A 10‐year analysis of the united network for organ sharing registry: Insights from machine learning to prevent pediatric waitlist deaths

Pediatric liver transplantation is the definitive treatment for children with end‐stage liver disease, malignancy, and selected metabolic disorders, with excellent post‐transplant outcomes in contemporary series. ¹ , ² Despite these advances, children continue to die while awaiting transplantation because of the limited availability of suitable, size‐matched donor organs. ¹ , ² Infants and adolescents are particularly vulnerable, reflecting both physiologic fragility and allocation constraints.

Risk stratification for pediatric liver transplant candidates has traditionally relied on the pediatric end‐stage liver disease (PELD) and model for end‐stage liver disease (MELD) scores, which estimate short‐term mortality risk. ³ While effective for population‐level prioritization, these tools do not fully capture dynamic clinical trajectories, center‐level practices, or logistical factors such as organ availability and regional sharing. ² , ⁴ Consequently, there has been increasing interest in applying machine learning (ML) methods to transplant data to improve prediction and identify high‐risk subgroups. ⁵ , ⁶

ML models can, in principle, model nonlinear relationships and complex interactions among many variables. However, it remains uncertain whether ML offers meaningful advantages over traditional regression when applied to structured national registry data, which are limited in granularity and lack unstructured clinical information. ⁷ , ⁸ , ⁹

In this study, we analyzed over a decade of United Network for Organ Sharing (UNOS) registry data to characterize age‐based disparities in pediatric liver transplant waitlist outcomes and to directly compare the performance of contemporary ML models with multivariable logistic regression for predicting waitlist mortality. This Research Report focuses on the performance and limitations of ML within the constraints of national registry data, rather than on specific preservation or perfusion technologies.

We performed a retrospective cohort study using de‐identified data from the UNOS/Organ Procurement and Transplantation Network database. Pediatric candidates aged 0–17 years listed for primary, single‐organ liver transplantation in the United States between January 1, 2010 and December 31, 2021 were included. Candidates listed for multi‐organ transplantation or re‐transplantation were excluded. For candidates with multiple listings, only the first listing was retained to standardize time‐to‐event analyses and reduce survivorship bias.

Because only de‐identified registry data were used, informed consent was not required and did not required Institutional Review Board approval.

Waitlist outcomes were categorized as: (1) receipt of liver transplantation, (2) death on the waitlist, (3) removal from the waitlist, or (4) remaining on the waitlist at the end of follow‐up (censored). For age‐based analyses, candidates were stratified by age at listing into 0–2 years (infants and toddlers), 3–10 years (early and middle childhood), and 11–17 years (adolescents), reflecting clinically relevant differences in graft size compatibility and disease progression. ² , ⁴

From an initial set of 434 candidate variables, 27 were retained based on clinical relevance, completeness, and distribution. Variables with substantial missingness, collinearity, or free‐text formatting were excluded. Free‐text fields were not analyzed because of heterogeneity and the absence of standardized preprocessing in the registry. MELD and PELD scores were not included in multivariable models to avoid collinearity with their component laboratory values. Age‐adjusted body surface area (BSA) Z‐scores were calculated using Centers for Disease Control growth references.

Categorical variables were compared using chi‐square tests. Continuous variables were non‐normally distributed and compared using Kruskal–Wallis tests with Dunn's post hoc testing and Bonferroni correction. A two‐sided p‐value < 0.05 was considered statistically significant.

Multivariable binary logistic regression was used to identify predictors of waitlist mortality, with death on the waitlist as the outcome and successful transplantation as the comparator. Least absolute shrinkage and selection operator regularization supported feature selection and mitigation of overfitting.

In parallel, ensemble ML models were trained to predict waitlist mortality using the same set of structured variables: Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Random Forest. Data were split into training (80%) and test (20%) sets, with five‐fold cross‐validation performed on the training set. Hyperparameters were optimized via grid search. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) in the held‐out test set. Analyses were conducted using Python (scikit‐learn, LightGBM, XGBoost) and R.

A total of 7864 pediatric liver transplant candidates met inclusion criteria. During the study period, 5975 candidates (76.0%) underwent transplantation, 1129 (14.4%) were removed from the waitlist, 383 (4.9%) died prior to transplantation, and 377 (4.8%) remained on the waitlist at the end of follow‐up. The median time on the waitlist was 61 days (interquartile range [IQR] 18–151). Among candidates who died on the waitlist, the median time to death was 72 days (IQR 17–245). The overall distribution of wait times was highly right‐skewed, with 353 candidates waiting longer than 500 days and a maximum observed wait time of 3442 days.

Waitlist mortality differed significantly by age group (chi‐square p < 2.2 × 10⁻¹⁶). Mortality was highest among candidates aged 0–2 years (5.35%), lowest among those aged 3–10 years (3.66%), and intermediate among adolescents aged 11–17 years (4.97%). Post hoc testing showed a statistically significant difference between the 0–2‐ and 3–10‐year groups (p = 0.0063), while differences between 0 and 2 versus 11–17 years (p = 0.57) and 3–10 versus 11–17 years (p = 0.06) were not significant.

Time to transplant also varied significantly across age groups (Kruskal–Wallis p = 0.0079). Adolescents (11–17 years) experienced significantly longer wait times than children aged 0–2 years (p = 0.0439) and 3–10 years (p = 0.0076), whereas wait times between the 0–2‐ and 3–10‐year groups did not differ significantly (p = 0.764). Among candidates who died on the waitlist, time to death differed by age (Kruskal–Wallis p = 0.0326), with infants and toddlers demonstrating a significantly shorter time to death than adolescents (p = 0.0326).

Logistic regression outperformed all evaluated ML models for prediction of waitlist mortality using structured registry variables. The logistic regression model achieved an AUC of 0.72 in the test set, whereas LightGBM, XGBoost, and Random Forest each yielded AUC values below 0.70.

In the final multivariable logistic regression model, independent predictors of waitlist mortality included higher age‐adjusted BSA Z‐score, the presence of ascites, ventilator dependence, hyperalimentation‐induced liver disease, and temporary inactive listing status. These predictors are consistent with established markers of disease acuity in pediatric liver transplant candidates. ³ , ⁴

In this national cohort of pediatric liver transplant candidates, contemporary ML models did not outperform traditional logistic regression for predicting waitlist mortality when both approaches were restricted to structured UNOS registry data. This negative finding is the most distinctive contribution of the study and underscores the importance of data quality and granularity when evaluating advanced analytic techniques. ⁷ , ⁸ , ⁹

Despite their theoretical ability to capture nonlinear relationships and higher‐order interactions, ML models are constrained by the same information limits as regression when applied to a finite set of relatively coarse variables. The UNOS registry lacks time‐varying physiologic data, detailed comorbidity profiles, and unstructured clinical narratives that have been shown to enhance ML performance in other healthcare domains. ⁵ , ⁶ , ⁷ , ⁸ , ⁹ In this setting, carefully specified regression with disciplined feature selection appears sufficient to extract most of the available information, leaving little room for ML to demonstrate incremental gains.

Clinically, our results confirm persistent age‐based disparities in pediatric waitlist outcomes that have been described in prior registry analyses. ¹ , ² , ⁴ Infants and toddlers experienced the highest mortality and shortest time to death, reflecting both physiologic vulnerability and scarcity of appropriately sized grafts. Adolescents, in contrast, faced significantly longer waits but lower mortality, likely due to difficulty finding adolescent‐sized livers, competition with adult candidates, and more indolent disease etiologies in some cases. ² , ³ These findings support age‐specific strategies rather than uniform adjustments in priority.

Several predictors identified in the regression model, ventilator dependence, ascites, hyperalimentation‐induced liver disease, and temporary inactive listing status, represent markers of advanced disease or modifiable aspects of clinical management. Inactive listing status has been associated with increased waitlist risk in prior studies and highlights the importance of timely reassessment and reactivation of candidates once transient contraindications resolve. ³ , ⁴

Although advanced organ preservation platforms and machine perfusion have generated substantial interest for expanding the donor pool, ² , ¹⁰ , ¹¹ our analysis predates widespread pediatric implementation and does not evaluate these technologies directly. Instead, the risk patterns described here can inform future trials and implementation efforts by clarifying which subgroups, especially infants and high‐acuity candidates, stand to benefit most from improved organ availability through split liver transplantation, living donation, or optimized use of marginal grafts. ² , ⁴ , ¹⁰ , ¹¹

In a 10‐year national cohort of pediatric liver transplant candidates, ML models did not outperform traditional logistic regression for predicting waitlist mortality when applied to structured UNOS registry data. Persistent age‐based disparities and identifiable clinical risk factors, particularly among infants and adolescents, continue to drive pediatric waitlist deaths. The most immediate opportunities to reduce mortality lie in early referral, proactive candidate management, careful attention to inactive listing status, and refined allocation and utilization strategies, while future work integrates richer data sources and evaluates pediatric‐specific applications of advanced analytic and preservation technologies (Figures 1 and 2).

Figure 1.

Distribution of time on the pediatric liver transplant waitlist. Histogram of wait time (days) for all pediatric liver transplant candidates listed in the United States between 2010 and 2021 (n = 7864). The distribution is truncated at 500 days for visualization; 353 candidates had wait times exceeding 500 days, and the maximum observed wait time was 3442 days. The median wait time was 61 days (interquartile range 18–151).

Figure 2.

Time to liver transplantation by age group. Box‐and‐whisker plots of time from listing to liver transplantation (days) stratified by age at listing: 0–2 years, 3–10 years, and 11–17 years. Boxes represent the interquartile range with median line; whiskers denote 1.5× the interquartile range. Overall comparison by Kruskal–Wallis test: p = 0.0079. Post hoc Dunn tests with Bonferroni correction demonstrated significantly longer wait times in adolescents compared with children aged 0–2 years (p = 0.0439) and 3–10 years (p = 0.0076); no significant difference was observed between the 0–2‐ and 3–10‐year groups (p = 0.764).

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.