A new statistical approach for joint modeling of longitudinal outcomes measured in electronic health records with clinically informative presence and observation processes

TL;DR

This paper introduces EHRJoint, a novel statistical framework that jointly models visit, observation, and biomarker processes in EHR data to address biases from informative presence and observation, improving association estimates.

Contribution

The paper develops a new joint modeling method, EHRJoint, that handles both informative presence and observation processes simultaneously in EHR-based longitudinal studies.

Findings

EHRJoint provides unbiased exposure effect estimates in simulations.

Existing methods fail under real-world IP and IO patterns.

Application to Michigan Genomics Initiative data reveals meaningful associations.

Abstract

Biobanks with genetics-linked electronic health records (EHR) have opened up opportunities to study associations between genetic, social, or environmental factors and longitudinal lab biomarkers. However, in EHRs, the timing of patient visits and the recording of lab tests often depend on patient health status, referred to as informative presence (IP) and informative observation (IO), which can bias exposure-biomarker associations. Two gaps remain in EHR-based research: (1) the performance of existing IP-aware methods is unclear in real-world EHR settings, and (2) no existing methods handle IP and IO simultaneously. To address these challenges, we first conduct extensive simulation studies tailored to EHR-specific IP patterns to assess existing methods. We then propose a joint modeling framework, EHRJoint, that simultaneously models the visiting, observation, and longitudinal biomarker processes to address both IP and IO. We develop a computationally efficient estimation procedure based on estimating equations and provide asymptotically valid inference. Simulations show that EHRJoint yields unbiased exposure effect estimates under both IP and IO, while existing methods fail. We apply EHRJoint to the Michigan Genomics Initiative data to examine associations between repeated glucose measurements and two exposures: genetic variants and educational disadvantage.