ML superior in predicting adverse events from CT angiograms

Machine learning models appear better than traditional scoring systems at forecasting from coronary CT angiography whether patients would die or have a heart attack years later.

Coronary CT angiography contains prognostic information to help physicians determine how best to treat patients. However, there are several methods used to interpret the data.

The study authors, from Yale University and elsewhere, hypothesized that machine learning could find a combination of arterial features that better identified the patients who would later experience an adverse event.

“The overall purpose of our study was to use machine learning to improve the contribution of noninvasive imaging to cardiovascular risk assessment,” the researchers said.

They reviewed coronary CT angiographies of 6,892 patients with various indicators of cardiac issues, such as stable atypical chest pain, family history, and indeterminate stress tests. Different vessel features were recorded, including plaque amount, plaque calcification, vessel remodeling and amount of stenosis.

The researchers applied four machine learning model types and compared their predictions against five conventional vessel scores, including the venerable Coronary Artery Disease Reporting and Data System (CAD-RAM) score created by a consensus of medical societies.

They then followed the patients for a median of nine years.

During that time, there were 380 deaths from all causes and 113 cardiac-related events.

All of the machine learning models were better than the traditional coronary CT angiography derived scores in predicting which patients would later have a major adverse event.

For instance, there is extensive evidence that plaque volume and vessel remodeling are important factors and visible on coronary CT angiography images, according to the researchers. However, that data is not included in CAD-RAM scores, which are based only on an analysis of stenosis.

The study also found that the use of the machine learning scores would have ensured that 93 percent of patients that had adverse events would have received preventive care with statins. If CAD-RAM scores were used, only 69 percent would have been so treated.

The study was published in Radiology.

“Vessel scores can be used as inputs into more comprehensive risk models that include non-imaging risk factors such as age, sex, hypertension and smoking. If machine learning can improve vessel scoring, the contribution of noninvasive imaging to cardiovascular risk assessment would also be improved,” the study authors stated.