Imputation of missing values for electronic health record laboratory data

Area of Science:

• Biomedical Informatics
• Health Data Science
• Clinical Prediction Modeling

Background:

• Electronic Health Records (EHR) contain valuable laboratory data for clinical prediction models.
• Missing laboratory data in EHR can introduce estimation bias and degrade model performance.
• Imputation methods are crucial for addressing missingness in EHR data.

Purpose of the Study:

• To demonstrate the utility of imputation methods for EHR laboratory data.
• To characterize missingness patterns and compare imputation algorithms.
• To assess the impact of comorbidity data on imputation performance.

Main Methods:

• Utilized two real-world EHR cohorts: ischemic stroke (Geisinger) and heart failure (Sutter Health).
• Characterized missingness patterns in laboratory variables.
• Simulated missing data under arbitrary and monotone mechanisms.
• Compared cross-sectional and multi-level multivariate imputation algorithms.
• Assessed the incorporation of latent comorbidity information.

Main Results:

• Missingness patterns in EHR laboratory data were non-random and associated with patient comorbidity data.
• Multi-level multivariate imputation algorithms exhibited lower imputation error compared to cross-sectional methods.
• Incorporating latent comorbidity information showed potential for improving imputation performance.

Conclusions:

• Imputation is essential for leveraging EHR laboratory data in prediction models.
• Multi-level imputation methods are more effective than cross-sectional approaches for EHR laboratory data.
• Patient comorbidity data provides valuable latent information that can enhance imputation accuracy.