the regression models also offer the flexibility to select risk thresholds which provide a practical balance between a desired GA conversion rate and the size of the population screened for clinical trial recruitment.
This PDF is available to Subscribers Only