In the Journals

AI with CTA may predict adverse events vs. other prediction scores

U. Joseph Schoepf
U. Joseph Schoepf

Coronary CTA assessed by machine learning was able to better distinguish patients who had an adverse event compared with the Coronary Artery Disease Reporting and Data System, or CAD-RADS, and other scores, according to a study published in Radiology.

“The risk estimate that you get from doing the machine learning version of the model is more accurate than the risk estimate you’re going to get if you rely on CAD-RADS,” Kevin M. Johnson, MD, associate professor of radiology and biomedical imaging at Yale School of Medicine, said in a press release. “Both methods perform better than just using the Framingham risk estimate. This shows the value of looking at the coronary arteries to better estimate people’s risk.”

Coronary CTA

Researchers analyzed data from 6,892 patients (mean age, 53 years; 65% men) who underwent CTA between February 2004 and November 2009. Indications for CTA included indeterminate stress test, stable atypical chest pain, strong family history, multiple risk factors and strong personal concern for CAD.

Four features from 16 coronary segments were assessed including plaque amount, plaque calcification, diameter stenosis and vessel remodeling. In addition, five convention vessel scores were calculated including CAD-RADS, segment stenosis score, segment involvement score, CT Leaman score and segment plaque burden score.

Outcomes of interest included CAD-related death, all-cause mortality or nonfatal MI.

During a median follow-up of 9 years, there were 380 deaths, 70 of which were related to CAD. There were also 43 nonfatal MIs during follow-up.

The area under the curve for all-cause mortality was 0.77 for machine learning (95% CI, 0.76-0.77) vs. 0.72 for CAD-RADS (95% CI, 0.72-0.72). The AUC for deaths related to CAD was 0.85 for machine learning (95% CI, 0.84-0.85) and 0.79 for CAD-RADS (95% CI, 0.78-0.8).

If the choice regarding statin therapy was made to tolerate the treatment of 45 patients and included one patient who would die of CAD, machine learning would determine that 93% of patients with events will be given statin therapy. CAD-RADS would only determine that 69% of patients would be treated.

“There is extensive evidence that the volume of plaque and remodeling are important prognostic factors and are uniquely visible on coronary CT angiographic images, yet are not included in the CAD-RADS score,” Johnson and colleagues wrote. “CAD-RADS is based on an analysis of stenoses alone, in the tradition of conventional coronary angiography. It should be conceded that CAD-RADS has several important goals, of which prognosis is only one, albeit an important one.”

Future implementation

“Before clinical deployment of [machine learning] algorithms, large multicenter evaluations of [machine learning] validation in dependent datasets will be essential to demonstrate generality in various patient cohorts,” U. Joseph Schoepf, MD, director of cardiovascular imaging at the Medical University of South Carolina in Charleston, and Christian Tesche, MD, head of cardiovascular imaging at Heart Center Munich-Bogenhausen in Germany, wrote in a related editorial. “Another critical point is the cost-effectiveness of [machine learning] models and the costs associated with extensive software development and the information technology infrastructure. This must be evaluated further in large-scale multicenter trials.” – by Darlene Dobkowski

Disclosures: The authors of the study and editorial report no relevant financial disclosures.

U. Joseph Schoepf
U. Joseph Schoepf

Coronary CTA assessed by machine learning was able to better distinguish patients who had an adverse event compared with the Coronary Artery Disease Reporting and Data System, or CAD-RADS, and other scores, according to a study published in Radiology.

“The risk estimate that you get from doing the machine learning version of the model is more accurate than the risk estimate you’re going to get if you rely on CAD-RADS,” Kevin M. Johnson, MD, associate professor of radiology and biomedical imaging at Yale School of Medicine, said in a press release. “Both methods perform better than just using the Framingham risk estimate. This shows the value of looking at the coronary arteries to better estimate people’s risk.”

Coronary CTA

Researchers analyzed data from 6,892 patients (mean age, 53 years; 65% men) who underwent CTA between February 2004 and November 2009. Indications for CTA included indeterminate stress test, stable atypical chest pain, strong family history, multiple risk factors and strong personal concern for CAD.

Four features from 16 coronary segments were assessed including plaque amount, plaque calcification, diameter stenosis and vessel remodeling. In addition, five convention vessel scores were calculated including CAD-RADS, segment stenosis score, segment involvement score, CT Leaman score and segment plaque burden score.

Outcomes of interest included CAD-related death, all-cause mortality or nonfatal MI.

During a median follow-up of 9 years, there were 380 deaths, 70 of which were related to CAD. There were also 43 nonfatal MIs during follow-up.

The area under the curve for all-cause mortality was 0.77 for machine learning (95% CI, 0.76-0.77) vs. 0.72 for CAD-RADS (95% CI, 0.72-0.72). The AUC for deaths related to CAD was 0.85 for machine learning (95% CI, 0.84-0.85) and 0.79 for CAD-RADS (95% CI, 0.78-0.8).

If the choice regarding statin therapy was made to tolerate the treatment of 45 patients and included one patient who would die of CAD, machine learning would determine that 93% of patients with events will be given statin therapy. CAD-RADS would only determine that 69% of patients would be treated.

“There is extensive evidence that the volume of plaque and remodeling are important prognostic factors and are uniquely visible on coronary CT angiographic images, yet are not included in the CAD-RADS score,” Johnson and colleagues wrote. “CAD-RADS is based on an analysis of stenoses alone, in the tradition of conventional coronary angiography. It should be conceded that CAD-RADS has several important goals, of which prognosis is only one, albeit an important one.”

PAGE BREAK

Future implementation

“Before clinical deployment of [machine learning] algorithms, large multicenter evaluations of [machine learning] validation in dependent datasets will be essential to demonstrate generality in various patient cohorts,” U. Joseph Schoepf, MD, director of cardiovascular imaging at the Medical University of South Carolina in Charleston, and Christian Tesche, MD, head of cardiovascular imaging at Heart Center Munich-Bogenhausen in Germany, wrote in a related editorial. “Another critical point is the cost-effectiveness of [machine learning] models and the costs associated with extensive software development and the information technology infrastructure. This must be evaluated further in large-scale multicenter trials.” – by Darlene Dobkowski

Disclosures: The authors of the study and editorial report no relevant financial disclosures.