Biases in Race and Ethnicity Introduced by Filtering Electronic Health Records for "Complete Data": Observational...

Area of Science:

• Biomedical Informatics
• Health Equity Research
• Data Science in Healthcare

Background:

• Integrated clinical databases from national biobanks are crucial for disease research.
• Data quality filters are essential for building clinical cohorts but can introduce systemic biases correlated with race and ethnicity.
• These biases may unintentionally exclude minoritized populations from research.

Purpose of the Study:

• To examine race and ethnicity biases introduced by applying common data filters to clinical records databases.
• To evaluate if these filters disproportionately exclude minoritized groups from research cohorts.
• To understand the impact of data filters on sample representation across racial and ethnic groupings.

Main Methods:

• Applied 19 common data filters to electronic health record datasets from 4 geographically diverse locations, totaling nearly 12 million patients.
• Analyzed the variation in sample drop-off across self-reported racial and ethnic groups for each filter individually.
• Included filters covering demographics, medication records, visit details, and observation periods.

Main Results:

• The observation period filter significantly reduced data availability for all groups.
• White individuals consistently had higher data availability compared to other racial groups after filter application.
• Black or African American individuals were most impacted by filters across three of the four datasets.
• The 'All of Us' dataset showed minimal deviation in sample representation after filter application.

Conclusions:

• Data filters can disproportionately affect the availability of data for minoritized racial and ethnic populations.
• Researchers must consider and mitigate unintentional biases introduced by data filters in health research.
• Recommends using only necessary filters, exploring bias minimization techniques, and disclosing sample sizes by race/ethnicity before and after filtering.