Using AI to help predict cardiac arrests

Perelman School of Medicine cardiologist Rajat Deo has been studying electrocardiographic (ECG) data and cardiac rhythms for nearly two decades at Penn.

He says that every second, hospitals generate “enormous streams of ECG data—electrical traces of the heart that accumulate into one of medicine’s richest, most underused archives.” For decades, most of this information had been treated as crucial in the moment, but it goes unused later.

Now, Deo and other clinicians from Penn Medicine have partnered with computer scientists at Penn’s School of Engineering and Applied Science to change that by harnessing the power of the cardiac data that hospitals already collect.

Cardiologists like Deo bring clinical insight into how small electrical irregularities can foreshadow serious cardiac events, while computer scientists like Rajeev Alur bring decades of work on systems that find patterns in complex, constantly evolving streams of information and use them to make predictions.

“At Penn, you can walk across the street and find a clinician who can challenge you to develop an AI-based solution to a problem that they want to solve,” Alur says.

Together with students and faculty across the University, the team developed the Cardiac Autoregressive Model for ECG Language-Modeling (CAMEL), an artificial intelligence model that treats ECG less like isolated snapshots and more like language.

Rather than simply identifying abnormalities after they appear, CAMEL analyzes longer stretches of heart rhythm to recognize patterns that may signal what comes next, paving the way for warnings of arrhythmias or cardiac arrest 10 to 15 minutes before they happen.

A new kind of model

“Existing AI models focus on classifying the ECG signals,” says Alur. To shift from classification to forecasting, the engineering team had to treat the heart’s electrical rhythms like text, figuring out how to best encode the ECG signal.

CAMEL converts segments of ECG waveforms into a format that can be interpreted alongside clinical text, such as doctor’s notes or lab results, allowing the system to reason—like a clinician—about how subtle variations in rhythm may signal future changes. Unlike traditional approaches that rely on 10-second snippets, the model is designed to handle hours of telemetry (the continuous remote monitoring of patient vital signs in hospitals), thereby expanding the window in which risk can be detected.

View large image

For clinicians, the difference could be meaningful, as cardiac deterioration rarely appears out of nowhere; warning signs often exist but remain too faint, diffuse, or complex for conventional tools to interpret.

Deo notes that the model has shown promising results in analyzing normal sinus rhythms and detecting subtle indicators that a patient may be at high risk of in-hospital cardiac arrest caused by dangerous arrhythmias such as ventricular fibrillation or ventricular tachycardia.

Working toward real-world use

The researchers hope to soon test the model by processing real-time patient information in the background—but without alerting medical staff.

“The last thing I want to do is alert nurses and technicians on the floor at 2 a.m. to intervene based on a false signal,” Deo says. “Every time we trigger an alarm, we are diverting a finite resource from another patient who may be in need. In a clinical setting, we have to be certain.”

The researchers will also weigh the model’s predictions against patient outcomes to determine if CAMEL’s foresight exceeds the current standard of care.

Beyond its use in the hospital, the researchers are optimistic about the potential of this technology in the general population using consumer wearable devices. 

Rajeev Alur is the Zisman Family Professor in the Department of Computer and Information Science at the University of Pennsylvania’s School of Engineering and Applied Science. He is also Director of the ASSET Center for Trustworthy AI. 

Rajat Deo is a professor of Medicine in the Electrophysiology Section and Division of Cardiovascular Medicine at the University of Pennsylvania’s Perelman School of Medicine.

Other authors Seewon Choi, Mayank Keoliya, Sameed Khatana, Mayur Naik, Alireza Oraii, Alaia Solko-Breslin, Neelay Velingker, and Eric Wong of the University of Pennsylvania. 

This research was supported by the ARPA-H program on Safe and Explainable AI (Award D24AC00253-00), the National Institutes of Health (Award R01-EB036016, and Amazon Web Services. The researchers also acknowledged the Penn Advanced Research Computing Center (PARCC) for providing high-performance computing infrastructure, and Penn Medicine's CIRCE project (The Complete Inpatient Record using Comprehensive Electronic data) for providing high-quality dataset for training and testing.