Abstract 313: Deep Learning to Assess Cardiovascular Age From Chest Radiographs

Introduction: Chronological age is a well-known risk factor for cardiovascular disease, but measures of vascular age may enable more personalized care. We hypothesize that a convolutional neural network (CNN) can assess vascular age from a chest radiograph image. Methods: The CNN model, CXR-Age, was developed using data from over 100,000 indviduals from publicly available cohorts and was validated in 1) a subset of the Prostate, Lung, Colorectal, and Ovarian Cancer screening trial’s chest x-ray arm (PLCO, N = 40,967) and 2) the chest radiograph arm of the National Lung Screening Trial (NLST, N = 5,414). The primary outcome was 13-year cardiovascular mortality defined by ICD9 codes for ischemic heart disease, myocardial infarction, and stroke. Results are provided for independent testing datasets only. Results: After adjusting for sex, a 5-year increase in CXR-Age was a better predictor of cardiovascular mortality than a 5-year increase in chronological age in PLCO (CXR-Age aHR 2.69 per 5 years [95% CI 2.55-2.84] vs. chronological age aHR 1.84 per 5 years [95% CI 1.75-1.93], p < 0.001) and NLST (CXR-Age aHR 2.06 per 5-years [95% CI, 1.78-2.39] vs. chronological age aHR 1.64 per 5 years [95% CI, 1.44-1.86], p = 0.06). This association with cardiovascular mortality was robust to adjustment for baseline cardiovascular risk factors (chronological age, sex, diabetes, hypertension, smoking) in PLCO (CXR-Age aHR 1.58 per 5 years [95% CI, 1.54-1.63], p < 0.001) and NLST (CXR-Age aHR 1.48 per 5 years [95% CI, 1.36-1.61], p < 0.001). Kaplan-Meier curves (Figure 1) stratified by chronological age groups show CXR-Age has a graded association with cardiovascular mortality in individuals with similar baseline chronological age. Conclusions: A CNN model, CXR-Age, can assess vascular age from a chest radiograph image, and this CXR-Age predicts cardiovascular mortality better than chronological age.