Algorithms have been developed in recent years to automatically derive coronary artery calcium (CAC) scores from cardiac CT scans to determine patient risk for major adverse cardiovascular events. Now, new research demonstrates that deep learning-based algorithms can automatically derive coronary artery calcium (CAC) scores from thoracic CT scans previously performed for other reasons, including lung cancer screening. Automatically calculating CAC scores on routine chest CT scans using deep learning could facilitate opportunistic screening and allow for earlier prevention and management of patients with cardiovascular disease. Informing clinicians and their patients when significant incidental findings are detected—even if these scans occurred in the past—could lead to important conversations with patients about their previously undiagnosed heart disease.

Coronary Artery Calcium Scoring in Clinical Practice

Coronary artery calcium scoring is widely recognized as one of the strongest predictors of cardiovascular risk in patient populations [1]. Clinical trials, including the Multi-Ethnic Study of Atherosclerosis, have demonstrated the value of CAC scoring in reliably measuring patient risk for adverse cardiovascular events. As a result, societies including the American College of Cardiology (ACC), Society for Cardiovascular Computed Tomography (SCCT), and American Heart Association (AHA) have published guidelines recommending CAC scoring to be used in guiding primary prevention of atherosclerotic cardiovascular disease.

According to guidelines published by the ACC and the AHA in 2018, a calcium score of zero indicates that patients are at very low risk of developing heart disease and are therefore unlikely to benefit from statin therapy, while a calcium score above 100 indicates that a statin should be considered. Additionally, the 2019 ACC/AHA Guidelines on the Primary Prevention of Cardiovascular Disease state that risk-enhancing factors, such as CAC scanning, can help guide decisions about preventive interventions in select individuals and can be a reasonable tool to reclassify risk either upward or downward as part of shared decision-making.

Despite these recommendations, CAC scoring CT has not yet become the standard of care. Currently, CAC scoring CT is not reimbursable by Medicare and other U.S. insurance companies. Additionally, as stated by Dr. Michael W. Vannier of the University of Chicago in a 2019 editorial published in Radiology, quantifying CAC scores on routine chest CT exams would be time-consuming for the clinician, and the extra work would not be reimbursable. Therefore, the automatic calculation of CAC scoring would be necessary for it to become routine [2]. This presents an opportunity for deep learning and other advanced informatics to automatically derive CAC scores from chest CT scans.

Quantification of Coronary Artery Calcium on Routine Non-gated Chest CT

Millions of patients undergo routine chest CT scans each year, with this number continuing to rise annually. This growing number of chest CT scans presents an opportunity for AI to retrospectively screen patients for coronary calcification without patients having to undergo additional tests.

In a recent study, researchers developed a deep learning model to automate CAC scoring for dedicated gated coronary CT exams and routine non-gated chest CT scans performed for other reasons to allow for opportunistic screening. In the study published in NPJ Digital Medicine, the non-gated deep learning model was validated internally using 42 chest CT scans from patients at Stanford and 46 chest CT scans from patients in MESA.

The study results demonstrated that the deep learning model had an almost perfect agreement with the manually derived CAC scores on the Stanford dataset, and a moderate agreement with the MESA dataset. The non-gated deep learning model demonstrated a sensitivity of 100% and a positive predictive value (PPV) of 100% on the Stanford dataset [3]. Sensitivity was lower in the MESA dataset as the deep learning model demonstrated a sensitivity of 85% and a PPV of 100%.

This study reveals the potential that applying deep learning to pre-existing data can have in unlocking additional data from a CT scan that can help to inform cardiovascular disease prevention.

Automated Coronary Artery Calcium Scoring on Lung Cancer Screening CT

Additional research published in Nature Communications demonstrates the value of using AI algorithms to concurrently derive CAC scores from a single lung cancer screening CT.

Researchers developed a deep learning model that could automatically calculate CAC scores from cardiac CT scans and chest CT scans in under two seconds without human input. In the publication titled, “Deep Convolutional Neural Networks to Predict Cardiovascular Risk from Computed Tomography,” researchers evaluated 20,084 patients, including asymptomatic patients from the Framingham Heart Study (FHS) and the National Lung Screening Trial (NLST), as well as patients with stable and acute chest pain from the Prospective Multicenter Imaging Study for Evaluation of Chest Pain (PROMISE) and the Rule Out Myocardial Infarction Using Computer Assisted Tomography (ROMICAT II) trials.

To evaluate the value of coronary calcium in heavy smokers who underwent lung cancer screening CT, the researchers applied the deep learning system to 14,959 patients who had undergone low-dose chest CT from the NLST database. According to the publication, the NLST low-dose chest CT was performed at 33 institutions with various CT scanners using a non-ECG-gated low-dose chest CT protocol. Study results demonstrate that the automated score is a strong predictor of cardiovascular events, showing a high correlation with manual quantification [4].

By automatically deriving CAC scores from previously performed thoracic CT scans, AI can help discovered incidental findings that can better inform previously undiagnosed cases of heart disease without requiring patients to undergo additional testing.

Opportunistic Screening of Cardiovascular Disease: Creating Value via the Network Effect

The economics of population health and value-based care is based on the network effect of small savings aggregated across thousands of individuals being realized as major savings over time. A testable hypothesis is that the results of earlier coronary artery disease treatment could improve the bottom line for healthcare institutions bearing financial risk for patient outcomes, and for those who want to demonstrate their value to self-insured employers. Leveraging AI may unlock valuable information currently locked in routine chest CT scans to inform cardiovascular disease population health.

At Keya Medical, we are developing deep learning-based solutions that will deliver value throughout the care continuum, from early screening and diagnosis to treatment and post-treatment maintenance. We invite healthcare professionals to collaborate with us on future projects to improve cardiovascular disease population health. Contact us to learn more about collaboration opportunities.

[1] Thanassoulis G, Peloso GM, Pencina MJ, Hoffmann U, Fox CS, Cupples LA, Levy D, D’Agostino RB, Hwang SJ, O’Donnell CJ. A genetic risk score is associated with incident cardiovascular disease and coronary artery calcium: the Framingham Heart Study. Circ Cardiovasc Genet. 2012 Feb 1;5(1):113-21. doi: 10.1161/CIRCGENETICS.111.961342. Epub 2012 Jan 10. PMID: 22235037; PMCID: PMC3292865.

[2] Vannier MW. Automated Coronary Artery Calcium Scoring for Chest CT Scans. Radiology. 2020 Apr;295(1):80-81. doi: 10.1148/radiol.2020192718. Epub 2020 Feb 11. PMID: 32053061.

[3] Eng D, Chute C, Khandwala N, Rajpurkar P, Long J, Shleifer S, Khalaf MH, Sandhu AT, Rodriguez F, Maron DJ, Seyyedi S, Marin D, Golub I, Budoff M, Kitamura F, Takahashi MS, Filice RW, Shah R, Mongan J, Kallianos K, Langlotz CP, Lungren MP, Ng AY, Patel BN. Automated coronary calcium scoring using deep learning with multicenter external validation. NPJ Digit Med. 2021 Jun 1;4(1):88. doi: 10.1038/s41746-021-00460-1. PMID: 34075194; PMCID: PMC8169744.

[4] Zeleznik, Roman & Foldyna, Borek & Eslami, Parastou & Weiss, Jakob & Alexander, Ivanov & Taron, Jana & Parmar, Chintan & Alvi, Raza & Banerji, Dahlia & Uno, Mio & Kikuchi, Yasuka & Karady, Julia & Zhang, Lili & Scholtz, Jan-Erik & Mayrhofer, Thomas & Lyass, Asya & Mahoney, Taylor & Massaro, Joseph & Vasan, Ramachandran & Aerts, Hugo. (2021). Deep convolutional neural networks to predict cardiovascular risk from computed tomography. Nature Communications. 12. 10.1038/s41467-021-20966-2.