Privacy-enhancing sequential learning under heterogeneous selection bias in multi-site electronic health records data

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Distributed health platforms face challenges in centralized analysis due to patient privacy concerns.
• Heterogeneous selection mechanisms across electronic health record (EHR) sites complicate disease risk parameter estimation.
• Sharing individual-level patient data is often infeasible in multi-site health research.

Purpose of the Study:

• To develop and validate privacy-enhancing statistical methods for disease risk estimation across multiple EHR sites.
• To address challenges posed by heterogeneous selection mechanisms and avoid individual-level data sharing.
• To illustrate the utility of proposed methods in a cross-biobank analysis of smoking and cancer subtypes.

Main Methods:

• Proposed Sequential Pseudo-Likelihood (SPL) and Sequential Augmented Inverse Probability Weighting (SAIPW) methods.
• Utilized summary statistics shared across sites and external population information to adjust for selection bias.
• Compared SPL and SAIPW against unweighted and centralized/meta-learning benchmarks using simulated and real-world data from the NIH All of Us (AOU) and Michigan Genomics Initiative (MGI) biobanks.

Main Results:

• Unweighted estimators showed significant bias, while SPL and SAIPW produced unbiased estimates with valid coverage.
• SAIPW demonstrated robustness to selection model misspecification.
• Both SPL and SAIPW exhibited minimal efficiency loss compared to centralized methods, outperforming unstable meta-learning approaches for rare outcomes.
• Analyses identified strong associations between smoking and lung, bladder, and larynx cancers.

Conclusions:

• Site-specific selection biases must be accounted for in distributed health networks.
• SPL and SAIPW provide practical, scalable, privacy-enhancing solutions for harmonizing diverse biobanks.
• The developed framework facilitates valid inference across EHR sites, enabling scalable, distributed research with real-world data.