Abstract 4141180: Opportunistic assessment of cardiovascular risk using deep learning of the heart and aorta on non-contrast chest computed tomography

Purpose: Primary prevention of cardiovascular (CV) disease relies on accurate risk estimation. The coronary artery calcium score (CAC) can be measured on routine chest computed tomography (CT); however, it is unclear whether the heart and thoracic aorta regions on CT can provide additional information about future risk beyond calcium and risk factors. Methods: We applied the open-source TotalSegmentator tool to extract heart and thoracic aorta volumes from 27,943 non-contrast non-ECG gated chest CTs from 17,241 National Lung Screening Trial participants. We used the segmented volumes from 10,356 participants as input to train a new 3D Densenet169 deep learning model to predict 12-year CV mortality. Independent testing was performed in 5,165 individuals not used for model training with no history of type 2 diabetes, myocardial infarction, or stroke (eligible for primary prevention). The deep learning score was compared to a regression model using demographics, smoking, BMI, comorbidities, and the radiologist’s findings from the CT. We stratified the CAC and deep learning scores into high vs. low-risk groups using a CAC>100 and deep learning >10% risk threshold, respectively. Results: In the independent testing dataset, (N=5165; median age 60.0±8.0 years; 62.6% male), 3.3% died of CV disease. The deep learning score had higher discrimination for CV mortality than risk factors and CT findings (deep learning AUC 0.72 [0.68,0.76] vs. baseline AUC 0.66 [0.61,0.69], p < 0.001). Similar results were found for secondary outcomes of fatal myocardial infarction and fatal stroke. In a subset of 2,972 testing set individuals with measured CAC, those with high CAC and a high deep learning score had a 10.8% rate of 12-year CV mortality, higher than those with high CAC alone (4.1%, p<0.001). Conclusion: Based on the heart and thoracic aorta regions from CT, a deep learning model predicted 12-year CV mortality beyond prevalent risk factors, CAC, and imaging findings. Clinical Relevance: Opportunistic screening of chest CT using deep learning may enhance risk assessment to guide decisions for preventing CV events. External validation in population-representative cohorts is needed before clinical use.