AI Model Boosts Cardiac Event Prediction

In a large retrospective study conducted at Emory University, an interpretable machine learning (ML) model combining CT and clinical variables achieved high accuracy in predicting major adverse cardiovascular events (MACE).

By combining all the clinical risk factors for heart disease, as identified by the American Heart Association, the model significantly outperformed traditional risk scores and coronary artery calcium (CAC) scores, according to Xinyue (Fay) Yan, a student at Northwestern University in Chicago, who shared the results Wednesday morning.

“Clinically, single-risk source prediction may miss signals, so combining clinical and imaging variables will help to better predict major adverse cardiac events,” she said. “Our goal was to determine which input contributes most heavily to predicting events and how they relate to each other.”

Yan’s research team used patient data from Emory University from multiple facilities covering the metro Atlanta area. Nearly 8,000 asymptomatic individuals, averaging 58 years of age (45% female and 12% Black) who received a non-contrast cardiac CT for calcium scoring between 2010 to 2023 were included in the study cohort. Patients with known cardiac disease were excluded.

“Using an interpretable model allows us to see how each measure contributes,” explained research team member Marly Van Assen, MSc, PhD, before the meeting. “It gives us insight into further developing more advanced AI models.”

A total of 52 clinical variables and risk factors such as high cholesterol and hypertension, lab values, EKG metrics and PCE/PREVENT risk scores were extracted from medical records within six months of the CT. 

The researchers also quantified all CT imaging variables, including CAC score and volume, vessel involvement and cardiac chamber volume. They collected MACE outcomes, including MI, stroke and revascularization at 90 days.

Model Outperforms Traditional Risk Scores

The ML model’s accuracy in predicting MACE was compared with that of clinical risk score models and CAC scores. The AI model achieved an AUC of .861 for predicting MACE, outperforming PCE and PREVENT (AUC .627 and .630, respectively) and CAC score alone (AUC .628).

The PREVENT and PCE risk scores and CAC scores had the highest gains in our model,” Yan said. “Imaging features such as total CAC volume, myocardial volume and right arterial volume were also major contributors in the model.”

The most prevalent MACE predictor in the study cohort was high cholesterol, present in 80% of patients. 

“By combining CT imaging and clinical variables, our machine learning model showed good accuracy in predicting MACE,” Yan said. “Our findings can be used to improve risk prediction and to develop more advanced AI models.”

Dr. Van Assen said there’s much more work to be done to achieve an acceptable level of accuracy while maintaining interpretability. “I think we have a good baseline from which we can refine, evaluating whether each risk factor needs further investigation and adding other factors we don’t use clinically, such as right atrial and left ventricular volumes.”

The research team is partnering with collaborators from the Georgia Institute of Technology in Atlanta to build a more advanced AI model. “We’re trying to remove the co-linearity to make the model simpler,” Dr. Van Assen explained.

Access the presentation, “Predicting Adverse Cardiac Events Using Clinical and Imaging Features with an Interpretable Machine Learning Model,” (W3-SSCA07-5) on demand at RSNA.org/MeetingCentral.