Development and Validation of a Comprehensive Model to Estimate Early Allograft Failure Among Patients Requiring Early Liver Retransplant

Key Points Question Can the individual risk estimation for early allograft failure (EAF) be improved in view of liver retransplant? Findings In this multicenter cohort study investigating the association between donor-recipient factors and EAF, a novel Early Allograft Failure Simplified Estimation (EASE) score was developed. The score includes Model for End-stage Liver Disease score, transfused packed red blood cells, and hepatic vessel early thrombosis as well as transaminases, platelet, and bilirubin kinetics as variables on day 10 after transplant. The EASE score outperformed previous model scores, estimating EAF risk with 87% accuracy on day 90 after transplant; EASE was developed on a multicenter Italian database (1609 recipients) and validated on an external UK database (538 recipients). Meaning In this study, the EASE score rated the EAF risk (0%-100%) and identified cases at unsustainable risk to be listed for retransplant.


1.Statistical analysis
The study was performed according to current recommendations for retrospective observational analysis reporting in transplant population. 1-2 Continuous variables were presented as medians and interquartile ranges (IQR) or means ± standard deviations (SD), whilst categorical variables were summarized as numbers and percentages. Categories of patients who could be confounders due to peculiarities and/or low prevalence were excluded (Figure 1). Missing data were not managed by imputation methods because of their exiguous number (eTable 1).
Graft and patient survival curves were performed according to Kaplan-Meier and compared using the log-rank test. The goodness of fit was assessed using the Hosmer-Lemeshow test. 3 We evaluated also the calibration of the final model using the Calibration-BELT test of the final model. 4 In the derivation set, the Cstatistic comparison of the final model at 90 days with Model for Early Allograft Failure (MEAF), 5 L-GrAFT, 6 EAD, 7 Donor age x Model for End-stage Liver Disease (D-MELD), 8 new Theoretical Euro-Transplant Donor Risk Index (ET-DRI), 9 and Donor Risk Index (DRI) 10 was performed through non-parametric method. 11 The P value <.05 was considered significant. Statistics were performed using SPSS (ver. 25.0) and STATA (ver. 14.0) packages.

Work-flow to develop the final model
We initially replicated the methodology adopted in the seminal L-GrAFT study. 6 This score was derived through a kinetic approach using the area under the curve (AUC), direction and steepness of the curve (SLOPE). AUC and SLOPE were calculated using 10 evaluations (one a day from day 1 to 10). This methodology was adopted for AST, platelets and bilirubin. The highest value of INR, recorded from day 1 to 10, was included. The logarithmic trapezoidal method was used to calculate AUC and SLOPE for AST and platelets, while standard linear trapezoidal method was used for bilirubin. 6 In this study, we aimed to build a comprehensive model available at the 10 th postoperative day. An extensive set of variables was considered, including pre-operative and intraoperative parameters. Due to the timedependence incidence of EAF, variables were first analyzed by univariate Cox regression, adopting the same methodology used to develop the L-GrAFT. In addition to other significant parameters, not relevant for subsequent analysis, PRBC, THV, AST-AUC 2 , platelets-AUC, platelets-SLOPE, and bilirubin-SLOPE were significant at all time spans. MELD was significant at POD 2-30, 2-60, and 2-90 evaluation times (eFigure 1).
Single values of AUC and SLOPE were calculated for each case. Some variables in the original L-GrAFT model were expressed as their squared forms, and for these we adopted the square elevation. Such variables were then analyzed by univariate and multivariate logistic regression. Only variables with a P value <.2 at univariate logistic regression were included in multivariate analysis. Interestingly, not all the variables included in the L-GrAFT were significant.
We initially tested the same beta-coefficients of the original L-GrAFT model derived from 40 data entries validating L-GrAFT in our population. Following evaluation of the entire set of lab data we reduced the number of entries by recording only data at specified PODs.
In details, the number of lab data entries was reduced (fixed POD determinations instead of each day determinations from POD 1 to POD 10). In total, there were 4 entries for bilirubin and 4 entries for PLT (POD 1, 3, 7, and 10) and 5 for AST (POD 1, 2, 3, 7, and 10). The timing of data entries was chosen to best include relevant changes. In order to capture the cytolysis peak, the inclusion of day-2 AST data was necessary. Next, the number of calculated variables was reduced in order to maintain an adequate proportion between parameters and events in the logistic models. Furthermore, additional donor-and recipient-related parameters, not originally included in the L-GrAFT model, were investigated.
In summary, four subsequent logistic models (1,2,3,4) were developed in the derivation set, to reduce the number of data entries, improving C-statistic and including additional factors. Five additional models (5,6,7,8,9) were tested in order to investigate the impact of THV, DCD and MP grafts in the derivation and validation sets. The models 5-9 were adjusted for Center volume.