Purchase this article with an account.
J. L. Haines, K. L. Spencer, A. Agarwal, E. A. Postel, A. Iannaccone, S. Satterfield, K. C. Johnson, S. B. Kritchevsky, W. K. Scott, M. A. Pericak-Vance; Progress in Predicting Risk for Age-Related Macular Degeneration. Invest. Ophthalmol. Vis. Sci. 2009;50(13):1600.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Interest in genetic testing for age-related macular degeneration (AMD) has grown as the number of confirmed AMD loci has risen. We evaluated 3 modeling strategies which could be used to create a genetic test using age, smoking history, and CFH Y402H and ARMS2 A69S genotypes as predictor variables.
We used a subset of 352 AMD cases and 184 controls ascertained through Vanderbilt University (VU) and the University of Miami (UM) to create logistic regression (LR), decision tree (DT), and Bayesian classification (BC) models. We applied these models to an independent dataset of 89 cases and 48 controls from VU/UM and to an independent set of 86 cases (mainly AREDS cat. 3 AMD) and 149 elderly controls from the Age-Related Maculopathy Ancillary (ARMA) Study mainly drawn from the Health ABC cohort.
In the VU/UM dataset, the DT method performed best with 77% of the testing dataset correctly classified. Considering only those instances where all 3 models agreed (52% of individuals), 85% of individuals were classified correctly (88% of cases, 77% of controls). In the ARMA cohort, the BC model performed best with 54% of individuals correctly classified. The overall accuracy when all 3 methods agreed (51%) was less than for the BC model alone. Using the consensus of all 3 methods, the "affected" classification was much more likely to be correct than the "control" classification (79% vs 34%). The ARMA cohort and VU/UM dataset significantly differed in mean age and proportion of Y402H risk allele carriers, contributing to decreased accuracy in the ARMA cohort.
Using this risk profile will identify particularly high-risk individuals at the expense of a substantial false positive rate. Selectively applying this method to more homogenous populations may help improve accuracy. Methods to further refine this algorithm are currently being employed.
This PDF is available to Subscribers Only