Predictive value of CAC score combined with clinical features for obstructive coronary heart disease on coronary computed tomography angiography: a machine learning method.

Objective: We investigated the predictive value of clinical factors combined with coronary artery calcium (CAC) score based on a machine learning method for obstructive coronary heart disease (CAD) on coronary computed tomography angiography (CCTA) in individuals with atypical chest pain.

Methods: The study included data from 1,906 individuals undergoing CCTA and CAC scanning because of atypical chest pain and without evidence for the previous CAD. A total of 63 variables including traditional cardiovascular risk factors, CAC score, laboratory results, and imaging parameters were used to build the Random forests (RF) model. Among all the participants, 70% were randomly selected to train the models on which fivefold cross-validation was done and the remaining 30% were regarded as a validation set. The prediction performance of the RF model was compared with two traditional logistic regression (LR) models.

Results: The incidence of obstructive CAD was 16.4%. The area under the receiver operator characteristic (ROC) for obstructive CAD of the RF model was 0.841 (95% CI 0.820-0.860), the CACS model was 0.746 (95% CI 0.722-0.769), and the clinical model was 0.810 (95% CI 0.788-0.831). The RF model was significantly superior to the other two models (p < 0.05). Furthermore, the calibration curve and Hosmer-Lemeshow test showed that the RF model had good classification performance (p = 0.556). CAC score, age, glucose, homocysteine, and neutrophil were the top five important variables in the RF model.

Conclusion: RF model was superior to the traditional models in the prediction of obstructive CAD. In clinical practice, the RF model may improve risk stratification and optimize individual management.