Cardiovascular Disease (CVD), including heart attacks and strokes, is the most common cause of death in the world. Although strategies to prevent cardiovascular disease are available, such as a healthier lifestyle and statin treatment, the use of which may not always be optimal, identification of risk may be difficult. A new study led by the Cardiovascular Imaging Research Core (CIRC) at Massachusetts General Hospital (MGH) describes how an AI model, utilizing a routine imaging exam, can be used to rapidly and non-invasively screen for CVD risk factors.

Challenge In Current Cardiovascular Disease Detection 

The traditional tools used to predict cardiovascular disease risk, such as the Pooled Cohort Equations, which are endorsed by the American College of Cardiology and the American Heart Association, rely on several inputs — including lipid profiles, blood pressure readings, a history of diabetes, and more — many of which are missing or degraded in data from electronic medical records. This not only hinders systematic identification of high-risk patients for pMPD prophylaxis, but it may also lead to unintended consequences such as increased cost and incidence of adverse effects.

Objective

To address these barriers, the MGH team created a deep learning (Artificial Intelligence) AI Model — called CXR-CVD. It is used to estimate cardiovascular risk using only chest X-ray images, which are far more common and cost-effective compared to clinical risk–scoring tools.

Solution & Methodology

The scientists trained the model on hundreds of thousands of already available chest radiographs by applying a Convolutional Neural Network (CNN). The AI model taught itself how to recognize patterns indicative of longer-term cardiovascular risk, without the help of conventional clinical inputs. The group’s results were shared at the Radiological Society of North America (RSNA) meeting and received national recognition via the RSNA daily bulletin and CNN.

Key Findings

• The AI model for predicting 10-year cardiovascular risk was similar in accuracy to the Pooled Cohort Equation.
• The model could accurately predict the responders and non-responders of statin therapy, even in the absence of classical data.
• There are hidden markers of aging and cardiovascular risks in chest radiographs that AI can capture.

Innovation & Novelty

This study is an unconventional use of AI, taking ordinary diagnostic images to predict future health events. While gated scans are needed in CT calcium scoring, standard chest radiographs are used in the current model, greatly expanding the possibility of broadly applying this approach.

Extension: AI-CAC for CT Scans

Expanding on this idea, researchers from Mass General Brigham have also created AI-CAC, a deep learning algorithm that can detect the presence of coronary artery calcium (CAC) on nongated chest CT scans. The tool achieved:

• 89.4% sensitivity for the presence of CAC
• 87.3% for CAC score above or below the clinical threshold of 100
•  There was a 3.49X increased 10-year mortality with scores > 400, marked predictive value.

Clinical Impact

Both AI tools, CXR-CVD Risk and AI-CAC, allow for a shift from reactive care to preventative health based on these new opportunities. Through the use of available imaging data, healthcare systems may more conveniently identify high-risk patients for interventions that may reduce morbidity, mortality, and long-term healthcare expenses.

Limitations & Future Scope

AI-CAC was trained using veterans only. Subsequent work will aim to confirm the model in outside populations and evaluate its capacity to serially monitor CAC progression, especially concerning lipid-lowering therapy.

Conclusion

This groundbreaking effort by the CIRC and Mass General Brigham demonstrates the potential AI has to change the game in the early detection of cardiovascular diseases. These advances preserve, but extract additional value from, standard imaging examinations, and can enable earlier interventions and improved patient outcomes, with a concomitant reduction in the workload on, and resource requirements of, the clinic.