Deep learning-based prediction for significant coronary artery stenosis on coronary computed tomography angiography in asymptomatic populations

Figure 1. Schematic flowchart of the study population. AUC, area under the curve; ROC, receiver operating characteristic; SHAP, SHapley Additive ExPlanations.

Figure 2. Schematic diagram for the prediction model development. This is a schematic diagram of a newly developed DL-based prediction model for significant coronary artery stenosis on CCTA in 11,180 asymptomatic populations who underwent a routine health check-up. Using various parameters that physicians can readily access in clinical practice, the DL-based model could provide more precise indications of CCTA for the purpose of screening. CCTA, coronary computed tomography angiography; CHD, coronary heart disease; DL, deep learning.

Figure 3. Increase of predictive value upon forward feature selection. Features in Y-axis are arranged by forward selection in order. The bar graph, generated from the validation dataset, shows how performance gains when variables are added to the model, on a logarithmic scale. Error bar shows its 95% confidence interval. Scatter dot plot shows an increase in ROC-AUC of the test dataset when using top-k features, indicating that performance saturates shortly after adding a few variables to the model. CVD, cardiovascular disease; ECG, electrocardiography; eGFR, estimated glomerular filtration rate; HbA1, glycated hemoglobin; HDL, high-density lipoprotein; OH, hydroxy; SBP, systolic blood pressure; WBC, white blood cells; other abbreviations as Figure 1.

Figure 4. Outcome distribution of individuals with obstructive and non-obstructive CHD. The X-axis represents the probability of coronary artery stenosis, and the Y-axis represents the number of individuals in the test set. The blue and orange bars indicate individuals predicted to have coronary artery stenosis of <30% and ≥70%, respectively.

Figure 5. Comparison of predictive performances among a neural network-based model, other DL-based models, and clinical risk scoring models with selected features. Both plots show the ROC-AUC of various models with selected features. (A) Our neural network-based prediction model demonstrates significantly better performance than previously known clinical scoring systems, with ROC-AUC of 0.782, even in asymptomatic populations. (B) Among the machine learning-based models, a neural network with multi-task shows the best performance to predict significant coronary artery stenosis on CCTA. CAD, coronary artery disease; DL, deep learning; XGBoost, eXtreme gradient boosting; other abbreviations as Figures 1, 2.

Figure 6. SHAP value of a neural network-based model. The bar graph shows how each feature contributes to the model decision in test set. As the contribution of each input feature is described as an absolute value, the sign of the effect of each variable is not represented. Abbreviations as Figures 1, 3.