Machine Learning Helps Choose Best Test for Chest Pain

Using machine learning, clinicians may be able to choose which imaging test to use for patients who may have coronary artery disease, a condition caused by plaque buildup in the arterial wall.

Yale researchers describe the machine learning tool, called ASSIST, in a study published in the European Heart Journal. The algorithm aims to focus on the long-term outcome for a given patient.

Functional testing, known as a stress test, examines patients for coronary artery disease by detecting reduced blood flow to the heart. The second test is anatomical testing, or coronary computed tomography angiography (CCTA), which identifies blockages in the blood vessels. Using machine learning techniques, ASSIST provides recommendations for each patient.

"There are strengths and limitations for each of these diagnostic tests," said Rohan Khera, MD, MS, an assistant professor of cardiology at Yale School of Medicine. "If you are able to establish the diagnosis correctly, you would be more likely to pursue optimal medical and procedural therapy, which may then influence the outcomes of patients."

Recent clinical trials have attempted to determine if one test is optimal. The PROMISE and SCOT-HEART clinical trials have indicated that anatomical imaging has similar outcomes to stress testing, but may improve long-term outcomes in certain patients.

"When patients present with chest pain you have two major testing strategies. Large clinical trials have been done without a conclusive answer, so we wanted to see if the trial data could be used to better understand whether a given patient would benefit from one testing strategy or the other," said Khera.

To develop the ASSIST tool, researchers gathered data from 9,572 patients who were enrolled in the PROMISE trial through the National Heart, Lung and Blood Institute. The team then created a novel strategy that embedded local data experiments within the larger clinical trial.

"A unique aspect of our approach is that we leverage both arms of a clinical trial, overcoming the limitation of real-world data, where decisions made by clinicians can introduce bias into algorithms," said Khera.

The tool proved effective in a distinct population of patients in the SCOT-HEART trial. Among 2,135 patients who underwent functional-first or anatomical-first testing, researchers saw a two-fold lower risk of adverse cardiac events when there was agreement between the test performed and the one recommended by ASSIST.

The group expects that this tool will offer clinicians further insight when they make the choice between anatomical or functional testing in chest pain evaluation.

"While we used advanced methods to derive ASSIST, its application is practical for the clinical setting. It relies on routinely captured patient characteristics and can be used by clinicians with a simple online calculator or can be incorporated in the electronic health record," said Evangelos Oikonomou, MD, DPhil, a resident physician in Internal Medicine at Yale and the study's first author.

Researchers have recently aimed to develop machine learning-driven clinical decision support tools.

A team from Columbia University developed a machine learning algorithm that can quickly analyze EHR data to identify chronic kidney disease, a condition that often goes undetected until it causes irreversible damage.

The model can automatically scan a patient’s EHR for results of blood and urine tests and uses a combination of established equations and machine learning techniques to process the data.

“Identifying kidney disease early is of paramount importance because we have treatments that can slow disease progression before the damage becomes irreversible,” said study leader Krzysztof Kiryluk, MD, associate professor of medicine at Columbia University Vagelos College of Physicians and Surgeons.

“Chronic kidney disease can cause multiple serious problems, including heart disease, anemia, or bone disease, and can lead to an early death, but its early stages are frequently under-recognized and undertreated.”