Artificial intelligence could refine CAD prognosis

By Eleanor McDermid

A study shows that machine learning can predict mortality in patients with coronary artery disease (CAD) with greater accuracy than models based on coronary computed tomographic angiography (CCTA) or clinical variables.

"Appreciating and integrating the myriad risk predictors in an individual patient is a challenge for the clinician", say Piotr Slomka (Cedars-Sinai Medical Center, Los Angeles, California, USA) and co-researchers.

They say the complexity increases along with the number of known risk factors, and "the potential influence of unexpected interactions between several weaker predictors in an individual patient is often overlooked."

But the team shows that machine learning, in which computers identify patterns in large and complex datasets, "is able to overcome these challenges, by providing deep integration of the comprehensive CCTA and clinical data."

When applied to data for more than 10,000 patients with suspected CAD, machine learning distinguished between the 745 patients who died during 5 years of follow-up and those who did not with 79% accuracy. Machine learning was initially based on 44 CCTA and 25 clinical variables, ranked in order of information gain, although 14 variables did not add information and were not used in the model-building and cross-validation phases.

Various CCTA-based models - the segment stenosis score, the segment involvement score and the modified Duke index - had predictive accuracies for mortality ranging from 62% to 64%, the team reports in the European Heart Journal. They also note that they previously achieved 68% accuracy over 2 years with a clinical risk score plus two CCTA variables in the same dataset.

And the Framingham Risk Score (FRS) was 61% accurate. Slomka et al caution that the FRS was originally validated for 10-year outcomes in asymptomatic patients, but argue that it is often used regardless and that "it frames the accuracy of the [machine learning]-model in the context of a widely used and understood clinical score."

Machine learning categorised 39% of patients as low mortality risk (<3.8% risk), 48% as medium risk (between 3.8 and 14.0%) and 13% as high risk (>14.0%). The observed all-cause mortality rates in these three groups were 1.8%, 6.7% and 27.0%, respectively.

"The observed efficacy suggests [machine learning] has an important clinical role in evaluating prognostic risk in individual patients with suspected CAD", conclude the researchers.

Licensed from medwireNews with permission from Springer Healthcare Ltd. ©Springer Healthcare Ltd. All rights reserved. Neither of these parties endorse or recommend any commercial products, services, or equipment.