Censoring-adjusted tree-based policy learning for estimating dynamic treatment regimes with censored outcomes

Area of Science:

• Biostatistics
• Machine Learning
• Clinical Data Science

Background:

• Dynamic Treatment Regimes (DTRs) are crucial for sequential treatment decisions adapting to patient characteristics, especially for survival outcomes.
• Learning optimal DTRs from observational data is challenging due to complexities like censored data.
• Existing methods may not fully address the nuances of censored data in DTR estimation.

Purpose of the Study:

• To develop and evaluate a novel framework, Censoring-Adjusted Tree-based Reinforcement Learning (CA-TReL), for learning effective DTRs from observational data.
• To address the challenges posed by censored data in estimating optimal DTRs.
• To provide robust and interpretable personalized treatment strategies.

Main Methods:

• CA-TReL enhances traditional tree-based reinforcement learning with augmented inverse probability weighting (AIPW) and censoring-adjusted estimation.
• The framework is validated through extensive simulation studies.
• Real-world performance is assessed using the SANAD epilepsy dataset.

Main Results:

• CA-TReL demonstrated statistically superior performance compared to state-of-the-art methods (OWL, RWL, TBWL, ASCL, DWsurv).
• Improvements were observed across key metrics including restricted mean survival time (RMST) and decision-making accuracy.
• The method provides robust and interpretable treatment strategies.

Conclusions:

• CA-TReL offers a significant advancement in learning optimal DTRs from observational data, particularly with censored outcomes.
• The framework enhances personalized and data-driven treatment strategies in healthcare.
• This approach represents a step forward for adaptive clinical decision-making.