Smoothed Estimation on Optimal Treatment Regime Under Semisupervised Setting in Randomized Trials

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Personalized Medicine

Background:

• Optimal treatment regimes are crucial for personalized medicine but often rely on limited labeled data.
• Existing semisupervised methods for treatment regimes face challenges with model assumptions and high computational costs.
• Unlabeled data, abundant in healthcare, holds valuable information for improving treatment regime estimation.

Purpose of the Study:

• To propose a model-free semisupervised framework for estimating optimal treatment regimes.
• To leverage large amounts of unlabeled data to enhance treatment regime estimation.
• To address the limitations of existing methods regarding model assumptions and computational burden.

Main Methods:

• Dimension reduction using a single-index model.
• Imputation of missing outcomes in unlabeled data via kernel regression.
• Development of semisupervised value functions incorporating labeled and unlabeled data.
• Derivation of optimal treatment regimes by maximizing semisupervised value functions.

Main Results:

• The proposed framework demonstrates consistency and asymptotic normality of estimators.
• A perturbation resampling procedure is introduced for asymptotic variance estimation.
• Simulations confirm the benefits of incorporating unlabeled data for optimal treatment regime estimation.

Conclusions:

• The novel semisupervised framework effectively utilizes unlabeled data for optimal treatment regime estimation.
• The approach offers a computationally efficient and model-free alternative to existing methods.
• The methodology is applicable to both randomized trials and observational studies.