Predictive Models for Kidney Offer Acceptance: Challenges and Strategies.

Area of Science:

• Transplant medicine
• Machine learning in healthcare
• Predictive modeling

Background:

• Organ offer acceptance prediction is challenging due to high volumes, imbalanced data (more declines than acceptances), and limited insight into human decision-making.
• Existing offer acceptance models are used for program evaluation and policy development, but best practices and baselines are not well-established.
• This study investigates the impact of various machine learning models, feature sets, and sampling procedures on organ offer acceptance prediction.

Purpose of the Study:

• To compare the performance of different machine learning models for predicting kidney organ offer acceptance.
• To evaluate the impact of incorporating additional features, such as transportation logistics, on model performance.
• To assess the effectiveness of different data sampling procedures in improving prediction accuracy.

Main Methods:

• Evaluated multiple kidney offer acceptance models, ranging from logistic regression to gradient boosted trees (XGBoost).
• Trained models using donor and candidate characteristics, then augmented the best-performing model with transportation-related features and sampling procedures.
• Compared model performance using metrics like average precision and AUROC (Area Under the Receiver Operating Characteristic curve).

Main Results:

• The XGBoost model demonstrated the best performance improvement over the baseline logistic regression model (average precision increased from 0.0645 to 0.0907).
• Incorporating transportation-related features further enhanced model performance (average precision reached 0.0940).
• No substantial performance differences were observed based on the sampling procedures used.

Conclusions:

• Advanced machine learning models and non-clinical data, like transportation distances, can improve organ offer acceptance prediction.
• Significant trade-offs between precision and recall were observed, indicated by low average precision scores despite high AUROCs.
• Current models, even optimized ones, may not offer clear advantages over existing organ allocation policies, necessitating further research for clinical applicability.