Abstract
An improved method for detecting coronary heart disease (CHD) could have substantial clinical impact. Building on the idea that systemic effects of CHD risk factors are a conglomeration of genetic and environmental factors, we use machine learning techniques and integrate genetic, epigenetic and phenotype data from the Framingham Heart Study to build and test a Random Forest classification model for symptomatic CHD. Our classifier was trained on n = 1,545 individuals and consisted of four DNA methylation sites, two SNPs, age and gender. The methylation sites and SNPs were selected during the training phase. The final trained model was then tested on n = 142 individuals. The test data comprised of individuals removed based on relatedness to those in the training dataset. This integrated classifier was capable of classifying symptomatic CHD status of those in the test set with an accuracy, sensitivity and specificity of 78%, 0.75 and 0.80, respectively. In contrast, a model using only conventional CHD risk factors as predictors had an accuracy and sensitivity of only 65% and 0.42, respectively, but with a specificity of 0.89 in the test set. Regression analyses of the methylation signatures illustrate our ability to map these signatures to known risk factors in CHD pathogenesis. These results demonstrate the capability of an integrated approach to effectively model symptomatic CHD status. These results also suggest that future studies of biomaterial collected from longitudinally informative cohorts that are specifically characterized for cardiac disease at follow-up could lead to the introduction of sensitive, readily employable integrated genetic-epigenetic algorithms for predicting onset of future symptomatic CHD.