Implicit bias of encoded variables: frameworks for addressing structured bias in EHR-GWAS data.

Area of Science:

• Genomic Medicine
• Biomedical Informatics

Background:

• Genome-wide association studies (GWAS) historically required large, homogeneous cohorts for discovery.
• Translating genomic discoveries into clinical utility necessitates understanding disease presentation within real-world data.

Purpose of the Study:

• To explore the potential of electronic health record (EHR) data for scalable, lifelong phenotyping.
• To identify and address inherent biases within EHR data for accurate phenotypic characterization.
• To propose frameworks for minimizing and uncovering encoded bias in large-scale EHR analyses.

Main Methods:

• Reviewing the evolution of cohort requirements in genetic studies.
• Analyzing the nature of clinical decision-making and its encoding in EHR data.
• Illustrating potential biases and their compounding effects over time within EHR analyses.

Main Results:

• EHR data enables scalable phenotyping, capturing comprehensive patient context.
• Clinical judgments and decisions outside direct system input introduce bias into EHR data.
• Algorithmic approaches assuming unbiased EHR variables risk generating compounded biased conclusions.

Conclusions:

• Phenotype definition in EHR and biobank-based analyses demands rigorous attention to potential biases.
• Frameworks are needed to systematically minimize and uncover encoded bias in large-scale EHR data.
• Addressing bias is essential for leveraging EHR data to its full potential in clinical and genomic research.