Machine Learning to Identify Patients at Risk of Developing New-Onset Atrial Fibrillation after Coronary Artery Bypass

Figures

• 

  Figure 1. Iterations of hyperparameters search performed with MARS (A), NN (B), RF (C), and SVM (D) models after the last run, respectively. The graphs report the variations of the ROC AUC as a function of one or more parameters of each model. With the MARS model, the maximum ROC AUC value is obtained by degree and term hyperparameters equal to 4 and 5, respectively (A). With the NN model, the maximum ROC AUC value is obtained by layer1, layer2, and layer3 hyperparameters equal to 7, 9, and 6, respectively (B). With the RF model, the maximum ROC AUC value is obtained by predictors hyperparameter numbers equal to 1 (C). With the SVM model, the maximum ROC AUC value is obtained by sigma and cost hyperparameters equal to 0.002 and 2, respectively (D).

• 

  Figure 2. Statistically significant estimated differences of the metrics values with confidence intervals and a vertical line indicating the points with zero difference. The most statistically significant difference in sensitivity and specificity is between the NN and RF and MARS and RF models.

• 

  Figure 3. Relevant features selected by single models: MARS (A). NN (B). RF (C). SVM (D). The relevant features rank has to be interpreted from the top (most important) to the bottom (less important). An acceptable cut-off value is 35%.

• 

  Figure 4. ROC curves by color scale threshold on held-out set. MARS (A). NN (B). RF (C). SVM (D) models. Right side: Color scale for different classification threshold values. The best ROC AUC value is with the RF model.

• 

  Figure 5. Precision-Recall (PR) prediction curve by color scale threshold on held-out set. MARS (A). NN (B). RF (C). SVM (D) models. Right side: Color scale for different classification threshold values. The best PR AUC value is with the RF model.