2.1 Data source

This retrospective study utilized donor data from the Organ Procurement and Transplant Network (OPTN) which includes data on all donors, waitlisted candidates, and transplant recipients in the US, submitted by the members of OPTN. The Health Resources and Services Administration (HRSA), U.S. Department of Health and Human Services provides oversight of 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 authors' responsibility and in no way represent an official policy of or interpretation by the OPTN or the US Government. This study was approved by the University of Virginia's (UVA) Institutional Review Board (IRB). Both the OPTN and UVA consider research using decedent data as ‘non-human research’ and do not require full IRB approval.

2.2 Study population and data collected

Data were evaluated on all pediatric donors (<18 years of age at the time of declaration of death by neurologic criteria (DDNC)) whose hearts were offered to pediatric patients (<18 years of age at the time of offer) on the waitlist for heart transplantation between January 1, 2010 and December 31, 2019. Donors following declaration of cardiac death, offers made to a candidate who had already accepted another organ, and candidates requiring heart-lung or retransplantation were excluded. In addition to the de-identified data available within the provided DonorNet database files, all attachments (PDF format) associated with the donor's terminal hospitalization were also provided for evaluation. These attachments included, but were not limited to: cardiology consultation forms, echocardiographic (echo) interpretations both within the native echo interpretation software and subsequent electronic medical record formats, and all formal donor documentation produced by the Organ Procurement Organization (OPO) personnel responsible for donor management. PDF data extraction focused entirely on echo-related documentation, specifically: study date and time; qualitative assessments of right ventricular systolic function; objective and qualitative assessments of left ventricular systolic function; objective assessments of tricuspid valve insufficiency gradients; qualitative assessments of atrioventricular and semilunar valve insufficiency; and any structural abnormalities, as outlined below. Data were extracted by MM, SW, and FH and a team of 24 undergraduate University of Virginia students personally instructed on the technique by MM prior to any data extraction. After initial instruction, MM reassessed extraction accuracy by reviewing the first 50 donor files of each student and provided repeat training as necessary. MM continued to over-read studies until satisfactory performance was achieved by every student. Data were entered into a REDCap database produced and maintained by UVA. REDCap (research electronic data capture) is a secure, web-based application designed to support data capture for research studies, providing: (1) an intuitive interface for validated data entry; (2) audit trails for tracking data manipulation and export procedures; (3) automated export procedures for seamless data downloads to common statistical packages; (4) procedures for importing data from external sources. The REDCap echo database was subsequently merged with the DonorNet database using individual DONOR IDs.

Data were collected on the donors, candidates, offers, listing centers, donor hospitals, and OPOs. Static donor data included gender, race, weight, height, date and time of DDNC, cause of death, duration of cardiopulmonary resuscitation (CPR), center for disease control (CDC) high-risk status, and blood type. Dynamic donor variables include the time since DDNC, vital signs, inotrope administration, and echocardiogram reports. Candidate variables included status, diagnosis, calculated panel reactive antibody (cPRA), weight, height, gender, race, blood type and need for mechanical ventilation, extracorporeal membrane oxygenation (ECMO), or ventricular assist device (VAD). The donor/candidate weight ratio and age differences were also included. Offer variables included distance between donor hospital and transplant center, offer time of day, and the number of prior donor refusals. We also calculated listing centers' prior year number of pediatric heart transplants, prior year acceptance rate by candidate status, and prior 18-month adverse outcome rate (defined as the censoring adjusted proportion of transplants occurring in the last 18 months that required re-transplant or resulted in patient death within 1 year). The Organ Procurement Organization's (OPO) prior year number of pediatric donors managed and donor hospital's prior year number of pediatric donors produced were also included. Since we did not have offers from 2009, 2010 data were utilized as a proxy for the prior year's values.

2.3 Echocardiographic data

We collected date and time of donor echocardiogram reports as well as qualitative and quantitative assessments of right and left ventricular function and both semilunar and atrioventricular valve function. An echocardiogram received a composite categorization of ‘Normal’ if: qualitative assessments of right and left ventricular wall motion were considered normal AND if present, ejection fraction measurements were ≥55% and/or shortening fraction was ≥28%; AND no evidence of left ventricular segmental wall motion abnormalities (LVSWMA); AND no more than mild insufficiency of either semilunar or atrioventricular valve; all other studies were considered ‘Abnormal’. Any structural abnormalities aside from atrial septal defects were noted when relevant (e.g., bicuspid aortic valves, ventricular septal defects) but were not considered in the composite endpoint. For a more detailed description of the echocardiographic data assessment, see Appendix S1.

2.4 Primary outcome

The primary outcome for each offer was whether a pediatric heart transplant candidate's institution decided to ‘accept’ or ‘decline’. A heart accepted for the first candidate on a donor list was determined to have produced 1 offer and if accepted by the 10th candidate, that donor heart generated 10 different offers. In both cases, the final offer was considered ‘accepted’ and the donor heart ‘utilized’ by a candidate in our study population. If a donor heart was never accepted by a candidate in our study population during the study period, that organ was considered ‘non-utilized’. It is possible that non-utilized organs were eventually accepted for a candidate outside of our study population (e.g., adult or Canadian candidate) or after December 31, 2019. It is also possible that an accepted organ was never procured or transplanted.

2.5 Statistical analysis and mathematical modeling

Descriptive statistics utilized median and interquartile ranges for continuous and proportions for categorical variables, respectively. Comparisons between groups utilized Wilcoxon rank sum or chi-square as appropriate. Considering the large number of variables assessed, statistical significance was assessed with a p-value <.01.

Multivariable logistic regression (LR) and random forest (rf) models were fit to determine patterns associated with the decision to ‘Accept’ versus ‘Decline’ an offer. Both models used the same 44 predictor variables that were available to decision-makers at the time of offer. Median imputation was used for numeric predictors and an ‘Unknown’ category was added to categorical predictors with missing values.⁷ Predictive performance and variable importance scores were evaluated using stratified Monte Carlo cross-validation. Stratification of offers was based on the listing center transplant rates. Predictions on the hold-out offers were used to compute the out-of-sample area under the ROC curve (AUROC). SHAP (SHapley Additive exPlanations) values were used to measure variable importance.^{8, 9} Additional details are provided in Appendix S1.

3.1 Donor demographics

During the study period, hearts from 5625 pediatric donors resulted in 30 156 offers to 4905 pediatric heart transplant candidates in the US. Of these offers, 88.7% (26 738/30 156) were Declined and there was a final Non-Utilization rate of 39.2% (2207/5625). Table 1 demonstrates donor demographics both for individual offers and the final decision. Donor age demonstrated marked differences in acceptance practices; donors <2 years resulted in the highest utilization rates as compared to a significant, progressive decline in utilization with increasing age. Though individual offer decisions were affected by blood type, final Utilization rates were only slightly higher for O blood type donors (p = .02). Anoxia, head trauma, or stroke were the cause of death listed in 97% of donors. CNS tumor and infection of any type (CNS or ‘general’) were routinely declined and collectively utilized from only 38% of donors, which was less common than all other causes of death (p < .0001). History of CPR was present in 58% of donors and although it did decrease offer acceptance, it did not negatively impact Utilization. The median CPR duration was 24 (12, 40) minutes and there was no statistically detectable difference between whether or not a heart was utilized (utilized 25 (12, 40) minutes versus non-utilized 24 (12, 39), p = .6).

The maximum vasoactive inotropic score prior to the current offer was higher for offers that were declined (declined 15 (5, 40) versus accepted 10 (3, 29.8); p < .001), but this calculated variable was not considered for the predictive models as it is not presently available in DonorNet. Maximum troponin-I values were lower in accepted/utilized (0.1 (0, 0.6)) versus both declined offers (0.3 (0.1, 1.4); p < .001) and non-utilized donors (0.4 (0.1,1.9); p < .001).

3.2 Echocardiographic data

A total of 7810 echocardiograms were evaluated from the 5394 donors with echocardiographic data. Each donor underwent a median of 1 (1, 2) echocardiograms during their terminal hospitalization, with the first occurring 10.8 h (2.1, 20.2) and the final 18.1 h (8.5, 25.3) after DDNC. Composite assessments of ‘normal’ were given to 79% of all final studies in donors with echocardiograms, comprising 89.7% (2943/3280) of all utilized donors and 62.9% (1330/2114) of non-utilized donors (Table 2). Of the 14 287 offers Declined using refusal code 830 (donor age or quality) and 2376 offers Declined using refusal code 837 (organ-specific donor issues), 48% and 34% had normal echocardiograms, respectively.

A total of 63 donors demonstrated echocardiographic abnormalities aside from ventricular function or valvar insufficiency, of which 33 were utilized for transplant. Findings of subaortic stenosis, a left ventricular mass, possible rhabdomyosarcoma, prosthetic aortic valve, situs inversus totalis, a history of an ASD repair, history of a VSD repair, possible left ventricular non-compaction cardiomyopathy (4), pulmonary valve stenosis (2) and a right ventricular mass (2) were found in a total of 15 donors, which all were in the non-utilized group. Conversely, 4 different donors in the utilized group demonstrated evidence of a single right coronary artery, prominent left coronary artery, pulmonary artery sling, and circumflex artery arising from the right coronary artery. Of 5 donors on ECMO, 2 were utilized, as were 8/10 donors with small muscular ventricular septal defects, 4/5 with ‘echobright regions’ demonstrated within the left ventricle, 5/7 described as having left ventricular hypertrophy, 2/3 with a ‘right atrial mass’, 2/4 with a ‘mitral valve abnormality’, and 6/10 with a bicommissural aortic valve.

3.3 Offer data

Figure 1 demonstrates the univariate relationships between select donor/candidate data and likelihood to accept. When comparing accepted versus declined offers, age difference and weight ratios were statistically but not clinically different, though there were clear inflection points when donors and candidates had similar ages and weights (Figure 1A,B). Figure 1C demonstrates acceptance inflection points at approximately 250 and 1000 nautical miles, coinciding with two of the four geographical location zones impacting pediatric donor allocation (https://optn.transplant.hrsa.gov/media/eavh5bf3/optn_policies.pdf). The median distance between donor and candidate centers was significantly shorter for accepted offers. Figure 1D demonstrates the marked decline in the likelihood of offer Acceptance with increasing numbers of prior offer refusals.

3.4 Candidate data

Table 3 demonstrates baseline demographics for the 4905 candidates. Candidate race demonstrated a significant intragroup difference for offer Acceptance rates and post hoc analyses demonstrated a trend toward less Black candidates being transplanted as compared to all other races (p = .02). Age <2 years was associated with a higher rate of offer Acceptance but not final transplantation rates as compared to all other age ranges (p = .07), whereas candidates aged >11 were associated with lower offer Acceptance and transplantation rates. Transplantation rates trended lower for candidates with O blood type (p = .011) but higher for those with AB blood type (p = .038). Candidates with cardiomyopathy as their listing diagnosis experienced higher offer Acceptance and transplantation rates than those with congenital heart disease. Of the 1775 candidates with a documented cPRA, 38% (682/1775) were ≥10%; of these, 57% listed CHD as their listing diagnosis.

3.5 Modeling results

Table 4 demonstrates the predictive performance of the LR and RF models using different dataset variables. Using donor, candidate, and offer variables, both models demonstrate good predictive performance of likelihood to Accept (AUROC ≥0.83). The performance difference between models is not statistically significant (p = .09), suggesting the added flexibility of RF to incorporate non-linear and interaction effects did not substantially improve prediction. Limiting the dataset to include only candidate or donor variables markedly decreased predictive performance.

To assess the importance of each predictor variable on the estimated probability of offer acceptance, we produced SHAP variable importance scores. Figure 2 demonstrates the 10 variables with median importance scores >0.01 for either the RF or LR models. The number of prior donor refusals and our derived variable assessing a listing center's prior year acceptance rates were the two most important variables predicting offer acceptance practices, followed closely by whether an echocardiogram was interpreted as normal and distance between donor and candidate centers. As a prior number of refusals was a stronger predictor using the more flexible RF modeling, we evaluated its independent influence on acceptance probability. Figure 3 demonstrates that on initial offers, where the number of prior refusals is 0, acceptance probability increases between 5 (LR) and 8% (RF) as compared to a donor on its 10th offer, in which the likelihood of acceptance has decreased between 7% and 12%. This is further supported by the finding that both RF and LR models lost acceptance practice predictive performance when using only first offers, as seen in Table 4, but largely regain it when using a random subset of subsequent offers, allowing the reintroduction of prior decision bias back into the model.

4.1 Limitations

Our study has two important sets of limitations. The first relates to echocardiographic data. No studies were found for 4.5% of donors. As it is highly unlikely institutions would utilize a heart without an echocardiogram and over 60% of these donors, or 4.7% of all donors whose hearts were ultimately accepted did not have a documented echocardiogram, it is almost certain these studies were simply missed or not available, which also suggests studies may be missing for other donors. Additionally, all echocardiogram interpretations are taken directly from the local cardiologist's interpretations. Several studies have demonstrated marked discordance in functional assessments when local studies are reviewed by an experienced core lab,^{22, 23} but this was not possible for our study.

The second set of limitations relates to the challenges of analyzing and modeling observational databases, aside from the inherent issues of data reliability and missing values. We limited our population to pediatric donors offered to pediatric candidates, necessarily removing any information regarding donors accepted for US adult or Canadian pediatric candidates, which occurred in almost 40% of the 2207 pediatric donors not utilized for a US pediatric candidate. Secondly, it is important to note that the SHAP importance scores reflect an aggregate acceptance practice. Variables with high importance do not necessarily imply that all centers consider them strongly and conversely, variables with lower importance scores may be strongly utilized by other centers. Finally, after careful consideration of variable correlation and collinearity, we chose to model 44 of the more than 150 different donor, candidate and offer specific variables and as a result, may not have included variables some institutions are using for decision-making; conversely, this should have helped minimize the risk of model overfitting.

In conclusion, our data confirm prior findings that pediatric heart donor pool utilization remains low in the United States. For the first time, however, we have demonstrated both the potential viability of those organs being discarded and the negative downstream influence institutions' decisions have on subsequent offers. We believe more research is needed to assess the utility of removing this information on both donor acceptance practices and post-transplant outcomes.