Incremental prognostic value of downstream positron emission tomography perfusion imaging after coronary computed tomography angiography: a study using machine learning

Figures

• 

  Figure 1. Flow diagram of the presented work. After removal of incomplete data entries, data from 2284 patients were retained for further analysis. In the feature selection process, 100 XGBoost models were trained using all the considered input variables, and based on the testing of these models, 9 most important features were selected. These nine variables were considered in training and testing of two sets of 100 XGBoost models (using clinical, CCTA, and PET variables vs. using only clinical and CCTA variables). Finally, we considered the subsets of the test data, restricted to patients who did not undergo early revascularization.

• 

  Figure 2. Mean AUCs and 95% CIs for the XGBoost models trained with all variables, and without PET variables in the test sets randomly selected from the whole cohort. Per each observation year, 100 XGBoost models were trained and tested for each input variable set; an independent and random training and test data split was performed for each of these models. The AUCs are calculated using the test sets (N = 571). The difference between these curves is statistically significant for the first four observation years; this is highlighted by the black asterisks.

• 

  Figure 3. Mean AUCs and 95% CIs for the restricted set of patients who did not undergo early revascularization. The XGBoost models trained with all variables, and without PET variables using the whole cohort data. The AUCs are calculated using the same test sets as in the whole cohort case, restricted to patients who did not undergo early revascularization.

• 

  Figure 4. Mean AUC differences (all variables vs. no PET variables) for XGBoost models trained and tested with data corresponding to all patients, and with data corresponding to patients without early revascularization.