AI model could reduce biopsies after heart transplants

Artificial intelligence could help doctors identify heart transplant patients who are unlikely to be experiencing organ rejection, potentially reducing the number of invasive biopsies they need to undergo. Researchers at NYU Langone Health have developed an AI model that combines electrocardiogram (ECG) data with two existing blood biomarkers to assess the risk of transplant rejection.

In a small test group, the combined approach correctly identified 94 percent of patients who were not experiencing clinically significant rejection. The researchers emphasise that larger, multicentre studies are now needed before the technology could become part of routine transplant monitoring.

Biopsies remain the standard

Rejection is an important complication following heart transplantation. The recipient's immune system can attack the donated organ, making early detection important because rejection can be treated, including through adjustments to immunosuppressive medication. The current diagnostic gold standard is an endomyocardial biopsy. During this invasive procedure, a small sample of heart muscle is removed and examined microscopically for inflammation and other signs of rejection. Heart transplant recipients can undergo multiple biopsies during follow-up.

Blood tests provide a less invasive alternative for estimating rejection risk. Two biomarkers are already used for this purpose: one measures gene activity associated with cellular rejection, while the other detects fragments of donor-derived DNA circulating in the recipient's bloodstream. These tests can identify patients at risk, but false-positive results remain a limitation. Patients can consequently be referred for a biopsy even when significant rejection is ultimately not found.

ECG adds another source of information

The NYU researchers investigated whether information contained in routine ECG recordings could improve these blood-based assessments. ECGs measure the heart's electrical activity through electrodes placed on the skin and are widely available in clinical practice. The researchers trained AI models using around 5,300 ECG recordings from 2,357 adult heart transplant recipients treated between 2018 and 2024. One model analysed ECG information alone. A second combined the ECG data with results from the two blood biomarkers.

Each ECG was linked to a biopsy performed within the preceding month. The biopsy findings provided the reference against which the models were trained and evaluated. Rejection was divided into two groups: no or mild rejection and moderate or severe rejection. The distinction is clinically relevant because treatment changes are generally made for more serious rejection. The researchers say this is the first study to combine the two blood biomarkers and ECG information within a single AI model and directly compare its predictions with biopsy findings.

Fewer false alarms

The combined model was subsequently evaluated in an additional group of 38 male and female heart transplant recipients. It correctly identified 94 percent of patients who were not experiencing significant rejection. According to the researchers, using the blood biomarkers without the additional ECG analysis would have incorrectly identified 19 patients as potentially requiring a biopsy. The combined model classified those patients as not experiencing significant rejection, meaning they potentially could have avoided the invasive procedure.

The results suggest that ECG recordings contain physiological information that is relevant to transplant rejection but may not be apparent from conventional interpretation alone. AI can analyse subtle patterns across large numbers of ECG measurements and combine them with molecular information from blood tests. Rather than introducing another diagnostic test, the approach therefore attempts to extract additional information from tests already used in transplant follow-up.

Larger studies needed

The findings should nevertheless be interpreted cautiously. The model was developed using retrospective data and its additional test group consisted of only 38 patients. Its performance therefore does not yet demonstrate that AI-supported monitoring can safely replace biopsies in routine care. The researchers plan to evaluate the model in larger patient populations across multiple transplant centres. Such studies will be important for establishing whether its performance remains consistent across different hospitals and patient groups and, ultimately, whether using it to reduce biopsies is clinically safe.

If those studies confirm the findings, the potential value of the approach could lie particularly in ruling out significant rejection. Combining routinely available ECG data with established blood biomarkers could help clinicians determine more accurately which patients really need an invasive biopsy. For transplant medicine, that would represent a shift from relying on repeated tissue sampling towards combining different non-invasive signals. AI would not diagnose rejection independently, but could help integrate those signals into a more precise assessment of when further investigation is warranted.

FAU research

Earlier this month, we reported about a research, conducted at Florida Atlantic University (FAU) exploring whether AI combined with Raman spectroscopy could help identify skin cancer without immediately requiring a biopsy. The technique analyses how light interacts with molecules in tissue, producing a molecular fingerprint that can distinguish healthy from potentially cancerous tissue.

The approach is still at an early stage and is not intended to replace conventional pathology, which remains the diagnostic gold standard. Larger studies and further algorithm development are needed. If validated, however, AI-assisted Raman spectroscopy could provide clinicians with additional information before removing tissue, helping determine which suspicious lesions genuinely warrant a biopsy and potentially reducing unnecessary invasive procedures.