by Pennsylvania State University

Visualization of an individual data sample as a P-QRS-T wave progression viewed from 12 leads. Credit: Heart Rhythm (2024). DOI: 10.1016/j.hrthm.2024.07.061

Researchers at Penn State are using machine learning and existing electrocardiogram (ECG) data to help doctors make more accurate predictions. A team of artificial intelligence engineers, in collaboration with a team of physicians from Penn State Heart and Vascular Institute, is working to develop novel algorithms for point-of-care, in-house use and for technology licensing.

“With stronger algorithms and a larger database, we can predict cardiovascular outcomes at significantly less cost,” said Ankit Maheshwari, assistant professor of medicine at Penn State and lead researcher on the project. “And there are several other examples out there where it can make health care more efficient and more effective.”

A promising pilot

An initial pilot study published in the journal Heart Rhythm in September reported a model that could predict whether a patient with stroke of unknown cause would develop atrial fibrillation (AFib)—or irregular heartbeat—by analyzing one heartbeat from a relatively cheap and common heart test, the standard 12-lead ECG, which measures the electrical activity in the heart.

A significant proportion of strokes of unknown cause are related to underlying subclinical, paroxysmal AFib, which is an irregular heartbeat that may last for a few minutes but the patient doesn’t experience any symptoms.

AFib-related strokes can be prevented by blood thinners. The current standard of care is to implant a loop recorder, a device placed under the skin that tracks heart activity to check for AFib. This helps doctors decide if patients should take blood thinners to prevent future strokes, Maheshwari explained.

The research team wanted to see if they could use the standard 12-lead ECG to predict AFib instead of relying on a loop recorder. The team compiled a small data set using existing ECG data from Penn State comprising patients with cryptogenic stroke, or stroke without a clear cause, who had loop recorders implanted, as well as data from 12-lead ECGs.

Using machine learning algorithms, the team built a model that could take a patient’s 12-lead ECG and predict whether they would or would not develop AFib. The model correctly classified 80% of patients in the test cohort.

“This pilot study shows that even with a smaller group of 200 to 300 patients, we could create a useful predictive model,” Maheshwari said.

Scaling up: Expanding the database, extending the application

Next, the team aims to expand the database, allowing for broader applications, Maheshwari explained. He added that the researchers accomplished high levels of accuracy through data augmentation techniques, which improved predictive performance.

“Our goal is to organize the 1.8 million ECGs in the University’s medical record system into a searchable database to facilitate large volume ECG analysis to support future projects aimed at utilizing a 12-lead ECG to predict cardiovascular outcomes and improve patient care,” Maheshwari said.

Key to this development is collaboration with other institutions, Maheshwari said. Collaboration provides an opportunity to validate their model across independent datasets, which is important for confirming that it can be used effectively in larger clinical settings, he explained.

“I’m grateful for the leadership of our CTSI Informatics Core and their engagement of our outstanding collaborators,” said Jennifer Kraschnewski, director of Penn State CTSI. “Their important efforts have leveraged expertise across our university to bring the power of AI and machine learning to opportunities with our electronic health records to take clinical and translational sciences into the future.”

The broader implications of predictive models

The research team’s predictive modeling could also be useful beyond just AFib, with applications in other heart-related areas, the researchers said. For instance, there is the possibility of using machine learning to predict when to use pacemakers in patients undergoing transcatheter aortic valve replacement procedures, Maheshwari explained. This type of procedure is minimally invasive and involves threading a catheter through the groin and to the heart to replace a failing valve with a human-made one.

“This could lead to better patient selection and outcomes, helping doctors identify who is most likely to benefit from the procedure and who might face complications,” he said.

According to the researchers, the ECG data could also help predict the presence of coronary lesions without the need for imaging stress tests. Such predictions could streamline diagnostic processes and reduce costs as well.

“These advancements highlight the great potential of machine learning to make heart care more efficient,” Maheshwari said.

He added that the human element of machine learning will always be crucial, emphasizing that human understanding of the biological signals within ECGs improves the effectiveness of machine-learning models.

“There’s quite a bit of human work that’s involved. And the more you understand about biology, you can kind of tailor the machine to help you more effectively,” he said.

Future directions

With initial funding secured, the team is now focused on building their ECG database and validating their predictive models. Their goal is to have a fully functional database enabling them to conduct larger-scale studies and potentially apply for additional funding to support randomized controlled trials.

“Powered by large data sets, AI offers unprecedented opportunities for advancing biomedical discoveries and individual and population health outcomes,” said Vasant Honavar, co-lead of Penn State CTSI’s informatics core, Dorothy Foehr Huck and J. Lloyd Huck Chair in Biomedical Data Sciences and Artificial Intelligence, and director of the Center for Artificial Intelligence Foundations and Scientific Applications at Penn State.

“Our efforts not only hold promise for improving patient care but also represent a paradigm shift in how cardiovascular diseases may be diagnosed and managed in the coming years,” Maheshwari said.

More information: R.S. Shah et al, ID: 4121654 Analysis of a Single Heart Beat with Deep Learning for Prediction of Atrial Fibrillation in Patients with Cryptogenic Stroke: A Novel Approach to Electrocardiogram Data Augmentation, Heart Rhythm (2024). DOI: 10.1016/j.hrthm.2024.07.061

Provided by Pennsylvania State University

---

Explore further

New AI-powered algorithm could better assess people’s risk of common heart condition