A dynamic machine learning model for predicting organ transplant success to optimize donor organ utilization and...

Organ recovery for transplantation costs approximately $40,000 per donor, yet nearly 20% of recovered kidneys ultimately go unused. Static viability scores (e.g., KDPI, DRI) provide baseline risk stratification but fail to capture dynamic physiological changes that occur during donor management.

We trained and evaluated six machine learning (ML) models using the ORCHID dataset derived from 56 U.S. organ procurement organizations (OPOs) between 2015 and 2021. To ensure generalizability, we implemented a spatial hold-out validation strategy, training models on 70% of OPOs and testing them on a geographically distinct 30% cohort. Predictors included prerecovery vital signs, blood chemistry, serologic markers, donor demographics, and circumstances of death. We compared logistic regression (linear baseline), random forest, XGBoost, LightGBM, CatBoost, and a multilayer perceptron (NN-MLP). Hyperparameters were optimized using Bayesian methods, and missing values were addressed using median imputation to prevent zero-inflation bias.

Nonlinear, tree-based models significantly outperformed both the linear baseline and KDPI. LightGBM achieved the highest discrimination (ROCAUC, 0.975; PRAUC, 0.809), followed closely by random forest (ROCAUC, 0.975) and CatBoost (ROCAUC, 0.974). At a classification threshold of 0.5, the top-performing model demonstrated 94.1% accuracy and 88.1% sensitivity. In contrast, logistic regression showed lower sensitivity (77.5%), suggesting that ML approaches more effectively identify borderline viable organs. The most influential predictors included serum creatinine, brain death status, blood urea nitrogen, and donor age. Compared with KDPI alone (ROCAUC, 0.830), the ML model improved discrimination by more than 0.14.

A dynamic ML model evaluated using spatial validation substantially outperforms static scoring systems for predicting kidney utilization. By providing interpretable, near-real-time assessments, this approach could help OPOs avoid failed recoveries and reduce unnecessary organ discards. Prospective trials are needed to assess its impact on downstream clinical transplant outcomes.