AI model flags heart rejection without biopsies, study finds

Nine out of ten patients who were not experiencing heart transplant rejection were correctly identified by a new artificial intelligence model that analyzes electrocardiograms alongside blood test results. This approach, detailed in a recent study led by researchers at NYU Langone Health, suggests that non-invasive monitoring could replace the current gold standard for diagnosis: the biopsy.

The limitations of traditional biopsies

Currently, diagnosing organ rejection requires surgically removing a small piece of heart muscle. Experts then examine this tissue under a microscope to look for inflammation or other cellular changes indicating that the immune system is attacking the donated organ. While effective, this procedure is invasive and carries inherent risks for patients.

The research team investigated whether combining data from electrocardiograms (EKGs), which record electrical activity via skin sensors, with standard blood tests could identify rejection patterns without surgery. They trained two distinct AI models using 5,300 EKG readings from 2,357 adult heart transplant recipients. One model relied solely on EKG data, while the second integrated these readings with results from two blood tests commonly used to predict rejection risk.

How does the combined model perform?

Biopsy records paired with corresponding EKG readings served as the benchmark for accuracy. In a separate test group comprising 38 male and female transplant recipients, the combined model demonstrated superior performance. It correctly identified 94 percent of patients who did not have rejection issues. Conversely, the model based only on blood tests incorrectly flagged 19 patients as potentially needing a biopsy, highlighting the prevalence of false positives in biomarker-only analysis.

"Our results highlight that electrocardiograms contain an abundance of physiological information that can be used to substantially improve the accuracy of detection and enable earlier diagnosis and treatment for patients with cardiac transplant rejection," said Lior Jankelson, MD, PhD, the study’s senior author. Jankelson serves as an associate professor in the Leon H. Charney Division of Cardiology at NYU Grossman School of Medicine and in biomedical engineering at NYU Tandon School of Engineering.

Study methodology and findings

Published online on Sept. 25 in the Journal of Heart and Lung Transplantation, this study is the first to merge blood biomarkers with EKG data into a single AI-enabled model and compare its output directly against biopsy results. The researchers noted that previous work established the effectiveness of blood biomarkers—measuring gene activity linked to cellular rejection and donor DNA fragments—in spotting rejection. However, these markers frequently produce false-positive results, leading to unnecessary invasive procedures.

To build the combined model, the team used data from heart transplant recipients treated between 2018 and 2024. Each EKG was matched to a biopsy performed within the preceding month, a timeframe supported by earlier research. Rejections were categorized into two groups: no or mild rejection versus moderate or severe rejection. Clinical guidelines typically recommend adjusting immune-suppressing medications only for the more serious cases.

"Heart transplantation is a major, resource-intensive procedure, and rejection is common but treatable if caught in time. Based on our findings, combining the availability of EKGs with the support of AI may help identify rejection earlier and save more lives," Jankelson stated. As a next step, the researchers plan to validate their model in a larger patient population across several centers.