Selection of 51 predictors from 13,782 candidate multimodal features using machine learning improves coronary artery disease prediction

Figures

• 

  Figure 1. Flow diagram illustrating exclusion criteria and 5-fold cross-validation procedure Prevalent cardiovascular disease included coronary artery disease, myocardial infarction, stroke, heart failure, and peripheral vascular disease. Five-fold cross-validation was used to select a range of models for subsequent clinician-review (Figure 2).

• 

  Figure 2. C-statistics in training and testing data as a function of the regularization hyperparameter The right white region represents an area of steep C-statistic growth on both the training and testing data, where adding predictors substantially improves prediction. In the left white region, the testing (green) and training (red) curves are diverging, representing a model that performs well in the training data but generalizes poorly to unseen test data. The blue region is an area of slow C-statistic growth, but continued rapid growth of the feature set. Using a single fold, models within this blue region were reviewed by an expert clinician panel and the model represented by the blue dot, corresponding to 51 features, was selected for further analyses. 95% confidence intervals are shaded around green testing and red training curves. Performance of the pooled cohort equations is drawn as a black line for reference.

• 

  Figure 3. Top 20 predictors selected by ML4HEN-COX and predicted 10-year risk of CAD as a function of GPSCAD, hip circumference, and waist circumference (A) Predictors are ranked by leave-one-out decrease in C-statistic and colored by category (Table 2). (B–D) Ten-year risk of CAD predicted by ML4HEN-COX plotted at ages 45, 55, and 65 years as a function of GPSCAD, hip circumference, and waist circumference, respectively. GPSCAD, genome-wide polygenic score for CAD; HDL-c, HDL cholesterol; SBP, systolic blood pressure; HbA1c, hemoglobin A1C; LDL-c, LDL cholesterol; Hip circ., hip circumference; ApoB, apolipoprotein B; Waist circ., waist circumference; Father heart dz, paternal history of heart disease; Lp(A), lipoprotein a; Sibling heart dz, sibling history of heart disease; PC3, genetic principal component 3; Lipid-lowering, history of taking lipid-lowering medication.

• 

  Figure 4. Comparisons of risk predictions from ML4HEN-COX, FRS, PCE, and QRISK3 (A) Observed 10-year risk of CAD plotted by quintiles of predicted risk for ML4HEN-COX, FRS, PCE, and QRISK3—a steeper gradient is observed with ML4HEN-COX. (B) Correlation plot between 10-year risks of CAD predicted by ML4HEN-COX, SimpleCox51, XGBoost, FRS, PCE, and QRISK3.