Prediction Models for LDLT Outcomes

Overview

Rationale: Living donor liver transplantation (LDLT) has emerged as an important option for patients with end-stage liver disease. To facilitate international and meaningful comparisons, our institution participates in the International LDLT Registry. Several models to predict outcomes post-LDLT have been developed to council and justify the major surgery that the living liver donors undergo. However, most proposed models are at high risk of bias and demonstrate suboptimal discriminative ability. This study aims to externally validate the most promising prediction models and subsequently, develop a new, clinically applicable prediction model for LDLT outcomes, using the International LDLT Registry. Objective(s): The main objective of this study is to develop a new, clinically applicable prediction model for LDLT outcomes, using the International LDLT Registry. The secondary objective is to externally validate the most promising existing prediction models for LDLT outcomes, using the International LDLT Registry. Study type: This is an observational, multicenter cohort study using prospectively collected data from the International LDLT Registry. Registry data will be analyzed retrospectively for the purposes of external model validation and prediction model development. Study population: The study population consists of living liver donors and their corresponding recipients recorded in the International LDLT registry. Methods: For external validation, parameters will be entered in the existing prediction models resulting in the predicted risks. Model discrimination will be measured using the area under the curve (AUC) and by the discrimination slope. The DeLong test will be used to test for difference between the AUC of the different prediction models. Calibration will be evaluated by comparing the observed with the predicted rate of events and graphically represented by calibration plots. For the development of a new prediction model, the outcome of interest is early graft failure, defined as graft loss within 90 days after transplantation. A multivariable logistic regression model will be developed to estimate the individual risk of early graft failure. Internal validation will be performed using bootstrapping, and model performance will be assessed in terms of discrimination and calibration. Model performance will also be tested in subgroups.

Research questions: 1. Can a new preoperative prediction model based on donor- and recipient-related variables be developed using the International LDLT Registry to reliably predict LDLT outcomes? 2. How do the most promising existing prediction models for LDLT outcomes perform when externally validated in the International LDLT Registry? Sample size calculation: Sample size for prediction model development was estimated using an events-per-parameter (EPP/EPV) approach. Based on previous research, the anticipated event rate for early graft failure was set at 17.5%. We prespecified 20 model parameters (including dummy variables for categorical predictors and any interaction terms) and targeted 15 events per parameter to reduce overfitting. This yields a minimum of 300 events, corresponding to a total sample size of approximately 1,715 patients. Statistical Analysis: Statistical analyses will be carried out using RStudio 2024.09.1, GraphPad Prism 10.6.1, and Microsoft Excel version 16.90.2. P \<0.05 indicates statistical significance. Normality will be tested using the Shapiro-Wilk test. Baseline characteristics will be divided based on the presence of early graft failure. Homogeneity of variances will be tested using the F-test. These results will be presented in a table. For external validation of existing prediction models, parameters will be entered in the prediction models resulting in the predicted risks. Model discrimination will be measured using the AUC and by the discrimination slope. The DeLong test will be used to test for difference between the AUC of the different prediction models. Calibration will be evaluated by comparing the observed with the predicted rate of events and graphically represented by calibration plots. For the development of a new prediction model, the outcome of interest is early graft failure, defined as graft loss within 90 days after transplantation (binary outcome). A multivariable logistic regression model will be developed to estimate the individual risk of early graft failure. Candidate predictors will be selected a priori based on clinical relevance, biological plausibility, and availability before transplantation, informed by existing literature and expert opinion. To reduce the risk of overfitting, the number of model parameters will be limited to approximately 20, in accordance with an events-per-parameter approach. Data-driven predictor selection procedures will be avoided. Internal validation will be performed using bootstrapping, and model performance will be assessed in terms of discrimination and calibration. Model performance will also be tested in the following subgroups: recipient sex, recipient continent of residence, indication for liver transplantation, actual donor hepatectomy performed, and approach to donor hepatectomy.