Enhancing the accuracy of solid-organ transplant compatibility predictions through machine learning technology could dramatically improve outcomes for patients awaiting life-saving surgeries. A recent study led by researchers at UNC Hospitals unveils promising advancements aimed at optimizing crossmatch compatibility between organ donors and recipients.
The study, published by Eric T. Weimer and K. A. Newhall on March 7, 2025, focuses on the increasingly significant role of machine learning (ML) to automate and refine the compatibility assessment process, thereby accelerating organ allocation and improving transplant success rates.
Currently, the demand for organ transplants is substantial, with over 114,000 patients waiting for organs across the United States. Specifically, more than 27,332 kidney transplants were performed last year, reflecting a 7.18% increase from 2022. This growing need highlights the necessity for more efficient and accurate tools to predict transplant compatibility.
To address this challenge, the researchers developed advanced models utilizing machine learning algorithms to predict crossmatch compatibility based on human leukocyte antigen (HLA) profiles. Their system particularly focuses on identifying donor-specific antibodies (DSA) through HLA allele imputation, which converts HLA antigens to alleles to optimize prediction performance.
According to the study, the machine learning models achieved an impressive receiver operating characteristic area under the curve (ROC-AUC) of 0.975, significantly outperforming traditional virtual crossmatch (VXM) methods, which have been reported at only 0.602. Such improvements suggest the ability of machine learning to fully leverage the complex biological interaction data previously overlooked through manual assessments.
From January 2015 to June 2024, the team analyzed flow cytometric crossmatches (XM) performed at UNC Hospitals, training their ML models on 9,892 XM cases. Of these, 9,292 were negative results, which constituted the vast majority (93.93%), and only 600 were positive (6.07%). To bolster their predictions and reduce false negatives, the researchers employed innovative techniques such as the Synthetic Minority Oversampling Technique (SMOTE) to balance data representation during training.
This approach enhances the accuracy of the model’s predictions by ensuring it learns effectively from both positive and negative outcomes, leading to fewer unnecessary physical crossmatches, which have historically increased the risk of delayed organ usage.
Machine learning is particularly adept at managing complex biological data, where patterns and correlations are difficult to discern through traditional methods. The HLA antibody model leverages this method to understand the interactions between donor organs and the recipient’s immune system, enabling personalized histocompatibility risk assessments.
Highlighting the impact of specific HLA antibodies, the study utilized permutation feature importance to assess how individual antibodies contributed to overall model predictions. This technique provided insight, illustrating which antibodies significantly affected crossmatch predictions, thereby transforming the assessment of transplantation risks.
While the machine learning models showcased high specificity (99.3%) and improved sensitivity compared to manual methods, the study noted challenges associated with determining the significance of certain antibodies, such as HLA-DQ or HLA-C. These observations are consistent with prior literature indicating discrepancies between serological and biological reactivity patterns, particularly with shared epitopes.
To summarize, the enhancements made through machine learning represent only the beginning of integrating advanced computational strategies within the field of transplant immunology. By employing these innovative models, transplant centers can streamline compatibility assessments, leading to potentially fewer delays and improved outcomes for patients awaiting transplants.
Future explorations may involve adapting these findings to broader datasets and incorporating additional parameters, such as eplet mismatches or antigen quality, which may refine predictions even more. This study serves as a foundation for developing technologies capable of revolutionizing the transplant process, ensuring more recipients can receive the timely care they desperately need.