Long-term Major Adverse Cardiac Event Prediction by Computed Tomography-derived Plaque Measures and Clinical Parameters Using Machine Learning.

ABSTRACT

Objectives The present study evaluated the usefulness of machine learning (ML) models with the coronary artery calcification score (CACS) and clinical parameters for predicting major adverse cardiac events (MACEs). Methods The Nationwide Gender-specific Atherosclerosis Determinants Estimation and Ischemic Cardiovascular Disease Prospective Cohort study (NADESICO) of 1,187 patients with suspected coronary artery disease (CAD) 50-74 years old was used to build a MACE prediction model. The ML random forest (RF) model was compared with a logistic regression analysis. The performance of the ML model was evaluated using the area under the curve (AUC) with the 95% confidence interval (CI). Results Among 1,178 patients from the NADESICO dataset, MACEs occurred in 103 (8.7%) patients during a median follow-up of 4.4 years. The AUC of the RF model for MACE prediction was 0.781 (95% CI 0.670-0.870), which was significantly higher than that of the conventional logistic regression model [AUC, 0.750 (95% CI, 0.651-0.839)]. The important features in the RF model were coronary artery stenosis (CAS) at any site, CAS in the left anterior descending branch, HbA1c level, CAS in the right coronary artery, and sex. In the external validation cohort, the model accuracy of ensemble ML-RF models that were trained on and tuned using the NADESICO dataset was not similar [AUC 0.635 (95% CI 0.599-0.672)]. Conclusion The ML-RF model improved the long-term prediction of MACEs compared to the logistic regression model. However, the selected variables in the internal dataset were not highly predictive of the external dataset. Further investigations are required to validate the usefulness of this model.