Purchase this article with an account.
G Buitendijk, Elena Rochtchina, Chelsea Myers, Sudha Iyengar, Paul Mitchell, Johannes Vingerling, Jie Wang, Ronald Klein, Caroline Klaver, ; Increasing the fidelity of AMD prediction models using population data from three continents. Invest. Ophthalmol. Vis. Sci. 2013;54(15):224.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Most prediction models for age-related macular degeneration (AMD) are based on case-control studies. Due to inclusion of only the extreme ends of the clinical spectrum, these studies have a tendency to over-estimate risks. The aim of this study is to develop a prediction model for late AMD based on data from three population-based studies that included both early and no AMD.
Participants included those with gradable fundus photographs, follow-up, and genotype data (n=8094) from the population-based Rotterdam Study (RS), the Beaver Dam Eye Study (BDES) and the Blue Mountain Eye Study (BMES) in the 3-Continent AMD Consortium. Incident late AMD (total 4-5 visits; median follow-up 11.1 years) was determined on fundus photographs at each visit using the AMD Consortium Harmonized 5-Level AMD Scale. An initial prediction model developed on the RS population included all available risk factors. This model was validated in BDES and BMES. A risk score was constructed using regression coefficients from Cox proportional hazards analysis; predictive values (AUC) were estimated by receiver operating characteristic (ROC) curves; and absolute risks were estimated using Kaplan-Meier product-limit analysis.
In all, 278 participants developed incident late AMD, 2720 early AMD, and 5096 remained free of any AMD. The predictive value was 0.88 in the initial model, and 0.85 in the validation (Z-test P>0.05). The best prediction was achieved with age, sex, 25 SNPs in AMD risk genes, smoking, BMI, and baseline AMD phenotype in the model. The risk score varied from -3 to 8 with the hazard of 22.0 (95%CI 15.2-31.8) for the highest risk scores (5-8). Life time risks up to age 90 were 60% for the highest risk scores and virtually nought for the lowest risk scores.
Our study distinguishes well between the risk of those with a low number of risk factors and those with many risk factors . The population-based setting may provide a more generalizable risk assessment measure than a case-control setting.
This PDF is available to Subscribers Only