Dynamic Treatment Regimes Using Bayesian Additive Regression Trees for Censored Outcomes

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Personalized medicine requires customized treatment plans, especially for progressive diseases like cancer.
• Dynamic Treatment Regimes (DTRs) formalize sequential decision-making in disease progression.
• Existing DTR optimization methods, such as Q-learning, are primarily frequentist.

Purpose of the Study:

• To adapt a Bayesian machine learning framework for optimizing DTRs with censored outcomes.
• To compare the proposed Bayesian approach with the frequentist Q-learning method.
• To develop a practical R tool for implementing the proposed method.

Main Methods:

• Utilized Bayesian additive regression trees (BART) within an accelerated failure time (AFT) modeling framework.
• Applied the approach to both simulated data and a real-world dataset.
• Developed an R wrapper function leveraging standard BART survival models.

Main Results:

• The proposed Bayesian DTR optimization method demonstrated effectiveness for censored outcomes.
• Simulations and real data analysis showed comparable or improved performance over Q-learning.
• The R wrapper function provides a user-friendly interface for applying the method.

Conclusions:

• Bayesian DTR optimization using BART offers a robust and flexible approach for censored data.
• This framework facilitates personalized treatment strategies in complex medical scenarios.
• The developed R tool enhances accessibility for researchers and clinicians.