Non-parametric individual treatment effect estimation for survival data with random forests.

Area of Science:

• Biostatistics
• Machine Learning
• Personalized Medicine

Background:

• Accurate estimation of individual treatment effects is crucial for personalized medicine.
• Time-to-event data with censoring presents unique challenges for treatment effect estimation.
• Existing methods may struggle to identify treatment effects in specific patient subgroups.

Purpose of the Study:

• To propose a novel random forest method for estimating individual treatment effects with censored survival data.
• To evaluate the performance of the proposed method against existing state-of-the-art techniques.
• To demonstrate the method's utility in identifying treatment effects within specific subpopulations using real-world cancer data.

Main Methods:

• Development of a random forest algorithm with a specialized splitting rule tailored for individual treatment effect estimation.
• Utilizing a dataset of similar subjects from the training data to compute individual treatment effect estimations.
• Validation through simulation studies comparing the proposed method with multiple competitors.

Main Results:

• The proposed random forest method demonstrates robust and stable performance in estimating individual treatment effects.
• Simulation results indicate superior or comparable performance against numerous competing methods.
• Application to colon and breast cancer data successfully identified treatment effects in subpopulations, even when overall effects were negligible.

Conclusions:

• The novel random forest approach provides a reliable tool for estimating individual treatment effects in the presence of censored survival data.
• The method has the potential to uncover treatment benefits in specific patient subgroups, advancing personalized cancer therapy.
• An R package for the method will be available soon, with code accessible upon request.