Development and validation of a machine-learning model to reduce futile procurements in donations after circulatory death in liver transplantation in the USA: a multicentre study.

The number of liver transplants from donors after the circulatory determination of death continues to increase, helping to alleviate the existing organ shortage. However, the rate of attempted but subsequently terminated procurements, known as futile procurements, remains high-mainly because many potential donors do not progress to death within a timeframe after extubation that maintains the suitability of the organ for donation. Futile procurements pose considerable financial and workload burdens to the transplant system. We aimed to develop and validate a machine-learning model to better predict progression to death and reduce futile procurements in cases of donation after circulatory death (DCD).This study included data from 2221 donors from six centres in the USA. Using a retrospective dataset obtained from 1616 donors between December 1, 2022, and June 30, 2023, we developed a prediction model using the Light Gradient Boosting Machine (LightGBM) framework, with neurological, biochemical, respiratory, and circulatory parameters as predictors. The model was validated retrospectively with data from 398 donors (July 1-Aug 31, 2023) and prospectively with data from 207 donors (March 1-Sept 30, 2024). The performance of the model was evaluated through the area under the receiver operating characteristic curve (AUC), accuracy, futile procurement rate, and missed opportunity rate. We also compared the performance of the model with that of two existing risk-prediction tools (the DCD-N score and the Colorado Calculator) and surgeon predictions.Of the 2221 DCD donors in this study, 1260 progressed to death, 927 of whom died within 30 min after extubation. Cross-validation of the LightGBM model yielded AUCs for predicting donor progression to death of 0·833 (95% CI 0·798-0·868) at 30 min, 0·801 (0·767-0·834) at 45 min, and 0·805 (0·770-0·841) at 60 min after extubation. This performance was maintained in both retrospective (0·834 [0·772-0·891], 0·819 [0·757-0·870], and 0·799 [0·737-0·855]) and prospective (0·831 [0·768-0·885], 0·812 [0·749-0·874], and 0·805 [0·740-0·868]) validation cohorts. Compared with surgeon predictions, the LightGBM model had lower futile procurement rates (0·195 vs 0·078, respectively), higher accuracy in cases of poor intersurgeon agreement (0·08 vs 0·29) at 30 min, and similar missed opportunity rates (0·155 vs 0·167). By contrast, the DCD-N score had AUCs of 0·799 (95% CI 0·730-0·860) at 30 min, 0·760 (0·695-0·824) at 45 min, and 0·739 (0·668-0·801) at 60 min, and the Colorado Calculator had AUCs of 0·694 (0·616-0·768), 0·669 (0·596-0·742), and 0·663 (0·585-0·736) at the same timepoints.We show that, compared with surgeon predictions and existing risk-prediction tools, our machine-learning model can enhance the accuracy of the prediction of progression to death in DCD donors and reduce futile procurements. Such improvements have the potential to alleviate some of the financial and operational burdens on the transplant community. Further improvements are needed to decrease missed opportunities and improve the overall accuracy of such models.None.

Duke Scholars

Author Daniel Firl