Increasing the Density of Laboratory Measures for Machine Learning Applications.

Area of Science:

• Biomedical Informatics
• Translational Medicine
• Data Science in Healthcare

Background:

• Missing laboratory values in Electronic Health Records (EHR) are common and impact translational medicine research.
• Existing imputation techniques often fail to account for the non-random nature of missingness in EHR data.
• Developing specialized imputation strategies for EHR is crucial for improving data quality and predictive modeling in healthcare.

Purpose of the Study:

• To develop and evaluate an adaptive imputation strategy for missing laboratory values in EHR.
• To address the non-random missingness of laboratory data by incorporating patient clinical profiles.
• To enhance the accuracy and predictive power of healthcare analytics by improving data imputation.

Main Methods:

• Analysis of laboratory measures from three distinct EHR cohorts: *Clostridioides difficile* (Cdiff) infection, inflammatory bowel disease (IBD), and osteoarthritis (OA).
• Extraction and filtering of Logical Observation Identifiers Names and Codes (LOINC), excluding those with over 75% missingness.
• Development of a hybrid adaptive imputation strategy involving clustering of patient comorbidity patterns and independent imputation within clusters, compared to standard imputation.

Main Results:

• The study analyzed 67,445 patients across the three cohorts.
• The adaptive imputation strategy demonstrated the most significant improvement for laboratory measures with higher levels of missingness.
• The best root mean square error (RMSE) differences achieved were -35.5 (Cdiff), -8.3 (IBD), and -11.3 (OA).

Conclusions:

• An adaptive imputation strategy leveraging patient clinical profiles can effectively improve the imputation of missing EHR laboratory values.
• This method is particularly beneficial when dealing with laboratory codes exhibiting high degrees of missingness.
• The proposed strategy enhances data quality for downstream analyses in translational medicine and healthcare applications.