A Machine Learning Model Utilizing a Novel SNP Shows Enhanced Prediction of Coronary Artery Disease Severity

Figures

• 

  Figure 1. Percentages of subjects with hypertension, diabetes, and obesity in Gensini score low and high groups. Subjects presented at the cardiac catherization clinic were observed to have high prevalence of hypertension (78.18%, with 53.45% in low and 24.73% in high Gensini score severity level), diabetes (35.28 with 20.73% in low and 14.55% in high Gensini score severity level) and obesity (52.37%, with 36.73% in low and 15.64% in high Gensini score severity level).

• 

  Figure 2. PCA and correlation matrix of patients with high and low severity score generated using conventional risk factors and Id3 SNP. PCA plots between PC1 and PC2-PC10 (A) and low fraction of variance captured by each principal component (PC) (B) suggested that patients with high and low severity score cannot be differentiated by using PCA based on their conventional risk factors and Id3 SNPs. Correlation matrix also demonstrates low-intermediate correlation between severity score and conventional risk factors and Id3 SNP (C). Both PCA and correlation matrix analysis indicates a challenge in developing a model/strategy to classify patients into high and low risk severity score.

• 

  Figure 3. Performance of high and low risk FRS in estimating high and low risk severity score. Framingham Risk Score (FRS) can accurately predict low risk group with 65% accuracy and high risk group with 66% accuracy (A). With addition of ID3 SNP to FRS, the model predicts low risk group with 66% accuracy and high risk group with 62% accuracy (B).

• 

  Figure 4. Structure and performance of our best machine learning based model in predicting high and low risk severity score evaluated by confusion matrix and AUC of ROC. The best machine learning model with the highest performance is a deep neural network model with 12 neurons in the input layer including age, total cholesterol, triglyceride, low-density lipoprotein (LDL), high-density lipoprotein (HDL), systolic blood pressure, HbA1c, glucose, sex and Id3SNP after all these inputs were normalized and transformed. The model contained 12 and 6 neurons in the next sequential layers and an output layer (A). The model performance was evaluated on 82 patients in the testing set. With our best model, the low risk group was predicted NPV of 81% and high risk group with PPV of 87% (B). ROC plot showing the performance of the model (blue line) with an AUC of 0.84 and random (dashed red line) for comparison (C).

• 

  Figure 5. Elimination of Id3 SNP as input variable reduces model performance. Performance of the model drops when Id3 SNP was removed. Positive predictive value (PPV) of predicting high risk group drops from 87% to 80% and NPV of predicting low risk group drops from 81% to 65%. AUC of ROC also drops from 0.84 to 0.72.

• 

  Figure 6. Importance score of each input feature demonstrated with different color boxplot (with square indicating outliers) on our best machine learning model. Importance score, defined as cross-validation score for model with actual values minus that using permuted values, for each of the 9 most important features where values were permuted 5 times for each feature. Age, total cholesterol, triglyceride, HDL, systolic blood pressure, HbA1c, glucose, sex, and Id3SNP are the 9 most important features out of 13 features chosen as inputs of our model input features. These 9 inputs contribute at different levels in enhancing precision and accuracy of the model with systolic blood pressure, HDL and age being the top 3 most important inputs.