Abstract
A major goal of personalized medicine is to pre-symptomatically identify individuals at high risk for disease using knowledge of each individual's particular genetic profile and constellation of environmental risk factors. With the identification of several well-replicated risk factors for age-related macular degeneration (AMD), the leading cause of legal blindness in older adults, this previously unreachable goal is beginning to seem less elusive. However, recently developed algorithms have either been much less accurate than expected, given the strong effects of the identified risk factors, or have not been applied to independent datasets, leaving unknown how well they would perform in the population at large. We sought to increase accuracy by using novel modeling strategies, including multifactor dimensionality reduction (MDR) and grammatical evolution of neural networks (GENN), in addition to the traditional logistic regression approach. Furthermore, we rigorously designed and tested our models in three distinct datasets: a Vanderbilt-Miami (VM) clinic-based case-control dataset, a VM family dataset, and the population-based Age-related Maculopathy Ancillary (ARMA) Study cohort. Using a consensus approach to combine the results from logistic regression and GENN models, our algorithm was successful in differentiating between high- and low-risk groups (sensitivity 77.0%, specificity 74.1%). In the ARMA cohort, the positive and negative predictive values were 63.3% and 70.7%, respectively. We expect that future efforts to refine this algorithm by increasing the sample size available for model building, including novel susceptibility factors as they are discovered, and by calibrating the model for diverse populations will improve accuracy.
