Penalized Spline-Involved Tree-based (PenSIT) Learning for estimating an optimal dynamic treatment regime using...

Area of Science:

• Causal Inference
• Machine Learning
• Personalized Medicine

Background:

• Dynamic treatment regimes personalize care using time-adaptive decision rules.
• Existing methods for estimating optimal regimes can be unstable with variable weights.

Purpose of the Study:

• Introduce Penalized Spline-Involved Tree-based Learning (PSITL) for improved dynamic treatment regime estimation.
• Develop a novel purity measure for flexible, interpretable multi-stage treatment regime identification.

Main Methods:

• PSITL predicts counterfactual outcomes using regression with penalized splines of propensity scores and covariates.
• A new purity measure is integrated into a decision tree framework.
• The method was evaluated using simulation experiments and applied to the MIMIC dataset.

Main Results:

• PSITL demonstrates good performance compared to competing methods in simulations.
• The method shows particular advantages in small sample sizes and high confounding.
• Application to the MIMIC dataset identified variables for tailoring fluid resuscitation in sepsis.

Conclusions:

• PSITL provides a flexible and interpretable approach for estimating optimal dynamic treatment regimes.
• This method advances personalized medicine by enabling tailored treatment strategies.
• PSITL is particularly beneficial in challenging data conditions like small sample sizes or significant confounding.