Brief
When a deceased donor kidney becomes available for transplant, transplant surgeons must rapidly determine whether the organ is an appropriate match for a specific recipient. While every offer must go through various filters to ensure a baseline level of organ quality, the final organ acceptance decision requires synthesizing donor characteristics, recipient medical complexity, and contextual factors, often under significant time pressure. To support this decision-making process, our team is developing a machine learning model that leverages donor and recipient data to predict the risk of short- and long-term kidney dysfunction following kidney transplantation. By providing objective, data-driven risk estimates, this tool aims to improve match quality, standardize decision-making, and ultimately improve patient outcomes.
Problem
Kidney transplant offer-evaluation is complex, high-stakes, and time-sensitive. Donor information is uploaded into the United Network for Organ Sharing (UNOS) UNET system and offered to transplant centers in sequence. At Duke, the on-call surgeon must quickly assess donor kidney quality and determine whether a given recipient is likely to achieve an acceptable outcome. The current process presents several challenges, including:
- Surgeon-to-surgeon variability in acceptance practices
- Inconsistent decision-making over time
- An effort-intensive evaluation process requiring review of structured and unstructured data
- Lost opportunities to transplant usable kidneys
- Occasional use of poor-quality kidneys
- Prolonged cold ischemia time due to delayed decisions
These limitations may contribute to patient harm, inefficiencies in organ utilization, and variability in transplant outcomes. Importantly, existing UNOS filters and structured data fields do not capture the full complexity of donor-recipient interactions, nor do they incorporate longitudinal recipient data or nuanced clinical narratives. A data-driven predictive approach offers the opportunity to support transplant surgeons with individualized risk estimates, helping optimize kidney acceptance decisions while preserving clinical judgment.
Solution
To address this problem, the Duke Transplant Center partnered with the Duke Institute for Health Innovation to develop a machine learning model that predicts short- and long-term graft dysfunction outcomes following kidney transplantation, namely 90-day graft failure, 1-year graft failure, and whether creatinine level exceeded 2.5 mg/dL 6-months post-transplant. Our goal is to provide transplant surgeons and coordinators with objective, patient-specific risk predictions at the time of organ offer, enabling more standardized and informed “go/no-go” decisions. By supporting optimal donor-recipient matching, this tool aims to improve graft outcomes, reduce unnecessary organ discard, and increase the number of successful kidney transplants performed at Duke.
We used data collected from 174,708 kidney transplant events across the United States between 2015 and 2025, curated from the United Network for Organ Sharing (UNOS) registry. The dataset was filtered for adult transplant recipients and deceased donors. Outcomes were derived from UNOS follow-up reporting. 90-day and 1-year graft failure were defined based on whether a reported graft failure date occurred within the respective time interval following the transplant date. Six-month renal dysfunction was defined using the reported recipient creatinine value at six months post-transplant, with values exceeding 2.5 mg/dL considered elevated.
Model inputs included structured donor and recipient variables capturing recipient demographics, comorbidities, and transplant status; waitlist history and dialysis characteristics; immunologic compatibility measures such as Human Leukocyte Antigens (HLA) mismatch and Calculated Panel Reactive Antibody (CPRA); donor demographics and cause of death; donor laboratory and infectious disease testing results; donor medical history and management prior to procurement; and transplant logistics including cold ischemia time and organ preservation methods. The model was trained using light gradient boosting machine (lgbm) with 70% of the data used for training and 15% each for validation and testing. Model performance was evaluated using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC-ROC).
Outcomes
Of the 174,708 kidney transplant events included in the modeling cohort, 3,723 (2.13%) were followed by 90-day graft failure, and 5,409 (3.10%) were followed by 1-year graft failure. For the analysis of renal dysfunction at six months post-transplant, the cohort was restricted to 152,892 kidney transplant events due to incomplete follow-up laboratory measurements in a subset of recipients. Within this subset, 7,262 (4.75%) recipients had a reported creatinine level greater than 2.5 mg/dL six months following transplantation. Model performance varied across the three predicted outcomes. Because graft failure and severe renal dysfunction are relatively low-prevalence events, the area under the precision-recall curve (AUPRC) provides a more informative measure of model performance than traditional discrimination metrics alone. AUPRC reflects how well the model identifies true positive cases relative to the expected baseline performance defined by outcome prevalence.
The model predicting 90-day graft failure achieved an AUPRC of 0.0454, representing approximately 2.1 times the baseline prevalence of the outcome. The 1-year graft failure model achieved an AUPRC of 0.059, approximately 1.9 times the baseline prevalence. The model predicting elevated creatinine (>2.5 mg/dL) six months post-transplant achieved an AUPRC of 0.130, approximately 2.7 times the baseline prevalence of the outcome. These results demonstrate that the models can identify transplant cases with elevated risk of graft dysfunction at rates substantially higher than random classification. Together, these findings suggest that donor-recipient characteristics captured in the UNOS registry contain meaningful predictive signal for clinically important post-transplant outcomes. The models provide a foundation for future clinical decision support tools that could assist transplant teams in evaluating donor-recipient matches and identifying higher-risk transplant scenarios
Next Steps (as of April 2026)
We are currently working with collaborators at the Duke Transplant Center to identify and prepare the appropriate data infrastructure required to retrieve donor and recipient information from the United Network for Organ Sharing (UNOS) in real-time. This will enable the predictive models to be executed at the time of organ offer, providing transplant surgeons and coordinators with risk estimates that can inform clinical decision-making during the donor evaluation process. In parallel, we are continuing to refine and expand the model feature set to evaluate whether we can construct more clinically meaningful variables to further improve predictive performance. Ultimately, our goal is to integrate these models into the transplant evaluation process to support more consistent donor–recipient matching decisions and improve kidney transplant outcomes.
