BACKGROUND: Accurate prediction of kidney graft failure at different phases post-transplantation is critical for timely intervention and long-term allograft preservation. Traditional survival models offer limited capacity for dynamic, time-specific risk estimation. Machine learning (ML) approaches, with their ability to model complex patterns, present a promising alternative.

METHODS: This study developed and dynamically evaluated phase-specific ML models to predict kidney graft failure across five post-transplant intervals: 0-3 months, 3-9 months, 9-15 months, 15-39 months, and 39-72 months. Clinically relevant retrospective data from deceased donor kidney transplant recipients were used for training and internal validation, with performance further confirmed on a blinded external validation cohort. Predictive performance was assessed using ROC AUC, F1 score, and G-mean.

RESULTS: The ML models demonstrated varying performance across time intervals. Short-term predictions in the 0-3 month and 3-9 month intervals yielded moderate accuracy (ROC AUC = 0.73 ± 0.07 and 0.72 ± 0.04, respectively). The highest predictive accuracy observed in mid-term or the 9-15-month window (ROC AUC = 0.92 ± 0.02; F1 score = 0.85 ± 0.03), followed by the 15-39-month period (ROC AUC = 0.84 ± 0.04; F1 score = 0.76 ± 0.04). Long-term prediction from 39 to 72 months was more challenging (ROC AUC = 0.70 ± 0.07; F1 score = 0.65 ± 0.06).

CONCLUSION: Phase-specific ML models offer robust predictive performance for kidney graft failure, particularly in mid-term periods, supporting their integration into dynamic post-transplant surveillance strategies. These models can aid clinicians in identifying high-risk patients and tailoring follow-up protocols to optimize long-term transplant outcomes.