who presented the results of their deep-learning model at the annual meeting of the Radiological Society of North America.
Current American College of Cardiologists and American Heart Association guidelines recommend estimating 10-year risk of major adverse cardiovascular events (MACE) to determine whether a patient should receive statins to help prevent atherosclerotic cardiovascular disease (ASCVD). Statins are recommended for patients with a 10-year risk of 7.5% or higher, the authors noted.
The current ASCVD risk score is determined with nine factors: age, sex, race, systolic blood pressure, hypertension treatment, smoking, type 2 diabetes, and a lipid panel.
Not all data points available in EHR
But not all of those data points may be available through the electronic health record, “which makes novel and easier approaches for population-wide screening desirable,” said lead researcher Jakob Weiss, MD, a radiologist affiliated with the Cardiovascular Imaging Research Center at Massachusetts General Hospital and the AI in medicine program at the Brigham and Women’s Hospital in Boston.
Chest x-ray images, on the other hand, are commonly available. The images carry rich information beyond diagnostic data but have not been used in this type of prediction model because AI models have been lacking, Dr. Weiss said.
The researchers trained a deep-learning model with single chest x-rays only.
They used 147,497 chest x-rays from 40,643 participants in the Prostate, Lung, Colorectal, and Ovarian Cancer (PLCO) Screening Trial, a multicenter, randomized controlled trial designed and sponsored by the National Cancer Institute.
Dr. Weiss acknowledged that the population used to train the model was heavily White and that should be a consideration in validating the model.
They compared their model’s ability to predict 10-year ASCVD risk with the standard ACC/AHA model.
“Based on a single chest radiograph image, deep learning can predict the risk of future cardiovascular events independent of cardiovascular risk factors and with similar performance to the established and guideline-recommended ASCVD risk score,” Dr. Weiss said.
Tested against independent group
They tested the model against an independent group of 11,430 outpatients (average age, 60 years; 42.9% male) who underwent a routine outpatient chest x-ray at Mass General Brigham and were potentially eligible to receive statins.
Of those 11,430 patients, 1,096 (9.6%) had a major adverse cardiac event over the median follow-up of 10.3 years.
There was a significant association of CXR-CVD risk and MACE among patients eligible to receive statins, the researchers found (hazard ratio, 2.03; 95% confidence interval, 1.81-2.30; P < .001), which remained significant after adjusting for cardiovascular risk factors (adjusted HR, 1.63; 95% CI, 1.43-1.86; P < .001).
Some of the variables were missing in the standard model, but in a subgroup of 2,401 patients, all the variables were available.
They calculated ASCVD risk in that subgroup using the standard model and the CXR model and found that the performance was similar (c-statistic, 0.64 vs. 0.65; P = .48) to the ASCVD risk score (aHR, 1.58; 95% CI, 1.20-2.09; P = .001).
Ritu R. Gill MD, MPH, associate professor of radiology at Harvard Medical School in Boston, who was not part of the study, said in an interview that “the predictive algorithm is promising and potentially translatable and could enhance the annual medical checkup in a select population.
“The algorithm was developed using the PLCO cohort with radiographs, which are likely subjects in the lung cancer screening arm,” she said. “This cohort would be at high risk of cardiovascular diseases, as smoking is a known risk factor for atherosclerotic disease, and therefore the results are expected.
“The algorithm needs to be validated in an independent database with inclusion of subjects with younger age groups and adjusted for gender and racial diversity,” Gill said.
David Cho, MD, a cardiologist at the University of California, Los Angeles, who also was not part of the study, said in an interview that “this work is a great example of AI being able to detect clinically relevant outcomes with a widely used and low-cost screening test.
“The volume of data needed to train these models is already out there,” Dr. Cho said. “It just needs to be mined.”
He noted that this tool, if validated in randomized trials, could help determine risk among patients living in places where access to specialized cardiac care is limited.
Dr. Weiss and Dr. Cho disclosed no relevant financial relationships. Dr. Gill has received research support from Cannon Inc and consultant fees from Imbio and WorldCare.
A version of this article first appeared on Medscape.com.