Hidden in plain sight: bias towards sick patients when sampling patients with sufficient electronic health record data for research
- Equal contributors
1 Department of Anesthesiology, Columbia University, New York, NY, USA
2 Department of Biomedical Informatics, Columbia University, New York, NY, USA
3 Department of Biostatistics, School of Public Health, Columbia University, New York, NY, USA
BMC Medical Informatics and Decision Making 2014, 14:51 doi:10.1186/1472-6947-14-51Published: 11 June 2014
To demonstrate that subject selection based on sufficient laboratory results and medication orders in electronic health records can be biased towards sick patients.
Using electronic health record data from 10,000 patients who received anesthetic services at a major metropolitan tertiary care academic medical center, an affiliated hospital for women and children, and an affiliated urban primary care hospital, the correlation between patient health status and counts of days with laboratory results or medication orders, as indicated by the American Society of Anesthesiologists Physical Status Classification (ASA Class), was assessed with a Negative Binomial Regression model.
Higher ASA Class was associated with more points of data: compared to ASA Class 1 patients, ASA Class 4 patients had 5.05 times the number of days with laboratory results and 6.85 times the number of days with medication orders, controlling for age, sex, emergency status, admission type, primary diagnosis, and procedure.
Imposing data sufficiency requirements for subject selection allows researchers to minimize missing data when reusing electronic health records for research, but introduces a bias towards the selection of sicker patients. We demonstrated the relationship between patient health and quantity of data, which may result in a systematic bias towards the selection of sicker patients for research studies and limit the external validity of research conducted using electronic health record data. Additionally, we discovered other variables (i.e., admission status, age, emergency classification, procedure, and diagnosis) that independently affect data sufficiency.