2.1 Data Source

In this retrospective study, we used data from the OPTN. The OPTN data system includes data on all donors, waitlisted candidates, and transplant recipients in the US, submitted by the members of the OPTN. The Health Resources and Services Administration (HRSA), US Department of Health and Human Services, provides oversight to the activities of the OPTN contractor. The data reported here have been supplied by the United Network of Organ Sharing (UNOS) as the contractor for OPTN. The interpretation and reporting of these data are the responsibility of the authors and in no way should be seen as an official policy of or interpretation by the OPTN or the US Government. This study was approved by the institutional review board of the University of Virginia.

2.2 Study Population

The OPTN database was queried for all pediatric (< 18 years) heart transplant candidates listed between January 1, 2010 and December 31, 2020. Candidates with a prior heart transplant, transferred to another center, listed in error, remained on the waitlist at the time of data collection, or with an unknown waitlist outcome were excluded. For candidates who were added to the waitlist multiple times during this period, only the initial listing period was included for analysis.

2.3 Data Collection

Prior to modeling, 42 candidate-specific variables at the time of waitlist registration were selected from the OPTN dataset based on a review of existing literature and clinical expertise (Data S1). Estimated glomerular filtration rate (eGFR) was calculated for each candidate using the Schwartz formula. Listing center code was included as a variable in the model to assess whether this influences the model's predictive ability. Three additional listing center variables were derived from the dataset through feature engineering: listing center transplant volume (number of pediatric heart transplants each listing center performed the previous year), median number of offer refusals per candidate at each listing center, and the overall proportion of offers refused at each listing center for all candidates during this period. We included listing year as well as the implementation date of the most recent pediatric heart allocation policy change (July 7, 2016) to assess whether they influence the model's predictive ability.

2.4 Statistical Analysis/Machine Learning Modeling

A comparative analysis of baseline demographics was made between groups. Categorical variables were compared using Chi-squared or Fisher's exact tests. Continuous variables were compared using the Wilcoxon rank-sum test. Several statistical and machine learning models were compared for overall performance using Area Under the Receiver Operating Characteristic Curve (AUC-ROC) as the evaluation metric. Models assessed include Lasso (logistic regression with Lasso penalty), generalized additive model (GAM), XGBoost, Random Forests, and CatBoost.

Further analysis was conducted to characterize variables most predictive of waitlist mortality or clinical deterioration resulting in waitlist removal. We utilized a binary classification model with CatBoost due to its improved predictive ability compared to the other models tested, as well as its improved handling of categorical variables [12]. The total population was split into training (n = 4523) and testing (n = 1000) data sets. We used the Python Optuna package [13] to optimize and select the best parameters using the training data. Optuna uses Bayesian optimization to search for the parameters that provide the best cross-validated predictive performance. Based on Optuna's tuning, the final model contained: depth 3 trees, 223 iterations, learning rate 0.04, and ridge regularization 11.95. We fit the final model using all the training data with Optuna's selected parameters and used the test data to calculate variable importance scores via SHAP values [14]. The SHAP importance scores indicate which variables have the most influence in and association for predicting the outcome (i.e., waitlist mortality or removal for clinical deterioration). The complete code for the final model, along with a detailed modeling report following the TRIPOD+AI checklist [15], is available in the Data S1.

4.1 Listing Center Impact on Waitlist Mortality

Several studies have previously shown an association between candidate risk factors and waitlist mortality for pediatric patients awaiting heart transplantation [2, 3, 6, 10, 11], and others have highlighted the role listing center offer acceptance practice plays in waitlist outcomes [9, 16]. To our knowledge, this is the first study combining these potential risk factors to assess their relative impact on waitlist mortality or removal for clinical deterioration.

In a single center retrospective analysis, Hernandez et al. [16]. demonstrated that implementing institutional donor acceptance guidelines significantly reduced waitlist time, offer refusals for hearts with normal ejection fraction, and overall offer refusal rates, while also showing a trend toward decreased waitlist mortality or deterioration. In their subsequent analysis of the OPTN database, they showed that listing centers with high organ offer refusal rates had longer waitlist times, higher rates of death or deterioration, and were less likely to transplant candidates compared to centers with lower offer refusal rates [9]. Additionally, Rana et al. [17]. found that patients listed in low volume transplant centers (less than 3 transplants per year) had a significantly higher rate of waitlist mortality compared to those at high volume centers (more than 10 transplants per year).

Similarly, our study demonstrated that listing center offer acceptance rate is a strong predictor of waitlist outcomes, even when accounting for candidate risk variables. Listing center variables were among the most predictive variables along with established risk factors such as diagnosis, age/size, and end-organ dysfunction. Consistent with previous findings, our results indicate that a higher number of offer refusals per candidate, a greater proportion of offer refusals at a listing center for all candidates, and low transplant volume are associated with worse waitlist outcomes. Our group previously reported that organ acceptance is influenced by the listing center receiving the offer and the number of times the offer has been refused by other institutions [7, 8]. This current study suggests that variable organ acceptance practice among listing centers is associated with waitlist mortality, even when accounting for candidate specific risk factors. Considering the number of times a heart offer is refused for donor quality is not associated with post-transplant outcomes [18], these findings highlight a critical opportunity for improving offer acceptance practices within the transplant community in order to increase organ utilization and decrease waitlist mortality.

4.2 Modifiable Risk Factors

Management of pediatric patients with end-stage heart failure refractory to guideline-directed medical therapy is aimed at achieving the highest probability of long-term survival. Clinicians are often confronted with the difficult task of determining the optimal timing of waitlist activation once all medical and surgical options are exhausted, balancing the risk of delayed listing, which may contribute to negative outcomes. Once a patient is listed for transplantation, there remains a role for further risk reduction by optimizing candidate risk factors.

In an analysis of the OPTN database utilizing machine learning to assess variables associated with waitlist mortality in adult candidates, Hsich et al. [19]. demonstrated low serum albumin level and eGFR to be strongly associated with waitlist mortality. Similarly, our results show modifiable risk factors such as malnutrition (low serum albumin, lower body mass index) and renal dysfunction (low eGFR, need for dialysis) are important predictors of waitlist mortality in the pediatric population. This emphasizes the importance of optimizing fragile patients through targeted interventions, including ventricular assist devices, which have been shown to improve functional status [20], nutritional status [21], and waitlist outcomes in select populations, especially if waitlist duration is expected to be prolonged [2, 22-24].

Unfortunately, as shown in previous studies, our model identified African American race to be predictive of poor waitlist outcomes, whereas white race was associated with positive outcomes—underscoring the continuing health disparities that exist in this population of patients.

4.3 Organ Allocation Policy

The current OPTN pediatric heart allocation policy introduced in 2016 aimed to reduce waitlist mortality by prioritizing patients with the highest clinical acuity based on previously identified risk factors [25]. These policy changes reduced urgency status for dilated cardiomyopathy patients on inotropic support, whereas congenital heart disease patients on inotropic support remained at the highest urgency status. Additionally, it removed primary allocation by blood type for ABO incompatible listings to reduce wait times for infants. However, this policy still only stratifies patients into three tiers (Status 1A, Status 1B, Status 2), with the opportunity to submit an exception request for higher acuity listing.

Listing status was not included as a variable in our model in order to exclusively test the relationship between candidate/listing center risk variables to waitlist outcome. However, our modeling did not identify candidate listing year or listing prior to vs. after the policy change as correlates to waitlist outcome (Figure 2). Power et al. [11]. recently showed the ongoing limitations of the allocation policy, which fails to adequately capture the mortality risk for patients within the same listing status group and even between listing status groups. Our results highlight the interplay of multiple candidate variables, each contributing varying degrees of risk. This supports a revision of the current allocation policy to one that allows for a more granular assessment of waitlist mortality risk, such as would be found in a continuous distribution system.