AI Model Predicts Cardiac Risk from Routine Chest CT with Greater Accuracy and Fairness

A new artificial intelligence model is showing strong potential for predicting major cardiovascular events using nothing more than routine, non-contrast chest CT scans—opening the door for earlier, more equitable heart disease detection across diverse patient populations.

Presented Thursday by Amara Tariq, PhD, a data science analyst at Mayo Clinic in Scottsdale, Arizona, the study introduces a 3D convolutional neural network (CNN) that estimates the risk of serious outcomes such as heart attacks and cardioembolic strokes directly from imaging, rather than relying solely on traditional risk scores based on demographic and clinical data.

“Traditional risk models for major adverse cardiovascular events like heart attacks and cardioembolic strokes do not incorporate radiologic imaging directly,” Dr. Tariq explained. “These models do not achieve optimal performance when applied to external institutions outside of the clinical and demographic distributions that they were trained on.”

To develop a more generalizable and equitable model, Dr. Tariq’s team—including collaborators from Emory University—trained their deep learning model using over 4,400 CTs from the Mayo Clinic and validated it externally on a set of 201 CTs from Emory, representing a population with notably different demographic characteristics. The model was further enhanced through the use of causal intervention, an approach that strengthens true cause-and-effect relationships while minimizing biased correlations between socioeconomic factors and health outcomes.

“Our goal was to quantify the risk of cardiovascular disease in a way that is both widely applicable and has fair performance for all demographic subgroups,” Dr. Tariq said.

In internal testing, the 3D CNN—with and without causal intervention—outperformed a standard machine learning model that relied on CT-derived body composition and demographic data. However, when tested on the external Emory dataset, only the version with causal intervention maintained strong performance—highlighting the technique’s value in making AI tools more adaptable and fair.

“The initial results are very promising,” Dr. Tariq said. “Our model is not only more accurate than comparative models but is also more equitable and fairer.”

Importantly, the model works using standard non-contrast chest CT scans, which are commonly performed for reasons unrelated to cardiac care—such as lung cancer screening or follow-up of incidental findings. This could allow for opportunistic cardiovascular risk assessment without any additional imaging, contrast injection, or cost.

“This not only provides accurate protection against future cardiovascular events,” Dr. Tariq added, “it also delivers fair and equitable performance across different subgroups.”

The team sees the model as a step toward broader, more inclusive cardiac screening, especially in patients who may not undergo dedicated cardiac imaging. It also provides a path to integrating cardiovascular risk evaluation into existing radiology workflows.

Next, the researchers plan to expand the project by combining multiple imaging modalities with clinical data to build a more comprehensive risk assessment framework.

“This is just the beginning,” Dr. Tariq said. “We’re working to build multimodal foundational models that will further improve accuracy and better reflect the complexity of cardiovascular risk in real-world patient populations.”