Algorithm tops physicians in predicting heart attacks, death

A machine learning algorithm can now predict death and MI more accurately than certified cardiologists, according to research presented at the International Conference on Nuclear Cardiology and Cardiac CT in Lisbon, Portugal, this May.

The algorithm, utilized by Luis Eduardo Juarez-Orozco, MD, PhD, and colleagues at the Turku PET Centre in Finland, uses an approach much like Netflix would to recommend movies to predict a person’s risk of dying or suffering a heart attack. Juarez-Orozco et al. fed the model imaging data from 950 heart patients with known six-year outcomes, allowing the algorithm to “learn” from the various ways data interacted to produce certain results.

“Doctors already collect a lot of information about patients, for example those with chest pain,” Juarez-Orozco said in a release. “We found that machine learning can integrate these data and accurately predict individual risk. This should allow us to personalize treatment and ultimately lead to better outcomes for patients.”

The authors’ study retrospectively enrolled patients with chest pain who underwent Turku PET Centre’s usual protocol for identifying coronary artery disease (CAD), which includes a coronary computed tomography angiography (CCTA) scan that yields 58 unique pieces of data about coronary plaque, vessel narrowing and calcification. Patients whose scans flagged them for potential CAD underwent additional positron emission tomography (PET) imaging that produced another 17 variables on blood flow

The researchers considered a further 10 clinical variables obtained from medical records—including sex, age, smoking status and presence of diabetes—for a total of 85 data points. They input those 85 variables into a machine learning algorithm called LogitBoost, which analyzed them alongside known patient outcomes to determine the best structure for predicting death and MI.

“The algorithm progressively learns from the data and after numerous rounds of analyses, it figures out the high-dimensional patterns that should be used to efficiently identify patients who have the event,” Juarez-Orozco said. “The result is a score of individual risk.”

During six years of follow-up, 24 patients in the pool had heart attacks and 49 died from any cause. Juarez-Orozco and his team said their algorithm was modestly successful in predicting death and MI from the 10 clinical variables alone, reaching an area under the curve (AUC) of 0.65 where 1.0 is a perfect test and 0.5 is a random result. Adding PET data increased the AUC to 0.69, while further adding CCTA data boosted the AUC to 0.82 and gave the model an accuracy of more than 90%.

“These advances are far beyond what has been done in medicine, where we need to be cautious about how we evaluate risk and outcomes,” Juarez-Orozco said. “We have the data but we are not using it to its full potential yet.”