Combining effects in genetic risk models for myocardial infarction

Abstract

Risk scoring models can be utilized as supportive elements within preventive strategies in healthcare settings through prediction of disease risk. The purpose of this study was to improve on and adapt an existing genetic risk scoring methodology to the Maltese population. Three main experiments have been proposed, each comprising the same designed model but using a variety of genetic and non-genetic features to generate the Genetic Risk Scores (GRSs). Linear regression analysis was the initial method used to develop GRSs with weightings for each genotype based on the effect on an intermediate phenotype, Triglycerides (TG). Odds Ratios (ORs) were calculated using a logistic regression model to compare the odds of having an MI in four GRS groups. Subsequently, three non-linear regression models replaced the linear element to observe differences in the ORs generated and to compare performance metrics. When all environmental risk factors were featured in the linear model, men in GRS 2 resulted in an OR of 2.7 (95% CI= 1.3- 5.9) which gradually increased to 14.1 (95% CI= 7.1- 28.4) in GRS 4. Similarly, women in GRS 2 had an OR of 5.0 (95% CI=0.8- 31.7) which also increased to 27.0 (95% CI= 4.8-151.5). Similar results were obtained when all risk factors and genetic variables were included in the linear model. The OR in men increased from 3.0 (95% CI= 1.5- 6.2) in GRS 2 to 11.9 (95% CI= 6.1- 23.1) in GRS 4. The women shared a similar trend where the OR was lowest in GRS 2 (OR= 1.0, 95% CI= 0.7-8.1) and highest in GRS 4 (OR= 5.0, 95% CI= 2.0-18.1). Replacing the linear component of the model with either Multi-Layer Perceptron (MLP) or Support Vector Regression (SVR) did in fact improve the performance in the majority by decreasing error and increasing the R2. Adding additional risk variables and genetic polymorphims to the currently validated clinical risk scores not only increased risk for MI but also increased model performance.