Researchers at NYU Langone Health have developed an artificial intelligence model that successfully detects heart transplant rejection by combining electrocardiogram readings with routine blood biomarkers, bypassing the need for invasive surgical biopsies in a majority of evaluated cases.
In Plain English: The Clinical Takeaway
- Non-Invasive Screening: The new tool uses skin sensors and standard blood draws to evaluate whether an immune system is attacking a donated heart, eliminating surgical tissue removal for many patients.
- Improved Accuracy: By analyzing 5,300 electrocardiogram readings alongside molecular biomarkers, the combined AI model correctly identified 94 percent of patients who were free of rejection.
- Fewer False Positives: Unlike models relying solely on blood tests, which incorrectly flagged 19 patients for potential biopsies in the test group, the multi-modal approach significantly reduces unnecessary medical procedures.
Training the Multi-Modal AI Model on Historical Patient Data
To detect cardiac allograft rejection today, physicians typically rely on a surgical biopsy—a procedure where medical professionals extract a minuscule portion of cardiac tissue to examine under a microscope for signs of inflammation and other abnormalities that point to an immune system attacking the organ. To evaluate a less invasive alternative, investigators at NYU Langone Health turned to machine learning. The team trained algorithms using historical clinical records from adult heart transplant recipients treated between 2018 and 2024.
The dataset comprised 5,300 electrocardiogram (EKG) readings paired with biopsy results gathered within a one-month window. Researchers grouped organ rejections into two distinct categories: no or mild rejection versus moderate or severe rejection. Treatment modifications, such as adjusting immunosuppressive medications, are typically reserved only for the latter, more serious cases.
Previous clinical work established that certain molecular blood biomarkers are effective at spotting cellular rejection and fragments of donor DNA circulating in a recipient’s bloodstream. However, those individual tests frequently yield false-positive results. By merging molecular biomarker data with the electrical activity recorded by EKGs, the NYU Langone team constructed a multi-modal AI framework designed to cross-reference physiological signals against actual biopsy outcomes.
Performance Comparison Against Biopsy Benchmarks
In a test group consisting of 38 male and female heart transplant recipients, the combined AI model demonstrated superior predictive capabilities compared to single-modality evaluations. While models relying exclusively on blood tests flagged numerous patients who ultimately did not require intervention, the integrated EKG-biomarker model correctly identified 94 percent of stable patients.
| Evaluation Method | Data Inputs | Clinical Performance & Limitations |
|---|---|---|
| Standard Surgical Biopsy | Microscopic tissue inspection | Current clinical gold standard; highly invasive. |
| Single-Modality Blood Test | Gene activity and donor DNA fragments | Effective at spotting cellular rejection, but prone to high false-positive rates leading to unnecessary procedures. |
| Combined AI Model | EKG readings + molecular biomarkers | Correctly identified 94 percent of patients without rejection in test groups, sparing them from invasive tissue sampling. |
Lior Jankelson, MD, PhD, who serves on the faculty at the Leon H. Charney Division of Cardiology within NYU Grossman School of Medicine’s Department of Medicine alongside a concurrent appointment as an associate professor of biomedical engineering at the NYU Tandon School of Engineering, pointed out the concealed diagnostic insights captured by the platform. 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,
Jankelson stated following the release of the findings.
The study detailing these findings was published online on September 25 in the Journal of Heart and Lung Transplantation. The authors noted that their project marks the first instance of merging such blood indicators with EKG data inside a unified machine learning algorithm while benchmarking the output directly against biopsy findings.
Funding for the study was provided by NYU Langone. Alongside Jankelson, the research team included Kevin Chen, MD; Robert Ronan, MS; Larry A. Chinitz, MD; and Randal I. Goldberg, MD. Looking ahead, the investigators intend to evaluate the performance of their algorithm across a broader cohort of patients spanning multiple transplant institutions.
References
- Chen, K., et al. (2026). Multi-modal AI: Integrating electrocardiography with molecular biomarkers for noninvasive detection of cardiac allograft rejection. JHLT Open. DOI: 10.1016/j.jhlto.2026.100698.
