This study was reported according to the STROBE guidelines for observational studies8.
Study population (derivation cohort)
We included all actual DCD kidney donors that have been transplanted, and their corresponding recipients transplanted in Australia between 9th April 2014 and 2nd January 2018. The Electronic Donor Record (EDR) system was introduced in 2014, therefore details of the pre-donation hemodynamic records were not available prior to 2014. The EDR contains pre-, peri- and post-mortem information of all donors from each state/territory in Australia who were consented for organ donation in Australia and is managed and held by Australian Organ and Tissue Authority in Canberra, Australia. Multi-organ DCD transplants were excluded. Other data were sourced from Australia and New Zealand Dialysis and Transplant (ANZDATA) Registry (recipient data and outcome measures). The institutional board and the human research ethics committee of the University of Western Australia approved the conduct of the study (ethics reference: RA/4/20/4743). Approval for data extraction from the EDR was granted by state and territory health departments and donor agencies. We have assigned this cohort as the derivation cohort because this dataset was provided to the investigator team during the first data request.
Data collection and linkage
Clinical data from the EDR includes hemodynamic parameters after WCRS such as systolic/diastolic BP (documented every 1-5 minutes using intra-arterial blood pressure monitoring, or in the absence of an arterial line, non-invasive oscillometric method (upper-arm cuff) was used), heart rate (documented every 10 minutes), prior requirement for inotropic/vasopressor support, the time to death, urine output over the 12 hours prior to death, cause of death and terminal serum creatinine. The EDR also contains the baseline characteristics of all actual deceased donors including age, sex, ethnicity, smoking history, personal history of diabetes, hypertension and body mass index (BMI). These data were entered into the EDR as part of routine data capture during the donation process by Donate Life Agency donor coordinator staff in each of the states and territories. The average (standard deviation [SD]) number of repeated hemodynamic measures per donor across the cohort was 9.5 (5.6).
In Australia, absence of circulation is evidenced by absent arterial pulsatility for a minimum of 3 minutes and a maximum of 5 minutes, using intra-arterial pressure monitoring and confirmed by clinical examination
(absence of heart sounds and/or central pulses). In cases without an arterial line, electrical asystole is observed for a minimum of 3 minutes and a maximum of 5 minutes on the electrocardiogram and confirmed by clinical examination.
The ANZDATA Registry contains recipient data and includes records of all recipient baseline characteristics of age, sex, ethnicity, smoking status, diabetes, body mass index (BMI), transplanting state, primary cause of kidney failure, comorbidities such as chronic lung disease, history of hepatitis, prior cancer, vascular disease and time on dialysis before transplantation. Other transplant characteristics such as the transplant date, number of human leukocyte antigen (HLA) mismatches and total ischemic time were also recorded in the registry.
Data on DCD donors and the corresponding recipients were collated to allow for the alignment of the donor pre-, peri- and post-mortem events with other donor and recipient characteristics and outcomes measures using unique donor identifier. The de-identified dataset was prepared by ANZDATA registry and all linked data between the ANZDATA registry and the EDR (donors and recipients) were not re-identifiable. During the timeframe of this study period, machine perfusion was not used routinely across all states and jurisdictions in Australia.
Outcome measures
DGF was defined as recipients who required dialysis within the first seven days of transplantation.
Statistical analyses
Continuous variables were described using means (SD) and medians (interquartile range [IQR]). Categorial variables were summarised with counts and percentages.
Latent class mixed-effect model
We fitted a latent class mixed model to identify classes of systolic BP trajectories9. This approach assumed that the individual patterns of systolic BP decline during the agonal phase could be grouped into several patterns of systolic BP decline (latent classes), sharing similar tendency. Latent class analysis (LCA) is a modelling technique based on structural equation model, which aims to identify subgroups of individuals with “unmeasured” class memberships of similar clinical characteristics or outcomes that are not directly observable from the data10. The individual was assigned to a particular latent class membership with the highest probability of the outcome measure. When the measurement or outcome of interest is longitudinal, the formulation of the problem will become the identification of the subgroups of the developmental trajectories.
In our model, a two stage-approach was applied to link the latent class with the risk of DGF: in the first stage, the latent classes are estimated based on clinical characteristics; in the second stage, the estimated latent classes were used as predictors for clinical outcomes. The latent class model was fitted by considering time as a fixed effect and donor characteristics (including sex, age, donor body mass index, donor terminal creatinine, warm ischemia, donor causes of death) as random effects, with systolic BP as the response variable, and took the form of a Gaussian distribution. In addition, the model was fitted with class-specific variance-covariance of the random-effects, with quadratic I-splines with 5 knots as link function to model the non-linear trend. Each trajectory was assigned to a certain latent class with the largest posterior likelihood. The number of latent classes was selected based on the Akaike information criterion (AIC), an estimator of in-sample prediction error, the entropy and by visual inspection of the patterns of trajectory decline in systolic BP11.
Characterisation of latent classes and the association with DGF
The comparison of the donor profiles between the latent classes was conducted using one-way analysis of variance (ANOVA) and chi-square tests for continuous and categorical variables, respectively. The association between classes of trajectories of systolic BP decline and DGF was examined using a multivariable logistic regression and adjusted for selected donor and recipient characteristics, and the transplanting states. These variables were selected using two different approaches: Least absolute shrinkage and selection operator (LASSO)12 and random forest13.
Variable selection using LASSO and random forest
For LASSO, the penalisation parameter was chosen by cross-validation and the covariates with non-zero coefficient were included in the logistic regression, together with class, and fitted using maximum likelihood (that is, LASSO was used solely for variable selection). For the random forest model, we trained the model to predict DGF with 500 trees, using five randomly chosen variables at each split. We examined the rankings of variables importance in random forest model and selected the variables with the importance scores of greater than 10 to be included in the final logistic models together with systolic BP trajectory classes.
Effects of the slope and intercept of systolic BP decline on the risk of DGF
We fitted a two-stage model to assess the association between donor-derived systolic BP trajectories and its association with DGF.
Details of the two-stage model
\(SB{P}_{it}\) was denoted as the systolic BP (mmHg) for transplant \(i\) at time point \(t\). First stage was written as \(SB{P}_{it}={\beta }_{0}+{b}_{0i}+\left({\beta }_{1}+{b}_{1i}\right)t+{ϵ}_{it}\), where \({\beta }_{0}\) and \({\beta }_{1}\) were the fixed effect parameters, and \({b}_{0i}\) and \({b}_{1i}\) were the donor specific random effect intercept and slope, respectively using lmer function in R package lme414. In the first stage, we used a linear mixed-effects model for systolic BP, with a random intercept and random effect for measurement time for each recipient. This longitudinal model allowed characterisation of individual systolic BP trajectories by the donor-specific intercept and rate of systolic BP decline (slope). We also tested the interaction between classes of systolic BP decline, donor age and DGF, and found donor age was not an effect modifier between the specific systolic BP latent classes and DGF.
In the next stage, the donor-specific intercept and rate of systolic BP decline were used in a multivariable logistic model for DGF, adjusted for other donor and recipient characteristics, and the transplanting states. These covariates were selected using similar methods to those described for the latent class approach (LASSO and random forest). The second stage model used the slope of systolic BP \({\widehat{\text{b}}}_{1\text{i}}\)estimated from Stage 1. This was written as: \(logit\left(P\left(DG{F}_{i}=1\right)\right)={\alpha }_{0}+{\alpha }_{1}{\widehat{b}}_{1i}+{{\gamma }}^{{T}}{{x}}_{{i}}\), where \({{\alpha }}_{1}\)was the coefficient for the estimated slope of systolic BP; and \({\mathbf{x}}_{\mathbf{i}}\) included the other variables in the multivariate logistic regression model. These variables were similar to those integrated in the latent class analyses and \({\gamma }\) were the vectors of the logistic regression coefficients.
Classification DGF and model evaluation
We compared the additional performance gained from the latent class trajectories, slope and intercept of systolic BP in addition to traditional clinical variables using random forest modelling. Each model was evaluated using 5-fold cross validation and repeated by 50 times. The accuracy rate and the Area Under the Receiving Operator Curve (AUC-ROC) were used to evaluate the performance of the models.
Validation cohort
Validation of the findings from the derivation cohort (the 8 latent classes, slopes and intercept of systolic BP) was conducted using data from all Australian DCD donors between 6th January 2018 to 24th December 2019. This was made available to the investigators during the second data request.
Sensitivity analyses
Using similar modelling strategies of latent class modelling, variable selections using Random Forest and LASSO, logistic regression and model evaluation; we also assessed the association between trajectories of diastolic BP and DGF. Given the AIC and entropy with 8 and 9 latent classes were similar, we then evaluated the association between 9 different classes of systolic BP trajectories and DGF in the sensitivity analyses.