Prospective Multicenter Observational Cohort Study on Time to Death in Potential Controlled Donation After Circulatory Death Donors-Development and External Validation of Prediction Models: The DCD III Study

Background. Acceptance of organs from controlled donation after circulatory death (cDCD) donors depends on the time to circulatory death. Here we aimed to develop and externally validate prediction models for circulatory death within 1 or 2 h after withdrawal of life-sustaining treatment. Methods. In a multicenter, observational, prospective cohort study, we enrolled 409 potential cDCD donors. For model development, we applied the least absolute shrinkage and selection operator (LASSO) regression and machine learning-artificial intelligence analyses. Our LASSO models were validated using a previously published cDCD cohort. Additionally, we validated 3 existing prediction models using our data set. Results. For death within 1 and 2 h, the area under the curves (AUCs) of the LASSO models were 0.77 and 0.79, respectively, whereas for the artificial intelligence models, these were 0.79 and 0.81, respectively. We were able to identify 4% to 16% of the patients who would not die within these time frames with 100% accuracy. External validation showed that the discrimination of our models was good (AUCs 0.80 and 0.82, respectively), but they were not able to identify a subgroup with certain death after 1 to 2 h. Using our cohort to validate 3 previously published models showed AUCs ranging between 0.63 and 0.74. Calibration demonstrated that the models over- and underestimated the predicted probability of death. Conclusions. Our models showed a reasonable ability to predict circulatory death. External validation of our and 3 existing models illustrated that their predictive ability remained relatively stable. We accurately predicted a subset of patients who died after 1 to 2 h, preventing starting unnecessary donation preparations, which, however, need external validation in a prospective cohort.