Semi-supervised approach to event time annotation using longitudinal electronic health records

Area of Science:

• Biostatistics
• Health Informatics
• Precision Medicine

Background:

• Large clinical datasets from electronic health records (EHRs) are crucial for precision medicine but lack precise clinical outcome timing.
• Manual annotation of event times is resource-intensive, hindering the development of personalized prediction models.
• Existing methods struggle to accurately extract event times from billing or procedure codes alone.

Purpose of the Study:

• To propose and validate a semi-supervised, multi-modal automated time annotation (MATA) method for deriving precise clinical event times from EHR data.
• To address the challenge of missing precise outcome information in large-scale clinical databases for research.

Main Methods:

• A two-step semi-supervised approach utilizing longitudinal EHR encounter records.
• Step I: Functional principal component analysis (FPCA) to estimate intensity functions from unlabeled patient data.
• Step II: Penalized proportional odds model with B-spline approximation for event time outcomes using features from Step I on labeled data.

Main Results:

• The proposed MATA method demonstrated superior performance compared to existing approaches in simulations.
• The method achieved root-n consistency for the feature effect vector estimator under regularity conditions.
• Successful application in a real-world example of annotating lung cancer recurrence in a Veteran Health Administration EHR cohort.

Conclusions:

• The MATA method offers an efficient and accurate solution for annotating clinical event times from EHR data.
• This automated approach facilitates the development of robust prediction models for precision medicine research.
• The findings highlight the potential of leveraging multi-dimensional EHR data for advancing clinical outcome research.