Phase-specific kidney graft failure prediction with machine learning model.

Area of Science:

• Nephrology
• Transplant Surgery
• Biomedical Data Science

Background:

• Accurate prediction of kidney graft failure is crucial for timely intervention and preserving allografts.
• Traditional survival models have limitations in dynamic, time-specific risk estimation.
• Machine learning (ML) offers a promising alternative for modeling complex patterns in transplant outcomes.

Purpose of the Study:

• To develop and evaluate phase-specific ML models for predicting kidney graft failure.
• To assess the dynamic, time-specific predictive accuracy of ML models across different post-transplant intervals.
• To explore the potential of ML in optimizing post-transplant surveillance and patient management.

Main Methods:

• Developed phase-specific ML models for kidney graft failure prediction across five intervals (0-3, 3-9, 9-15, 15-39, 39-72 months).
• Utilized retrospective data from deceased donor kidney transplant recipients for training and internal validation.
• Validated model performance on an external cohort using ROC AUC, F1 score, and G-mean.

Main Results:

• ML models showed varying accuracy across time intervals, with moderate short-term prediction (0-9 months).
• Highest predictive accuracy was achieved in the mid-term 9-15 month interval (ROC AUC = 0.92 ± 0.02).
• Long-term prediction (39-72 months) presented greater challenges (ROC AUC = 0.70 ± 0.07).

Conclusions:

• Phase-specific ML models provide robust predictions, particularly in mid-term post-transplant periods.
• These models can be integrated into dynamic surveillance strategies for kidney transplant recipients.
• ML models aid clinicians in identifying high-risk patients for tailored follow-up and improved outcomes.