Predicting expiration among potential donation after circulatory death donors at 30- and 120-min post extubation

Many potential donation after circulatory death (DCD) donors do not progress to circulatory death within meaningful timeframes, leading to wasted procurement resources. We therefore sought to develop robust machine learning models to predict donor expiration at 30- and 120-min following extubation.

We performed a retrospective cohort study using data from 3 organ procurement organizations, analyzing 4464 potential DCD donors from 2014-2025 across 61 clinical variables. Primary endpoints were binary expiration at 30- and 120-min post extubation. Three cohorts were constructed. Regression analysis utilized the "complete-case dataset", while "max sample" (n = 4464; 14 variables) and "max variable" (n = 2092; 61 variables) datasets were used for XGBoost modeling to maintain completeness >85%. Results were evaluated using receiver operating characteristic (ROC) and precision-recall (PR) area under the curve (AUC).

Regression modeling yielded modest results. Among the machine learning models, the max variable model had the highest performance and yielded ROC-AUCs of 76% and 89% for 30- and 120-min, respectively, while PR-AUCs reached 80% and 97%. In both the 30- and 120-min models, final cough, lung offer, arterial blood gas pH, and serum platelets were among the top 6 variables of importance.

These models represent the most sophisticated approach to predicting DCD donor expiration to date. By leveraging a large multicenter cohort and nonlinear machine-learning techniques, our model more accurately captures the complex physiologic dynamics governing donor progression. While further validation is warranted, this framework has potential to reduce futile procurements and more accurately allocate resources, expanding access to life-saving transplantation.