A self-supervised framework for laboratory data imputation in electronic health records.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Electronic Health Records (EHRs)

Background:

• Laboratory data in EHRs is valuable but underutilized due to missing values.
• Current imputation methods fail to fully leverage clinical history and scale effectively.
• High missingness in laboratory values limits their use in diagnostics, treatment, and research.

Purpose of the Study:

• To develop a scalable and effective framework for imputing missing laboratory values in EHRs.
• To leverage complete patient clinical histories for more accurate laboratory data imputation.
• To improve the utilization of real-world data (RWD) for healthcare applications.

Main Methods:

• Developed Laboratory Imputation Framework for EHRs (LIFE), a self-supervised learning framework.
• Utilized a multi-head attention architecture to jointly model all laboratory data.
• Incorporated additional EHR variables (diagnoses, medications) for clinical contextualization.

Main Results:

• LIFE achieved superior or equivalent performance across 23 out of 25 laboratory tests compared to baseline methods.
• Demonstrated improved performance in a downstream adverse event detection task (7 out of 9 cases).
• Validated on a large-scale dataset of over 1 million oncology patients.

Conclusions:

• LIFE accurately estimates missing laboratory values, enhancing RWD utilization in healthcare.
• This framework supports the development of better clinical models and informed decision-making.
• Potential for improved patient outcomes through more complete and reliable health data.