Predicting graft and patient outcomes following kidney transplantation using interpretable machine learning models

Area of Science:

• Nephrology
• Transplant Surgery
• Biostatistics

Background:

• Clinical decisions for deceased donor kidney offers are challenging due to lack of predictive tools.
• Accurate prediction of transplant outcomes is crucial for optimizing organ allocation.

Purpose of the Study:

• To develop and validate interpretable machine learning models for predicting kidney graft failure and patient death.
• To identify key clinical factors influencing kidney transplant outcomes.

Main Methods:

• Trained and compared survival analysis models, including neural networks and Cox proportional hazard models.
• Utilized over 25 years of kidney transplant outcome data.
• Employed post hoc interpretability techniques for clinical validation.

Main Results:

• Neural networks demonstrated comparable performance to Cox models, with concordance indices of 0.63 for graft failure and 0.79 for patient death.
• Identified donor/recipient age, DR locus mismatches, dialysis type, and primary renal disease as significant predictors.
• Models showed clinical relevance through post hoc interpretation.

Conclusions:

• Interpretable neural networks show promise for predicting kidney transplant outcomes.
• These models can serve as a foundation for clinical decision support systems in organ transplantation.
• Improved prediction accuracy can aid in accepting or declining organ offers.