Estimating Optimal Site-Specific Individualized Treatment Rules With Limited Data Sharing

Area of Science:

• Biostatistics
• Computational Biology
• Clinical Trials

Background:

• Estimating optimal individualized treatment rules (ITRs) is key for precision medicine.
• Multicenter trials present challenges due to site-level heterogeneity and data privacy constraints.
• Directly pooling data across centers can bias ITR estimation.

Purpose of the Study:

• To develop a privacy-preserving method for estimating site-specific ITRs in multicenter trials.
• To address site-level heterogeneity without accessing individual-level data.
• To propose Heterogeneous Distributed Learning (HD-learning) and Shrinkage HD-learning (SHD-learning).

Main Methods:

• HD-learning utilizes summary-level data from each site in a single communication round.
• A distributed mixed-effects model accommodates common fixed effects and site-specific random effects.
• SHD-learning is developed for scenarios with moderate- or high-dimensional covariates.

Main Results:

• Proposed HD-learning and SHD-learning methods estimate optimal site-specific ITRs.
• Asymptotic properties of the estimators are established under suitable conditions.
• Simulations demonstrate that the methods are lossless compared to pooled estimators and outperform methods ignoring heterogeneity.

Conclusions:

• HD-learning enables accurate estimation of site-specific ITRs using only summary data.
• SHD-learning extends this capability to complex, high-dimensional covariate settings.
• These methods offer a privacy-preserving solution for precision medicine in multicenter studies.