An artificial intelligence (AI) model has performed dramatically better than doctors using the latest clinical guidelines to predict the risk for sudden cardiac arrest in people with hypertrophic cardiomyopathy.
The model, called Multimodal AI for ventricular Arrhythmia Risk Stratification (MAARS), is described in a paper published online on July 2 in Nature Cardiovascular Research. It predicts patients’ risk by analyzing a variety of medical data and records such as echocardiogram and radiology reports, as well as all the information contained in contrast-enhanced MRI (CMR) images of the patient’s heart.
Natalia Trayanova, PhD, director of the Alliance for Cardiovascular Diagnostic and Treatment Innovation at Johns Hopkins University in Baltimore, led the development of the model. She said that while hypertrophic cardiomyopathy is one of the most common inherited heart diseases, affecting 1 in every 200-500 individuals worldwide, and is a leading cause of sudden cardiac death in young people and athletes, an individual’s risk for cardiac arrest remains difficult to predict.
Current clinical guidelines from the American Heart Association and American College of Cardiology, and those from the European Society of Cardiology, identify the patients who go on to experience cardiac arrest in about half of cases.
“The clinical guidelines are extremely inaccurate, little better than throwing dice,” Trayanova, who is also the Murray B. Sachs Professor in the Department of Biomedical Engineering at Johns Hopkins, told Medscape Medical News.
Compared to the guidelines, MAARS was nearly twice as sensitive, achieving 89% accuracy across all patients and 93% accuracy for those 40-60 years old, the group of people with hypertrophic cardiomyopathy most at risk for sudden cardiac death.
Building a Model
MAARS was trained on data from 553 patients in The Johns Hopkins Hospital, Baltimore, hypertrophic cardiomyopathy registry. The researchers then tested the algorithm on an independent external cohort of 286 patients from the Sanger Heart & Vascular Institute hypertrophic cardiomyopathy registry in Charlotte, North Carolina.
The model uses all of the data available from these patients, drawing on electronic health records, ECG readings, reports from radiologists and imaging technicians, and raw data from CMR.
“All these different channels are fed into this multimodal AI predictor, which fuses it together and comes up with the risk for these particular patients,” Trayanova said.
The inclusion of CMR data is particularly important, she said, because the imaging test can identify areas of scarring on the heart that characterize hypertrophic cardiomyopathy. But clinicians have yet to be able to make much use of those images because linking the fairly random patterns of scar tissue to clinical outcomes has been a challenge.
But that is just the sort of task that deep neural networks are particularly well-suited to tackle. “They can recognize patterns in the data that humans miss, then analyze and combine them with the other inputs into a single prediction,” Trayanova said.
Clinical Benefits
Better predictions of the risk for serious adverse outcomes will help improve care, by ensuring people get the right treatments to reduce their risk, and avoid the ones that are unnecessary, Trayanova said The best way to protect against sudden cardiac arrest is with an implantable defibrillator — but the procedure carries potential risks that are best avoided unless truly needed.
“More accurate risk prediction means fewer patients might undergo unnecessary ICD implantation, which carries risks such as infections, device malfunction, and inappropriate shocks,” said Antonis Armoundas, PhD, from the Cardiovascular Research Center at Massachusetts General Hospital in Boston.
The model could also help personalize treatment for patients with hypertrophic cardiomyopathy, Trayanova said. “It’s able to drill down into each patient and predict which parameters are the most important to help influence the management of the condition,” she said.
Robert Avram, MD, MSc, a cardiologist at the Montreal Heart Institute, Montreal, Quebec, Canada, said the results are encouraging. “I’m especially interested in how a tool like this could streamline risk stratification and ultimately improve patient outcomes,” he said.
But it is not yet ready for widespread use in the clinic. “Before it can be adopted in routine care, however, we’ll need rigorous external validation across diverse institutions, harmonized variable definitions, and unified extraction pipelines for each modality, along with clear regulatory and workflow-integration strategies,” Avram said.
Armoundas said he would like to see the model tested on larger sample sizes, with greater diversity in healthcare settings, geographical regions, and demographics, as well as prospective, randomized studies and comparisons against other AI predictive models.
“Further validation in larger cohorts and assessment over longer follow-up periods are necessary for its full clinical integration,” he said.
Armoundas, Avram, and Trayanova reported having no relevant financial conflicts of interest.
Brian Owens is a freelance journalist based in New Brunswick, Canada.
