Performance of risk prediction models for post-liver transplant patient and graft survival over time

Abstract

Given liver transplantation organ scarcity, selection of recipients and donors to maximize post-transplant benefit is paramount. Several scores predict post-transplant outcomes by isolating elements of donor and recipient risk, including the donor risk index, Balance of Risk, pre-allocation score to predict survival outcomes following liver transplantation/survival outcomes following liver transplantation (SOFT), improved donor-to-recipient allocation score for deceased donors only/improved donor-to-recipient allocation score for both deceased and living donors (ID2EAL-D/-DR), and survival benefit (SB) models. No studies have examined the performance of these models over time, which is critical in an ever-evolving transplant landscape. This was a retrospective cohort study of liver transplantation events in the UNOS database from 2002 to 2021. We used Cox regression to evaluate model discrimination (Harrell’s C) and calibration (testing of calibration curves) for post-transplant patient and graft survival at specified post-transplant timepoints. Sub-analyses were performed in the modern transplant era (post-2014) and for key donor-recipient characteristics. A total of 112,357 transplants were included. The SB and SOFT scores had the highest discrimination for short-term patient and graft survival, including in the modern transplant era, where only the SB model had good discrimination (C ≥ 0.60) for all patient and graft outcome timepoints. However, these models had evidence of poor calibration at 3- and 5-year patient survival timepoints. The ID2EAL-DR score had lower discrimination but adequate calibration at all patient survival timepoints. In stratified analyses, SB and SOFT scores performed better in younger (< 40 y) and higher Model for End-Stage Liver Disease (≥ 25) patients. All prediction scores had declining discrimination over time, and scores relying on donor factors alone had poor performance. Although the SB and SOFT scores had the best overall performance, all models demonstrated declining performance over time. This underscores the importance of periodically updating and/or developing new prediction models to reflect the evolving transplant field. Scores relying on donor factors alone do not meaningfully inform post-transplant risk.

INTRODUCTION

Organ scarcity in liver transplantation (LT) necessitates optimal selection of recipients and donors to minimize waitlist mortality and maximize post-transplant benefit.¹ Several scores have been developed to predict post-transplant patient and graft survival by isolating elements of donor risk, recipient risk, or both. The donor risk index (DRI), developed in 2006, focuses on 7 key donor characteristics associated with graft failure (ie, time to death or retransplantation) and is widely used in clinical and research applications.² More recent attempts to optimize the prediction of post-LT outcomes in the evolving allocation landscape led to the development of the Balance of Risk score,³ pre-allocation score to predict survival outcomes following liver transplantation (P-SOFT) and score to predict survival outcomes following liver transplantation (SOFT) scores,⁴ the survival benefit (SB) model, and most recently the improved donor-to-recipient allocation score for deceased donors only (ID2EAL-D)/improved donor-to-recipient allocation score for both deceased and living donors (ID2EAL-DR) scores.⁵ However, there is no consensus on key donor and recipient factors to consider nor the optimal risk prediction models to use. Furthermore, given that the transplant landscape has evolved substantially over the past decades, it is plausible that prediction model performance has changed with time. To address these questions, we evaluated the discrimination and calibration of LT risk scores over time and in the modern transplant era, with stratifications by key donor-recipient characteristics.

METHODS

Study design and cohort creation

This was a retrospective cohort study using the United Network for Organ Sharing (UNOS) database. Adult patients who received LT, including retransplantation, between January 2002 and December 2021 were included. Patients who underwent multiple organ transplants were excluded. Key recipient and donor characteristics at the time of transplant were extracted for each transplant event. For the recipient, this included age, gender, race, body mass index, etiology of liver disease, Model for End-Stage Liver Disease (MELD) at transplant, HCC exception case, prior liver transplant history, prior abdominal surgery, albumin, dialysis status at transplant, intensive care unit status pretransplant, life support pretransplant, encephalopathy, ascites, and PVT. For the donor, this included age, height, weight, donation after cardiovascular death (DCD) versus donation after brain death (DBD), organ location, cause of death, partial/split graft, creatinine, and insulin-dependent diabetes status. The year of transplant was also recorded.

Calculation of post-transplant risk scores

For each transplantation event, we computed the following clinician-facing risk prediction scores: DRI, P-SOFT, SOFT, Balance of Risk, ID2EAL-D, and ID2EAL-DR. We also computed scores from the SB model, for which no calculator is currently available, published by Schaubel et al.⁶ This entailed computing the linear predictor derived from the log hazards of the survival benefit model detailed in the study. Recipient and donor characteristics serving as inputs for each score are summarized in Table 1.

TABLE 1.

Summary of post-transplant prediction models and variable components

	DRI	P-SOFT	SOFT	Survival Benefit Model	Balance of Risk (BAR)	ID2EAL-D	ID2EAL-DR
References	Feng et al, 20062	Rana et al, 20084	Rana et al, 20084	Schaubel et al, 20096	Dutkowski et al, 20113	Asrani et al, 20225	Asrani et al, 20225
Years of included data	1998–2002	2002–2006	2002–2006	2001–2007	2002–2010	2013–2019	2013–2019
Operative/allocation variables	CIT, organ location	CIT	CIT, organ location	—	CIT	—	CIT
Donor variables	Age, COD, race, DCD, partial/split graft, height	—	Age, COD from CVA, creatinine>1.5	Age, race/ethnicity, COD from anoxia, COD from CVA, DCD	Age	Age, height, weight, COD, DCD, insulin-dependent diabetes, serum creatinine	Age, height, weight, COD, DCD, insulin-dependent diabetes, serum creatinine
Recipient variables	—	Age, BMI, 1 previous tx, 2 previous tx, previous abdominal surgery, albumin < 2.0, dialysis before tx, ICU pre-tx, MELD > 30, life support pre-tx, encephalopathy, PVT, ascites pre-tx	P-SOFT recipient variables plus: Portal bleed in the 48 hours before transplanta	Age, liver-related diagnosis, malignancy history, HCC, diabetes, dialysis, ICU pre-tx, hospitalization pre-tx without ICU, previous tx, life support, previous abdominal surgery, PVT, creatinine, albumin	Age, MELD score, retransplantation, life support	—	Age, MELD-Na score

^aNote: This variable is not captured in the UNOS database and thus is not incorporated in the score calculations in this study.

Abbreviations: BMI, body mass index; CIT, cold ischemia time; COD, cause of death; CVA, cerebrovascular accident; DCD, donation after circulatory death; DRI, donor risk index; ICU, intensive care unit; INR, international normalized ratio; MELD, Model for End-Stage Liver Disease; P-SOFT, pre-allocation score to predict survival outcomes following liver transplantation; tx, transplant; UNOS, United Network for Organ Sharing.

Statistical analysis

Descriptive statistics were presented as medians and IQRs for continuous variables and as frequencies and percentages for categorical variables. To evaluate model performance for patients who underwent post-transplant and graft survival over time, we fit univariable Cox regression models for each score in each transplant year. To evaluate model discrimination, we computed the Harrell’s C statistic at 1, 3, and 5-year post-transplant timepoints for post-transplant patient survival. For graft survival, we evaluated 90-day graft survival and 1-year graft survival conditional on graft survival to 90 days (ie, conditional graft survival), in line with current Scientific Registry of Transplant Recipients reporting metrics.⁷ For each outcome and time horizon, Harrell’s C values were plotted for each transplant year to display temporal trends in model performance. To visualize changes in prediction score performance for five-year post-transplant mortality by era, we also computed Harrel’s C statistics for the following periods: 2002–2004 (early MELD era), 2005–2013 (incorporates effects of regional share 15 and regional share for status 1 patients), and 2014–2021 (incorporates effects of regional share 35 and wide availability of direct-acting antiviral therapies for HCV infection).^8,9

To provide more in-depth prediction model performance in the modern transplant era, we performed subsequent analyses restricted to patients transplanted from 2014 onward. Discrimination through Harrell’s C was reported for each score for both patient and graft failure outcomes, and calibration was assessed by fitting a regression line through the calibration curve (a plot of observed vs. expected values). The beta coefficients for the intercept and slope of this line along with 95% CIs were reported for each score and follow-up timepoint, and a joint hypothesis test was performed to evaluate the null hypothesis of intercept = 0 and slope = 1. As an example, a poorly calibrated model could have a slope significantly > 1, which would indicate an underestimation of risk, or a slope significantly < 1, which would indicate an overestimation of risk. For all hypothesis tests, an alpha threshold of 5% was used to determine statistical significance. Finally, to identify transplant scenarios where the performance of scores might vary, we evaluated model discrimination across several a priori specified subgroups of interest. This included stratifications by recipient age (< 40, 40–59, ≥ 60 y), MELD at transplant (< 15, 15–24, ≥ 25), and DBD versus DCD donor. The Harrell’s C values were plotted for each score across these subgroups for 1-year post-transplant patient mortality, 90-day graft failure, and 1-year conditional graft failure.

Other consideration

This study was deemed Institutional Review Board exempt as the data are deidentified and publicly available. All data management and analyses were performed using STATA/BE 17.0 (College Station, TX).

RESULTS

Cohort characteristics

From 2002 to 2021, a total of 112,357 transplants were included in the study cohort (Table 2). The median recipient age was 56 years (IQR: 49, 62), the recipient cohort was 71.7% White, and with a median body mass index of 28.2 (IQR: 24.7, 32.4). Median MELD at transplant was 21 (IQR: 14, 30). The most common etiology of liver disease was HCV (32.7%), followed by alcohol-associated liver disease (20.7%) and NAFLD (17.5%). The median donor age was 42 years (IQR: 27, 55) and 6.2% of transplants were from DCD donors. The volume of transplants increased from 4266 in 2002 to 7400 in 2021.

TABLE 2.

Overall cohort characteristics

	Total N=112,357
Recipient age (y)	56 (49, 62)
Sex, n (%)
Female	37,495 (33.4)
Male	74,862 (66.6)
Race, n (%)
White	80,517 (71.7)
Black	9,740 (8.7)
Hispanic	15,708 (14.0)
Asian	4,958 (4.4)
Other	1,434 (1.3)
Recipient BMI	28.2 (24.7, 32.4)
Etiology of liver disease, n (%)
HCV	36,770 (32.7)
Alcohol	23,266 (20.7)
HBV	3,420 (3.0)
NAFLD	19,703 (17.5)
Primary Sclerosing Cholangitis	5,474 (4.9)
Autoimmune	3,471 (3.1)
Other	16,886 (15.0)
Primary Biliary Cholangitis	3,367 (3.0)
Laboratory MELD at transplant	21 (14, 30)
HCC Exception case, n (%)	23,711 (21.1)
Donor age (y)	42 (27, 55)
DCD donor, n (%)	6,978 (6.2)
Transplant year, n (%)
2002	4,266 (3.8)
2003	4,617 (4.1)
2004	5,008 (4.5)
2005	5,241 (4.7)
2006	5,413 (4.8)
2007	5,198 (4.6)
2008	5,118 (4.6)
2009	5,186 (4.6)
2010	5,089 (4.5)
2011	5,175 (4.6)
2012	5,054 (4.5)
2013	5,180 (4.6)
2014	5,372 (4.8)
2015	5,586 (5.0)
2016	6,208 (5.5)
2017	6,403 (5.7)
2018	6,602 (5.9)
2019	7,111 (6.3)
2020	7,130 (6.3)
2021	7,400 (6.6)

Note: Data are presented as median (IQR) for continuous measures and n (%) for categorical measures.

Abbreviations: BMI, body mass index; DCD, donation after cardiovascular death; MELD, Model for End-Stage Liver Disease.

Prediction of post-transplant patient survival

The SB model demonstrated the highest discrimination for patient survival at 3- and 5-year post-transplant timepoints, and the SB and SOFT models had generally similar high discrimination at the 1-year post-transplant timepoint (Figure 1A/B/C). All scores demonstrated declining performance over time and with longer post-transplant follow-up horizons. The DRI and ID2EAL-D scores, relying on donor factors alone, had declining and poor discrimination (C < 0.60) for post-transplant patient survival across all follow-up periods. Similar trends in discrimination were apparent when evaluated across transplant eras (2002–2004, 2005–2013, 2014–2021); the DRI and ID2EAL-D sores had declining and poor discrimination, whereas the SB model followed by the SOFT model had the highest discrimination among all scores; only the SB model demonstrated C > 0.60 in all eras for 5-year post-transplant patient survival (Figure 1D).

FIGURE 1.

Discrimination of post-transplant prediction models by transplant year for post-transplant patient mortality and graft failure. Abbreviations: BAR, Balance of Risk; DRI, donor risk index; ID2EAL-D, improved donor-to-recipient allocation score for deceased donors only; ID2EAL-DR, improved donor-to-recipient allocation score for both deceased and living donors; P-SOFT, pre-allocation score to predict survival outcomes following liver transplantation; SOFT, survival outcomes following liver transplantation; SB, survival benefit.

In the modern transplant era (2014 onward), the SB model had the highest discrimination for post-transplant survival at 1, 3, and 5 years (each with C > 0.60), followed by the SOFT and P-SOFT scores. These scores, however, had evidence of poor calibration at the 3- and 5-year timepoints (joint p < 0.05) such that these scores overestimated the risk of post-transplant mortality (Table 3). The ID2EAL-DR score had lower discrimination than SB, SOFT, and P-SOFT models but offered the best balance of discrimination and adequate calibration across all follow-up time horizons. Finally, while the DRI and ID2EAL-D scores were adequately calibrated at all follow-up timepoints, discrimination was very poor (C < 0.55).

TABLE 3.

Discrimination and calibration of prediction scores for post-transplant patient mortality and graft failure in the modern transplant era (Post-2013)

	Post-transplant patient mortality			Post-transplant graft failure
	One-year patient mortality	Three-year patient mortality	Five-year patient mortality	Ninety-day graft failure	One-year conditional graft failure
SB post-transplant model
Discrimination (Harrell’s C)	0.63	0.61	0.61	0.60	0.62
Intercept (95% CI)	−0.00 (−0.04, 0.03)	0.01 (−0.02, 0.03)	0.04 (0.02, 0.07) a	0.00 (−0.04, 0.04)	0.00 (−0.05, 0.05)
Slope (95% CI)	0.99 (0.92, 1.07)	0.92 (0.85, 0.99) a	0.86 (0.79, 0.93) a	1.05 (0.94, 1.17)	0.97 (0.85, 1.09)
Joint Test (p-value)	0.99	0.046 a	<0.001 a	0.64	0.86
DRI
Discrimination (Harrell’s C)	0.53	0.53	0.53	0.54	0.55
Intercept (95% CI)	−0.00 (−0.04, 0.03)	0.01 (−0.02, 0.03)	−0.01 (−0.03, 0.02)	0.00 (−0.04, 0.04)	0.00 (−0.05, 0.04)
Slope (95% CI)	0.95 (0.61, 1.29)	0.99 (0.73, 1.25)	0.81 (0.54, 1.09)	0.96 (0.64, 1.27)	0.93 (0.65, 1.21)
Joint test (p-value)	0.95	0.92	0.35	0.97	0.89
ID2EAL score-D
Discrimination (Harrell’s C)	0.53	0.52	0.52	0.54	0.56
Intercept (95% CI)	−0.00 (−0.03, 0.03)	0.01 (−0.02, 0.03)	0.00 (−0.03, 0.02)	0.00 (−0.04, 0.04)	0.00 (−0.05, 0.05)
Slope (95% CI)	0.98 (0.64, 1.32)	0.93 (0.61, 1.24)	0.79 (0.46, 1.11)	0.95 (0.70, 1.20)	0.99 (0.78, 1.19)
Joint test (p-value)	0.99	0.75	0.42	0.91	0.99
ID2EAL score-DR
Discrimination (Harrell’s C)	0.60	0.58	0.58	0.58	0.60
Intercept (95% CI)	0.02 (−0.02, 0.05)	0.01 (−0.01, 0.04)	0.00 (−0.03, 0.02)	0.01 (−0.03, 0.05)	0.01 (−0.04, 0.06)
Slope (95% CI)	1.04 (0.89, 1.18)	1.01 (0.88, 1.14)	0.94 (0.81, 1.08)	0.79 (0.55, 1.03)	0.69 (0.45, 0.94) a
Joint test (p-value)	0.56	0.62	0.70	0.18	0.04 a
BAR
Discrimination (Harrell’s C)	0.58	0.56	0.55	0.57	0.54
Intercept (95% CI)	0.00 (−0.03, 0.03)	0.00 (−0.03, 0.03)	−0.01 (−0.04, 0.01)	0.01 (−0.03, 0.04)	0.00 (−0.05, 0.05)
Slope (95% CI)	1.15 (1.03, 1.27) a	0.89 (0.74, 1.04)	0.71 (0.53, 0.89) a	1.20 (1.04, 1.35)	1.08 (0.73, 1.43)
Joint test (p-value)	0.02 a	0.36	0.01 a	0.04 a	0.90
P-SOFT
Discrimination (Harrell’s C)	0.62	0.59	0.59	0.61	0.57
Intercept (95% CI)	−0.01 (−0.04, 0.03)	0.00 (−0.03, 0.02)	−0.02 (−0.04, 0.01)	0.00 (−0.04, 0.04)	0.00 (−0.05, 0.05)
Slope (95% CI)	0.98 (0.91, 1.06)	0.89 (0.81, 0.97) a	0.78 (0.69, 0.88) a	1.02 (0.93, 1.12)	1.03 (0.84, 1.22)
Joint test (p-value)	0.87	0.03 a	<0.001 a	0.88	0.96
SOFT
Discrimination (Harrell’s C)	0.63	0.60	0.59	0.62	0.58
Intercept (95% CI)	−0.01 (−0.04, 0.03)	−0.01 (−0.04, 0.02)	−0.02 (−0.05, 0.00)	0.00 (−0.04, 0.04)	0.00 (−0.05, 0.05)
Slope (95% CI)	0.97 (0.90, 1.05)	0.88 (0.81, 0.96)a	0.78 (0.69, 0.87)a	1.00 (0.91, 1.09)	1.01 (0.84, 1.19)
Joint test (p-value)	0.75	0.01a	<0.001a	0.99	0.98

Note: Calibration was evaluated by joint hypothesis testing of the intercept and slope beta coefficients being equal to 0 and 1, respectively, from calibration curves.

^aStatistically significant at the alpha = 0.05 level.

Abbreviations: BAR, Balance of Risk; ID2EAL-D, improved donor-to-recipient allocation score for deceased donors only; ID2EAL-DR, improved donor-to-recipient allocation score for both deceased and living donors; P-SOFT, pre-allocation score to predict survival outcomes following liver transplantation; SOFT, survival outcomes following liver transplantation; SB, survival benefit.

Prediction of post-transplant graft survival

The SOFT score demonstrated the highest discrimination over time for 90-day graft survival, followed by the SB model and ID2EAL-DR score (Figure 1E). All models demonstrated declining performance over time. However, the SB model had the highest discrimination of all models for conditional 1-year graft survival, and performance remained generally consistent over time (Figure 1F). In the modern transplant era (2014 onward), all models were adequately calibrated with the exception of the ID2EAL-DR score, which tended to underestimate 90-day graft failure and the ID2EAL-D score, which tended to overestimate 1-year conditional graft failure (Table 3). The SOFT score demonstrated the highest discrimination for 90-day graft survival (C = 0.62), whereas the SB model had the highest discrimination for 1-year conditional graft survival (C = 0.63). The SB model was the only score with good discrimination (C ≥ 0.60) and adequate calibration for both graft outcomes.

Subgroup analyses

In subgroup analyses in the modern transplant era (2014 onward), discrimination for 1-year patient mortality and 90-day graft failure was notably higher for SB, SOFT, and P-SOFT scores in recipients aged < 40 years or MELD ≥ 25 (Figure 2). For both of these outcomes, the SOFT score yielded the highest discrimination of all scores in subgroup analyses of DBD donors and DCD donors. For 1-year conditional graft failure, the SB model generally demonstrated the highest discrimination of all scores and had fairly consistent performance across subgroups, with the exception of DCD donors, where the ID2EAL-DR score had the highest discrimination.

FIGURE 2.

Subgroup analyses of model discrimination for 1-year post-transplant patient mortality, 90-day graft failure, and 1-year conditional graft failure in the modern transplant era (2014–2021). Abbreviations: BAR, Balance of Risk; DRI, donor risk index; ID2EAL-D, improved donor-to-recipient allocation score for deceased donors only; ID2EAL-DR, improved donor-to-recipient allocation score for both deceased and living donors; P-SOFT, pre-allocation score to predict survival outcomes following liver transplantation; SOFT, survival outcomes following liver transplantation; SB, survival benefit.

DISCUSSION

In this large retrospective cohort study of UNOS LT data, we evaluated the performance of commonly used scores to predict post-transplant patient and graft survival. In general, scores accounting for both donor and recipient factors performed significantly better in predicting post-transplant outcomes. The SB model and SOFT scores yielded the highest discrimination for short-term patient and graft survival overall, followed by P-SOFT and ID2EAL-DR scores. The SB model and SOFT scores yielded the highest-fidelity predictions of 1-year patient survival and both performed well in predicting 90-day graft survival and 1-year conditional graft survival, in particular in the modern transplant era. By contrast, the SB model consistently had the highest discrimination of all models at longer timepoints (3- and 5-year post-transplant survival) and was the only model in the modern transplant era to have good discrimination (C ≥ 0.60) for all patients and graft outcomes. However, the overestimation of SB, SOFT, and P-SOFT models at longer timepoints for patient survival (3 and 5 years) may result from the fact that these models were developed in 2008–2009 in an era with poorer post-transplant outcomes. By contrast, the ID2EAL-DR score which was developed in 2022, demonstrated adequate calibration at all post-transplant timepoints for patient survival, though discrimination was inferior to SB, SOFT, and P-SOFT models. With regard to prediction of graft survival outcomes, ID2EAL-DR was superior to DRI, ID2EAL-D, and Balance of Risk scores but inferior to SB and SOFT models.

A major finding in this study was that scores relying on donor factors alone (ie, DRI and ID2EAL-D) yielded poor predictive performance, especially in the modern era. This calls into question the continued usage of the DRI in clinical and research settings as it is unclear if this score translates to meaningful benefit in post-transplant prediction of outcomes. This also underscores the hazards of considering donor risk in isolation during allocation decisions, as there is a significant improvement in predictive performance when considering donor-recipient elements together. Declining performance of prediction scores over time was a universal finding, likely due to changes in the modern transplant era, including the rise in DCD donors, development of techniques such as normothermic regional perfusion, changing liver disease epidemiology, and allocation policy changes, among others.¹⁰ Indeed, discrimination for post-transplant survival for all scores declined across successive transplant eras. In light of these findings, prediction models should be periodically updated to maintain adequate performance, or dynamic prediction models developed that may accommodate new data that reflect a changing transplant landscape. The SB model, in particular, which was generally high performing for all outcomes, would benefit from refitting and revision with modern transplantation data, as well as the development of a publicly available calculator for clinician use. Finally, we highlight that no existing scores can account for the current and future use of DCD machine perfusion technology, as cold ischemic time, as captured in the UNOS database, is not a relevant variable for this technique. Prediction scores will also need to be reevaluated and new scores developed to accurately predict risk in the era of MELD 3.0 and continuous distribution.

In subgroup analyses, we were able to isolate patient scenarios where particular scores performed better than others. For instance, for both 1-year post-transplant patient mortality and 90-day graft failure, the SB, SOFT, and P-SOFT scores had particularly high discrimination in recipients aged < 40 years or recipients with MELD ≥ 25. DCD versus DBD donor did not substantially affect discrimination for these outcomes for any score. These observations are clinically relevant as centers may have confidence that models such as SB, SOFT, and P-SOFT will perform especially well in younger patients and those with higher MELD with respect to the prediction of 1-year post-transplant mortality and 90-day graft survival. By the same token, these models may be less reliable when projecting risk for older recipients (age > 60 y) and those with low MELD scores (< 15). This knowledge may help centers with allocation decisions that may have a direct impact on improving center-specific metrics communicated through the Scientific Registry of Transplant Recipients Program-Specific Reports. Finally, the SB model is perhaps the only score to be recommended in predicting 1-year conditional graft failure at this time, given the generally poor overall performance of all other models in predicting this outcome.

There are several important limitations to highlight in this study. First, though we present subgroup analyses across recipient age, MELD, and DBD versus DCD, there are likely other key variable stratifications to consider that may be explored in future studies. Second, we recognize that comparing prediction models that account for recipient and donor factors versus donor factors alone is not an equivalent comparison; however, analyzing both groups together reveals the inadequacy of scores derived from only donor factors. Third, the SOFT score was calculated in the absence of the variable “portal bleed in the 48 hours before transplant,” which is not recorded in the UNOS database. Thus it is possible that prediction model performance could have been slightly underestimated for this score in this study. Fourth, we did not comprehensively assess the performance of all existing risk scores but rather focused on those that may be used to make patient-level allocation and transplant decisions and/or those frequently used for risk adjustment in research (eg, DRI). A notable exclusion from this study are the Scientific Registry of Transplant Recipients Program-Specific Reports risk adjustment models, which are intended for center-level risk adjustment in assessing the performance of transplant centers rather than being used as tools to facilitate allocation or transplantation decisions. Future studies may seek to leverage modern transplant data to create novel prediction scores to be compared to existing standards.

In conclusion, our study demonstrated superior discrimination of the SB and SOFT scores for post-transplant patient and graft survival versus all other scores; however, the ID2EAL-DR score represented the best balance of discrimination and calibration for patient survival over follow-up periods beyond 1 year. The SB and SOFT scores performed especially well in younger and higher MELD recipients. Finally, the inclusion of both donor and recipient factors in scores was critical to adequate prediction of post-transplant outcomes, emphasizing the importance of considering donor-recipient elements together during allocation decisions.

AUTHOR CONTRIBUTIONS

Nadim Mahmud, Lauren Shaffer, Samir Abu-Gazala, and Peter Abt: Intellectual genesis; Nadim Mahmud: Formal data analysis and visualizations; Nadim Mahmud, Lauren Shaffer, Douglas E. Schaubel, Samir Abu-Gazala, and Peter Abt: Data interpretation; Nadim Mahmud and Lauren Shaffer: Manuscript drafting; Nadim Mahmud, Lauren Shaffer, Douglas E. Schaubel, Samir Abu-Gazala, and Peter Abt: Critical review of the manuscript.

FUNDING INFORMATION

Nadim Mahmud is supported by a National Institute of Diabetes and Digestive and Kidney Diseases grant (K08-DK124577).

CONFLICTS OF INTEREST

Nadim Mahmud received grants from Grifols. The remaining authors have no conflicts to report.