The utility of machine learning for predicting donor non-use in lung transplantation

Efforts to reduce waitlist mortality in lung transplantation have been hindered by a high donor lung non-use rate, with approximately one-fifth of deceased organ donors ultimately providing lungs for transplantation. Rationale for organ decline varies across individual providers and transplant centers. Greater donor lung utilization could significantly reduce waitlist mortality and improve clinical outcomes for adult lung transplant candidates.

This study aimed to evaluate whether machine learning models can successfully predict donor lung non-use and characterize donor factors associated with current utilization patterns, with the long-term goal of informing more consistent and data-driven donor lung evaluation.

Using U.S. national registry data collected between 2012 and 2022, we generated five machine learning models - logistic regression, decision tree, random forest, XGBoost, and Naive Bayes - all of which were trained with 19 donor variables to predict donor lung non-use. Models were tested with both full donor feature sets and restricted donor feature sets comprised of the top 15, 10, and 5 predictors. Performance metrics, including area under the receiver operating characteristic curve (AUROC), accuracy, precision, recall, and F1 score, were assessed using cross-validation.

The random forest model incorporating the full feature set achieved the highest performance with an AUROC of 0.960 (95% CI 0.958-0.961) and an F1 score of 0.9489. However, the logistic regression (AUROC 0.886, 95% CI 0.883-0.888) and XGBoost (AUROC 0.906, 95% CI 0.902-0.908) models also demonstrated robust performance. A restricted random forest model that included only the top 15 predictors demonstrated similar efficacy (AUROC 0.956, 95% CI 0.954-0.958). Key predictors included PaO₂:FiO₂ ratio, donor age, body mass index, and serum aspartate aminotransferase levels. More restricted random forest models (i.e., those that only included 5 or 10 predictors) and any of the restricted logistic regression and XGBoost models had inferior but still acceptable performance (AUROC 0.867-0.958).

Machine learning models, particularly random forest, XGBoost, and logistic regression, accurately predicted donor lung non-use. Restricted models using fewer predictors also maintained strong performance, demonstrating technical feasibility for future evaluation in time-sensitive workflows. Because the models were trained on historical utilization decisions rather than recipient outcomes or independent measures of donor suitability, prospective validation is needed to determine whether they can improve appropriate donor lung utilization or support acceptance decisions without reinforcing practice-pattern bias.