Estimating summary statistics for electronic health record laboratory data for use in high-throughput phenotyping...

Area of Science:

• Biomedical Informatics
• Data Science in Healthcare
• Statistical Modeling

Background:

• Electronic health record (EHR) data are influenced by clinical care processes, introducing biases.
• Existing methods for summarizing laboratory data may not adequately capture underlying information.
• Understanding and mitigating these biases is crucial for accurate data interpretation.

Purpose of the Study:

• To develop a novel algorithm, PopKLD, for summarizing raw EHR laboratory data.
• To reduce and manage biases introduced by clinical care processes.
• To create an intuitive, continuous summary of laboratory data for downstream applications like phenotyping.

Main Methods:

• Constructed the PopKLD algorithm based on information criterion model selection.
• Developed PopKLD-CAT to transform continuous summaries into categorical data.
• Evaluated methodology using laboratory data from primary and intensive care settings.
• Assessed performance in a phenotyping task comparing PopKLD-derived predictions to clinical gold standards.

Main Results:

• PopKLD preserves known physiologic features often lost in traditional summaries (mean, standard deviation).
• The PopKLD-CAT algorithm effectively generates categorical summaries for applications like topic modeling.
• PopKLD-based summaries demonstrated substantially better disease state prediction accuracy compared to mean/SD methods.
• Different clinical contexts and laboratory measurements necessitate distinct statistical summaries.

Conclusions:

• The PopKLD algorithm provides a more informative summary of EHR laboratory data than traditional methods.
• This approach effectively reduces and copes with health care process biases.
• The PopKLD and PopKLD-CAT algorithms offer valuable tools for data analysis and phenotyping in healthcare.
• The choice of statistical summary should be tailored to specific clinical contexts and data characteristics.