Refining the Lung Allocation Score Models Fails to Improve Discrimination Performance.

Area of Science:

• Transplant medicine
• Biostatistics
• Health services research

Background:

• Broader geographic sharing in lung transplant allocation is implemented via the Composite Allocation Score (CAS) system.
• Accurate prediction models for waitlist (WL) and posttransplant (PT) survival are increasingly vital for organ access determination.

Purpose of the Study:

• To evaluate the performance of current CAS survival models.
• To determine if alternative statistical or machine learning models can enhance discrimination performance for lung transplant allocation.

Main Methods:

• Utilized Scientific Registry for Transplant Recipients (SRTR) data (2010-2020).
• Developed seven WL and PT survival models, including the current CAS model, re-estimated CAS models, and advanced machine learning approaches (e.g., random survival forests, gradient-boosted trees).
• Assessed model discrimination using Area Under the Curve (AUC) at multiple time points (WL: 1, 3, 6 months; PT: 1, 3, 5 years).

Main Results:

• Waitlist (WL) model performance was comparable across all tested models, with AUCs ranging from 0.93 (baseline) to 0.84-0.85 (6-month predictions).
• Posttransplant (PT) model performance was lower, with AUCs between 0.58-0.61, showing stability over time but slightly poorer performance in residual cohorts.
• The greatest variability in AUC was observed for individuals with Medicaid insurance across both WL and PT models.

Conclusions:

• Employing alternative modeling strategies and contemporary cohorts did not yield significant improvements in the predictive performance of models used for lung transplant access.
• The findings suggest that current modeling approaches for lung transplant allocation may have limitations in accurately discriminating survival outcomes, particularly for specific patient subgroups.