Integrated analysis for electronic health records with structured and sporadic missingness.

Area of Science:

• Health Informatics
• Data Science
• Clinical Research

Background:

• Electronic Health Records (EHRs) often contain structured and sporadic missing data, hindering integrated analysis.
• Missingness in EHRs is a common challenge when combining heterogeneous datasets for clinical applications.

Purpose of the Study:

• To propose and validate a novel imputation method, Macomss, for addressing structured and sporadic missing data in EHRs.
• To enhance the utility of integrated EHR data for downstream clinical applications and population health research.

Main Methods:

• Demonstrated structured and sporadic missingness mechanisms in EHR data integration.
• Introduced the Macomss imputation framework with theoretical guarantees.
• Conducted extensive simulations and validated using Duke University Health System (DUHS) EHR data.

Main Results:

• Macomss outperformed existing imputation methods in simulation studies.
• Achieved lowest imputation errors and superior/comparable downstream prediction performance on DUHS datasets.
• Demonstrated robustness in preserving data integrity for integrated analyses.

Conclusions:

• Macomss effectively imputes structured and sporadic missing data in integrated EHR analyses.
• The method enhances the robustness and generalizability of clinical predictions.
• Provides a theoretically guaranteed and practically meaningful solution for multi-EHR dataset analysis, advancing population health research.