About 1 in 3 people with atrial fibrillation — the most common type of heart rhythm disorder — do not know they have the condition, which is problematic considering it increases the risk of stroke, heart failure and other cardiac complications. To address this problem, Cedars-Sinai developed an algorithm that can detect an abnormal heart rhythm in people not yet showing symptoms.
This week, the health system published research on the algorithm in JAMA Cardiology. The study found that the AI model does a good job of identifying hidden signals in common medical diagnostic testing and could help physicians better prevent strokes and other heart-related complications in people with atrial fibrillation.
Researchers within Cedars-Sinai’s Smidt Heart Institute began developing the algorithm a couple of years ago to take routine electrocardiograms (ECGs) and use them to predict different cardiovascular disease processes, said Dr. Neal Yuan, who is a cardiologist and one of the study’s authors. They trained and tested the model using nearly a million ECGs from patients seen at six different VA hospitals, as well as Cedars-Sinai Medical Center in Los Angeles.
The algorithm makes predictions using “a complicated equation involving the 20,000 values that make up an ECG,” Dr. Yuan explained. The numbers used to make this equation come from plugging in thousands of ECGs into the model one at a time. With each ECG, the model tests the prediction ability of the equation. Based on how right or how wrong the prediction was, the model then makes adjustments to the equation. Over time, after it sees many examples, the algorithm’s equation becomes finely tuned, he said.
“One of the ongoing challenges in deep learning and machine learning is ensuring that a model that is developed using data from one group of individuals still works just as well when it is used in other populations. The early history of machine learning is filled with cautionary tales of models that did not generalize well,” Dr. Yuan pointed out.
For example, early facial recognition algorithms were known to perform significantly worse when applied to individuals of color because they were trained on datasets of mainly white faces, he noted. Additionally, Amazon used a hiring algorithm a few years ago that systematically discriminated against women because it was trained on the resumes and profiles of existing employees who were predominantly men, Dr. Yuan added.
For this study, Cedars-Sinai was “intentionally rigorous” about testing the model on various patient subgroups, including women, patients younger than 65, patients older than 65, Black patients, patients with few medical comorbidities and patients with many comorbidities, Dr. Yuan explained.
“The model performed just as well in these different subgroups and across different medical sites, which reassured us that our model was picking up on features in the ECG that are widely generalizable across all patients,” he declared.
Cedars-Sinai hopes that in the future, anytime someone has an ECG performed, an AI model will quickly make a prediction as to whether that individual has a high risk of atrial fibrillation, Dr. Yuan said. If a patient’s risk is high, their doctor can prescribe a more intensive monitoring test such as a patch monitor, he explained.
“I think deep learning algorithms are incredibly powerful and will allow us to better use existing tests such as ECGs to predict diseases earlier and more accurately. It allows us to do more with the information that we may already have. This could help start treatments earlier and avoid downstream complications from these diseases,” Dr. Yuan noted.
He said that the next step is deploying AI models like Cedar-Sinai’s into large clinical settings and studying how they affect physicians’ practice and patient outcomes. These sorts of studies are currently being planned and some are already underway, Dr. Yuan added.
Photo: ismagilov, Getty Images