Methods in dynamic treatment regimens using observational healthcare data: A systematic review

Area of Science:

• Health Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Dynamic Treatment Regimens (DTRs) are crucial for personalized medicine, guiding treatment decisions over time.
• Estimating optimal DTRs from observational healthcare data presents significant methodological challenges.
• Existing reviews lack a comprehensive synthesis of current DTR estimation techniques and their application.

Purpose of the Study:

• To systematically review and summarize methods for estimating Dynamic Treatment Regimens (DTRs) using observational healthcare data.
• To evaluate the strengths, weaknesses, and application settings of various DTR estimation approaches.
• To identify common assumptions, validation strategies, and research objectives in DTR estimation studies.

Main Methods:

• Systematic literature search of PubMed and EMBASE databases (1950-2022).
• Inclusion of observational studies evaluating medical treatments and estimating DTRs.
• Categorization of methods including reinforcement learning, counterfactual models, classification, and g-methods.

Main Results:

• 83 studies met inclusion criteria; reinforcement learning (44.6%) and counterfactual models (18.1%) were prevalent.
• Most studies (71.1%) aimed to refine existing DTRs rather than replicate human experts (28.9%).
• Commonly reported assumptions (65.1%) included exchangeability and positivity; time-varying confounders were noted in 50.6% of studies.

Conclusions:

• Reinforcement learning and counterfactual-based models are leading methods for DTR estimation in observational data.
• Future research should focus on addressing time-varying confounders and estimating conditional average treatment effects.
• Robust validation using diverse methods, including simulated and real-world data, is essential for reliable DTR estimation.