Immunosuppressive regimens and long-term kidney transplant outcomes: a dual survival modeling framework

Technical Summary: Immunosuppressive Regimens and Long-Term Kidney Transplant Outcomes

Problem Statement
Optimizing immunosuppressive therapy is critical for improving long-term outcomes following kidney transplantation. While induction and maintenance therapies are standard practice, their comparative effectiveness across diverse recipient populations requires further evaluation using contemporary, large-scale data. Existing literature often relies on smaller cohorts, older datasets, or lacks direct comparisons between classical statistical inference and modern machine learning (ML) predictive capabilities. Furthermore, few studies have simultaneously evaluated both induction and maintenance strategies against death-censored graft failure and all-cause patient mortality using a unified analytical framework.

Methodology
This retrospective cohort study analyzed data from 228,855 adult recipients of deceased-donor kidney transplants from the United Network for Organ Sharing (UNOS) registry, spanning January 1, 2000, to October 30, 2024.

• Study Design: A dual survival modeling framework was employed, integrating classical statistical inference with ML-based prediction.
• Outcomes: Two primary endpoints were analyzed: (1) death-censored graft failure (return to dialysis or re-transplantation) and (2) all-cause patient mortality.
• Predictors: The models incorporated donor characteristics (e.g., age, KDPI, DCD status), recipient factors (e.g., diabetes, race, dialysis dependence), and specific immunosuppressive regimens (induction agents like ATG, IL-2R, Alemtuzumab; maintenance regimens like CNI+MMF with or without steroids).
• Statistical Analysis:
  • Classical Approach: Multivariable Cox proportional hazards (PH) models were fitted to estimate adjusted hazard ratios (HRs) for clinical inference. Models were adjusted for transplant era to account for secular changes in practice.
  • Machine Learning Approach: Four ML survival algorithms were developed using the same covariate structure: Random Survival Forest (RSF), Support Vector Machine (SVM), Penalized Cox regression (CoxNet), and Extreme Gradient Boosting optimized with Cox partial likelihood (XGBoost-Cox).
  • Evaluation: Model performance was assessed on an independent test set (25% of data) using the Concordance Index (C-index) and time-dependent Area Under the Curve (tdAUC) at 1, 3, 5, and 10 years.
• Data Handling: Missing data were handled via imputation (medians for continuous, modes for categorical, with specific handling for cPRA and cold ischemic time). Multicollinearity was assessed and managed via Variance Inflation Factors (VIF).

Key Results

• Immunosuppressive Regimens:
  • Maintenance: Regimens containing Calcineurin Inhibitors (CNI) and Mycophenolate Mofetil (MMF) were associated with significantly lower hazards for both graft failure and patient mortality. Specifically, CNI+MMF (without steroids) showed the strongest association (Graft Failure HR: 0.72; Mortality HR: 0.78). The addition of steroids (CNI+MMF+Steroids) also conferred benefit, though with slightly smaller effect sizes.
  • Induction: Antithymocyte Globulin (ATG) was associated with lower hazards for both outcomes (HR ~0.93). IL-2R antagonists and Alemtuzumab demonstrated neutral associations. Notably, the combination of ATG+IL-2R was associated with a higher hazard of graft failure (HR 1.09).
• Risk Factors: Recipient diabetes, older age, dialysis dependence, and higher Kidney Donor Profile Index (KDPI) were the strongest adverse predictors. Recipients of Asian or Hispanic ethnicity showed lower adjusted hazards compared to White recipients.
• Model Performance:
  • The classical Cox PH model achieved robust discrimination (Graft Failure C-index: 0.687; Patient Mortality C-index: 0.703).
  • ML models (XGBoost-Cox, CoxNet, SVM, RSF) demonstrated performance comparable to or marginally superior to the Cox model (e.g., XGBoost-Cox C-index: 0.689 for graft failure; 0.705 for mortality).
  • The mean tdAUC for all models ranged between 0.703 and 0.727, indicating that while ML models offered slight improvements in predictive accuracy, the gains over classical Cox regression were modest.

Significance and Claims
The paper claims three principal contributions:

1. Comprehensive Evaluation: It provides one of the most extensive and temporally comprehensive national evaluations of immunosuppressive therapy to date, confirming the continued dominance of CNI/MMF-based maintenance regimens and the variable efficacy of induction strategies.
2. Methodological Framework: It introduces a comparative modeling paradigm that bridges interpretable hazard modeling (Cox) with predictive ML approaches. The study argues that while ML models offer complementary predictive value for individualized risk stratification, classical Cox models remain highly competitive and superior for clinical inference due to their interpretability (HRs).
3. Translational Potential: The study highlights the feasibility of using ML survival frameworks to support individualized post-transplant risk stratification, suggesting that future integration of richer clinical and biomarker data could further refine these tools.

The authors emphasize that the observed associations are adjusted for confounding but should be interpreted as hypothesis-generating rather than causal, given the retrospective nature of the registry data and the non-randomized allocation of immunosuppressive therapies. The study concludes that the dual analytical framework supports clinically informed regimen selection while demonstrating that advanced predictive algorithms currently offer incremental rather than transformative gains over traditional methods in this specific context.