Fighting Selection Bias in EHR Data Analyses at the 2020 AMIA Annual Symposium

Retrospective analyses of electronic health records (EHRs) and other health data sources are increasingly common as investigators seek to employ data collected during routine health care delivery to learn about health practices and outcomes. The data analysis process typically includes cohort selection, in which a group of patients (and their corresponding data) are identified from within a health organization’s overall population of patients. However, with such systems, the lack of randomization combined with the high level of expected interdependence between variables can produce cohorts that are highly skewed in ways that are unexpected to the analyst.

David Gotz will present the VACLab’s work on this topic at the 2020 AMIA Annual Symposium. In the team’s podium abstract, Gotz et al. describe their work developing new methods and tools to make selection bias effects more transparent to analysts. The extended abstract provides a high-level overview of a new form of visualization-based selection bias report which, when integrated within a cohort selection system, can make it easy for analysts to discover unexpected shifts in cohort variable distributions between a selected cohort and a baseline population which might otherwise go unnoticed.

More details and a PDF of the abstract can be found on the article’s page on this website.

Reference

David Gotz, Jonathan Zhang, Smiti Kaul, Georgiy Bobashev, David Borland. Visual Analytics to Combat Selection Bias in Retrospective EHR Data Analyses. American Medical Informatics Association (AMIA) Annual Symposium Podium Abstract (2020).