Geisinger, Tempus Develop AI Model to Predict Heart Disease

Danville, Pennsylvania-based Geisinger health system and biotechnology research company Tempus have developed an electrocardiogram (ECG)-based artificial intelligence (AI) model that can predict undiagnosed structural heart disease (SHD).

SHD is a group of various conditions that negatively affect the valves, walls, chambers, or muscles of the heart. SHD is a progressive disease, which means that the symptoms worsen over time. These worsening symptoms can debilitate patients and eventually lead to death, making early diagnosis and treatment key to improving patient outcomes.

To spur early diagnosis, researchers at Geisinger and Tempus developed an AI model that can flag high-risk SHD patients.

To create the model, they conducted a study for which they gathered data from 2.2 million ECGs from 480,000 patients who received care at Geisinger in the past 37 years. This data was used to train a deep neural network (DNN), a type of deep learning AI, to predict which patients, among those without a prior history of SHD, would develop the disease and could benefit from monitoring or treatment.

Their model, called rECHOmmend, achieved excellent performance at predicting the risk of developing any one of seven SHDs that are diagnosable by standard ECG. The model significantly outperformed previously published models predicting any single SHD.

“Past studies have shown the ability of artificial intelligence to enable single disease screening with echocardiography. The rECHOmmend study builds on those to further improve the feasibility of echocardiography as a screening tool for structural heart disease,” said Alvaro Ulloa Cerna, PhD, senior data scientist at Geisinger and a lead author of the study, in the press release. “This could allow for earlier diagnosis and potentially avoid further disease development and its debilitating symptoms.”

The study findings show that clinicians using rECHOmmend could identify more diseases using fewer diagnostic studies, building on previous AI-based cardiology research Geisinger and Tempus have collaborated on.

In May 2020, the two organizations partnered to examine whether AI can accurately predict mortality using ECG data. Their study, published in Nature Medicine, found that a DNN achieved high performance when tasked with predicting one-year all-cause mortality, even in a dataset with a large subset of ECGs interpreted as “normal” by clinicians.

The organizations published a second study in Circulation last year, which detailed an AI model they created to predict the risk of new-onset atrial fibrillation (AF) and AF-related stroke in patients without a history of AF. The model achieved excellent performance, and it was granted Breakthrough Device Designation by the US Food & Drug Administration.

 The AF model was designated for use in patients 40 years of age and older, who do not have pre-existing or concurrent AF or atrial flutter, and who are at elevated risk of stroke based on a commonly used clinical stroke risk assessment tool.