Pre-test probability for coronary artery disease in patients with chest pain based on machine learning techniques

Abstract

Background: A correct and prompt diagnosis of coronary artery disease (CAD) is a crucial component of disease management to reduce the risk of death and improve the quality of life in patients with CAD. Currently, the American College of Cardiology (ACC)/American Heart Association (AHA) and the European Society of Cardiology (ESC) guidelines recommend selecting an appropriate pre-diagnosis test for an individual patient according to the CAD probability. The purpose of this study was to develop a practical pre-test probability (PTP) for obstructive CAD in patients with chest pain using machine learning (ML); also, the performance of ML-PTP for CAD is compared to the final result of coronary angiography (CAG). Methods: We used a database from a single-center, prospective, all-comer registry designed to reflect real-world practice since 2004. All subjects underwent invasive CAG at Korea University Guro Hospital in Seoul, South Korea. We used logistic regression algorithms, random forest (RF), supporting vector machine, and K-nearest neighbor classification for the ML models. The dataset was divided into two consecutive sets according to the registration period to validate the ML models. ML training for PTP and internal validation used the first dataset registered between 2004 and 2012 (8631 patients). The second dataset registered between 2013 and 2014 (1546 patients) was used for external validation. The primary endpoint was obstructive CAD. Obstructive CAD was defined as having a stenosis diameter of >70% on the quantitative CAG of the main epicardial coronary artery. Results: We derived an ML-based model consisting of three different models according to the subject used to obtain the information, such as the patient himself (dataset 1), the community's first medical center (dataset 2), and doctors (dataset 3). The performance range of the ML-PTP models as the non-invasive test had C-statistics of 0.795 to 0.984 compared to the result of invasive testing via CAG in patients with chest pain. The training ML-PTP models were adjusted to have 99% sensitivity for CAD so as not to miss actual CAD patients. In the testing dataset, the best accuracy of the ML-PTP model was 45.7% using dataset 1, 47.2% using dataset 2, and 92.8% using dataset 3 and the RF algorithm. The CAD prediction sensitivity was 99.0%, 99.0%, and 98.0%, respectively. Conclusion: We successfully developed a high-performance model of ML-PTP for CAD which is expected to reduce the need for non-invasive tests in chest pain. However, since this PTP model is derived from data of a single medical center, multicenter verification is required to use it as a PTP recommended by the major American societies and the ESC.