Imputation of missing aggregate EHR audit log data across individual and multiple organizations

Area of Science:

• Health Informatics
• Data Science
• Electronic Health Records (EHR)

Background:

• Missing data in EHR systems poses challenges for accurate analysis.
• Imputation strategies are crucial for handling missing EHR data.
• Evaluating different imputation methods is essential for reliable healthcare analytics.

Purpose of the Study:

• To compare naive and machine learning imputation strategies for EHR data.
• To explore subgrouping criteria for optimizing imputation.
• To assess the performance and feasibility of in-house imputation implementation.

Main Methods:

• Conducted imputation experiments on EHR audit log data from diverse organizations.
• Classified variables based on coefficient of variation and missing percentage.
• Evaluated imputation performance using R²-values, selecting the most robust model.

Main Results:

• Rolling average time-series imputation showed more consistent R² across organization sizes.
• XGBoost (machine learning) performed slightly better in large organizations.
• Single-site organizations achieved higher R² using their own data for imputation compared to merged data.

Conclusions:

• A systematic variable classification and tailored imputation strategy is effective for EHR data.
• The rolling average time-series method outperformed non-time-series machine learning imputation.
• Organization size did not significantly impact the imputation process; merging diverse site data did not improve imputation.