Leveraging error-prone algorithm-derived phenotypes: Enhancing association studies for risk factors in EHR data.

Area of Science:

• Biomedical Informatics
• Health Data Science
• Clinical Epidemiology

Background:

• Multiple computable phenotypes are increasingly derived from electronic health records (EHR).
• EHR-based association studies commonly utilize a single phenotype, potentially introducing bias.
• Phenotyping error can impact the accuracy and efficiency of EHR-based research.

Purpose of the Study:

• To develop a novel method for simultaneously utilizing multiple EHR-derived phenotypes.
• To reduce bias stemming from phenotyping errors in EHR data.
• To improve the efficiency and accuracy of phenotype/exposure association studies.

Main Methods:

• A method combining multiple algorithm-derived phenotypes with validated outcomes was developed.
• The approach employs a statistically efficient seemingly unrelated regression framework.
• Performance was assessed via simulation studies and real-world EHR data analysis (colon cancer recurrence).

Main Results:

• Substantial bias reduction was achieved compared to single-phenotype methods, particularly when no single phenotype was uniformly superior.
• Estimation efficiency increased by up to 30% compared to using only one algorithm-derived phenotype.
• The method demonstrated effectiveness in integrating multiple phenotypes for improved statistical accuracy.

Conclusions:

• The proposed method effectively integrates multiple phenotypes from EHR data.
• It offers a robust alternative to single-surrogate bias correction methods.
• The approach enhances bias reduction, statistical accuracy, and efficiency in EHR-based association studies.