A deep learning model developed to estimate the risk of atherosclerotic cardiovascular disease (ASCVD) from chest CT may offer information, beyond traditional risk factors and coronary artery calcium score (CAC), that can aid in predicting cardiovascular events.
For patients at intermediate risk of heart attack and stroke, the CAC score derived using CT imaging is a strong predictor of future cardiovascular events. Treatment decisions depend on accurate risk estimation.
"Our vision is to build an opportunistic screening tool that can automatically sift through existing CT scans and identify patients at high-risk for heart attack and stroke that might benefit from preventive interventions, like statins to lower cholesterol," said Vineet K. Raghu, PhD, an instructor and postdoctoral research fellow at the Cardiovascular Imaging Research Center of Massachusetts General Hospital and Harvard Medical School in Boston. "We wanted to investigate whether there's more information on these CT scans that can be useful to predict future risk."
Dr. Raghu and colleagues developed a deep learning model, referred to as CT-CV-Risk, using chest CT data from 51,182 chest CTs of 12,433 smokers in the National Lung Screening Trial (NLST). They trained the model to predict the probability of cardiovascular mortality over 12 years.
The researchers then conducted independent testing to determine 12-year cardiovascular mortality for a hold-out group of 7,405 individuals who were eligible for primary prevention measures because they showed no history of Type 2 diabetes, myocardial infarction or stroke.
New Model Delivers Improved Results
The performance of the CT-CV-Risk model was compared to a baseline regression model using demographics, smoking history, BMI, comorbidities and CT imaging findings such as lung nodules, emphysema and CAC.
"In the testing dataset, the CT-CV-Risk model demonstrated improved discrimination for cardiovascular mortality compared to the baseline model. Similar improvements were observed for fatal myocardial infarction and stroke," Dr. Raghu said.
He added that in a subset of patients with CAC, the CT-CV-Risk tool successfully predicted 10-year cardiovascular mortality beyond prevalent risk factors, CAC and imaging findings.
With ongoing work focused on building AI tools to extract individual organs from the full scan, and developing AI models that analyze 3D images of each organ, Dr. Raghu acknowledged that a major challenge in developing AI tools to analyze CT data is dealing with the sheer size of each CT scan.
"We simplified the problem by converting the 3D scan into a 2D image projection similar to a chest X-ray," he said. "Though this loses a lot of valuable 3D information, we actually found that there is information on this simple X-ray-style image that predicts cardiovascular events beyond the CAC score and other known risk factors."
Dr. Raghu emphasized the importance of using high-quality training data in the development of deep learning tools like CT-CV-Risk.
"When we first attempted this project, only a portion of the National Lung Screening Trial CTs were available for researchers. Since AI models require a large amount of data for development, our initial attempts to predict risk from this smaller dataset were not very successful," Dr. Raghu said. "Recently, the National Cancer Institute has made all CTs conducted in the trial available to researchers, which is really what has enabled this work to be successful."
Access the presentation, "Cardiovascular Disease Risk Prediction Beyond Coronary Artery Calcium Using Deep Learning Applied at Chest Computed Tomography," (T7-SSCAC6-1) on demand at Meeting.RSNA.org