Modeling the Restricted Mean Survival Time Using Pseudo-Value Random Forests.

Area of Science:

• Biostatistics
• Medical Data Analysis

Background:

• Restricted Mean Survival Time (RMST) is a key metric for summarizing event times in longitudinal studies.
• RMST represents life expectancy within a specific time frame and is crucial for causal analysis of treatment effects.
• Existing methods for RMST estimation often rely on restrictive assumptions, limiting their applicability.

Purpose of the Study:

• To introduce a novel non-parametric approach for modeling RMST conditional on baseline variables.
• To develop a flexible method for estimating patient-specific RMST and confounder-adjusted treatment contrasts.
• To overcome limitations of existing RMST modeling techniques, particularly the proportional hazards assumption.

Main Methods:

• A direct modeling strategy for RMST using leave-one-out jackknife pseudo-values.
• Integration of pseudo-values within a random forest regression framework (termed PVRF).
• A model-free approach ensuring estimates are not affected by restrictive statistical assumptions.

Main Results:

• PVRF provides precise estimates of patient-specific RMST values.
• The method enables accurate estimation of confounder-adjusted treatment contrasts.
• Numerical experiments and application to the SUCCESS-A breast cancer trial confirm PVRF's accuracy.

Conclusions:

• PVRF is a flexible and accurate method for RMST estimation and causal inference.
• The model-free nature of PVRF enhances its reliability in diverse clinical settings.
• PVRF expands the capabilities of pseudo-value modeling for high-dimensional data analysis in medical research.