Machine Learning Predictions for Assessing Hard-to-Place Deceased Donor Kidneys.

Area of Science:

• Transplantation science
• Machine learning in healthcare
• Organ allocation optimization

Background:

• Nearly 20% of deceased donor kidneys are placed out-of-sequence to prevent nonuse.
• Standard allocation rules may not always identify kidneys at high risk of nonuse effectively.

Purpose of the Study:

• Develop machine learning (ML) models to predict kidney nonuse risk during allocation.
• Assess current out-of-sequence kidney placements using ML predictions.

Main Methods:

• Retrospective cohort study using Organ Procurement and Transplantation Network data (2022-2023).
• Developed ML models using clinical data, biopsy, and center refusal patterns.
• Evaluated model performance using AUC, accuracy, and feature importance; compared predicted nonuse probabilities.

Main Results:

• ML models incorporating refusal information outperformed those without it (AUC 0.90 vs 0.88).
• Center refusal data was a key predictor.
• Out-of-sequence kidneys had intermediate nonuse probabilities; ML identified hard-to-place kidneys within KDPI strata.

Conclusions:

• ML models can identify kidneys at high risk of nonuse earlier and more accurately than KDPI.
• ML offers data-driven tools for real-time identification of hard-to-place kidneys.
• This can standardize accelerated placement, improve evaluation of practices, and reduce organ wastage.