Estimation of Heterogeneous Restricted Mean Survival Time Using Random Forest

Area of Science:

• Biostatistics
• Machine Learning
• Survival Analysis

Background:

• Estimating heterogeneous restricted mean survival time (hRMST) is crucial for clinical interpretation of survival data with covariates.
• Current hRMST methods often depend on restrictive proportional hazards or parametric assumptions.
• There is a need for flexible and robust methods to model hRMST.

Purpose of the Study:

• To propose a novel random forest-based estimator for hRMST in the presence of right-censored survival data and covariates.
• To establish theoretical properties of the proposed estimator, including a central limit theorem.
• To develop an efficient method for constructing confidence intervals for hRMST.

Main Methods:

• Developed a random forest algorithm for estimating hRMST with right-censored survival data.
• Provided theoretical justification by proving a central limit theorem for the random forest estimator.
• Implemented a computationally efficient algorithm for confidence interval construction.

Main Results:

• Random forest-based hRMST confidence intervals demonstrate correct coverage probabilities in simulations.
• The proposed method exhibits smaller prediction errors compared to parametric models when models are misspecified.
• Application to ovarian cancer data from TCGA shows improved hRMST prediction performance.

Conclusions:

• The random forest approach offers a powerful and flexible alternative for hRMST estimation and prediction.
• This method provides accurate confidence intervals and superior predictive performance, especially under model misspecification.
• The developed method and software implementation (srf) are valuable tools for survival data analysis in clinical research.