Machine learning does not outperform traditional statistical modelling for kidney allograft failure prediction

Area of Science:

• Nephrology
• Transplantation
• Biostatistics

Background:

• Kidney allograft survival prediction is crucial for patient management.
• Machine learning (ML) models offer potential for improved prediction accuracy.
• The performance of ML models compared to traditional methods for kidney allograft outcomes is not well-established.

Purpose of the Study:

• To develop and validate ML-based prediction models for kidney allograft survival.
• To compare the performance of ML models against a validated Cox-Based Prognostication System (CBPS).

Main Methods:

• Developed tree-based (RSF, RSF-ERT, CIF), Support Vector Machine (LK-SVM, AK-SVM), and gradient boosting (XGBoost) models using a French derivation cohort (4000 patients).
• Included donor, recipient, and transplant-related parameters as predictors.
• Externally validated models using international cohorts (Europe, North America, South America) comprising 8422 kidney transplant recipients.
• Compared ML model performance (C-index) and calibration against the CBPS.

Main Results:

• Across 8422 recipients, 12.84% experienced graft loss after a median follow-up of 6.25 years.
• ML models achieved C-indices ranging from 0.527 to 0.788, while the CBPS achieved 0.808.
• ML models demonstrated discrimination performance comparable to the CBPS in validation cohorts, but calibration was similar or less accurate.

Conclusions:

• ML models, despite good performance, do not surpass the traditional CBPS in predicting kidney allograft failure.
• The study supports the continued use of traditional statistical approaches for kidney graft prognostication.
• Further investigation into ML model transparency and calibration is warranted for clinical application.