Abstract 11189: An Integrated Framework With Machine Learning and Radiomics Score for the Prediction of Coronary Artery Calcium Score From Chest Radiographs

Abstract

Introduction: Recently introduced radiomics technology enables the detection of hidden information from commonly disregarded medical images, including CXR. Hypothesis: The purpose of this study was to evaluate the feasibility of a novel technique involving radiomics features combined with machine learning to identify moderate-to-severe coronary artery calcium (CAC) using simple chest X-ray radiography (CXR). Methods: We included 559 patients (women, 44.9%; mean age, 62.4 ± 9.4 years) from two independent clinical studies who underwent a calcium scan and CXR within 6 months. The total cohort was allocated to the training and validation cohorts in a 7:3 ratio, with all clinical characteristics well-matched, including the CAC score. Radiomics features were extracted from manually delineated cardiac contours, and a radiomics score formulation for the prediction of a CAC score ≥100 was generated using the machine learning method in the training cohort. To evaluate the incremental performance of the radiomics score, a basic clinical model including age, sex, and body mass index (model 1) and a radiomics score added model (model 2) were utilized. Results: The radiomics score was the most prominent predictive factor for CAC score ≥100 (odds ratio [OR] = 2.33; 95% confidence interval [CI] = 1.62-3.44; p < 0.001). In the training cohort, model 2 demonstrated significant incremental validity in predicting CAC scores ≥100 compared to model 1 (area under the curve [AUC]; 0.73 vs. 0.69, p = 0.022). The performance of model 2 was also similar in both the training and validation cohorts (AUC 0.73, 95% confidence interval [CI] 0.68 - 0.78 vs. AUC 0.72, 95% CI 0.64 - 0.80). Conclusions: We developed a machine learning-based radiomics scoring model that could be utilized as a potential imaging marker for predicting CAC scores from CXR. This novel method may be widely applicable to clinical practice and can improve the pre-test probability of coronary artery disease.