Mitigating bias in the analysis and inferences from using longitudinal EHR data in disease outcomes research

Background The electronic health record (EHR) provides an opportunity for extracting a wealth of up-to-date real-world longitudinal data. Although the main limitations of using EHR data have been well-recognized and well-described, under-recognized factors may threaten the reliability of inferences made regarding the impact of EHR variables on chronic disease outcomes. A problem not well-recognized is the impact of routinely acquired variables that are documented with every patient encounter regardless of the reason for the encounter such as vital signs, height, and body weight. Methods We utilized a landmark approach to identify occurrence of 10 cardiovascular-related disease outcomes after a 5-year observation period during which all body weights recorded in the EHR were used in multivariate cox regression modeling to identify the strongest of 9 weight-based predictor variables for each of the 10 disease outcomes. Results We found that the number of recorded weights, as an independent variable, was the strongest predictor for all 10 cardiovascular-related disease outcomes when compared to all other weight-based variables (lowest weight, highest weight, average weight, last weight, absolute weight change, maximum weight change, weight fluctuation, and weight cycling) as well as BMI. The findings demonstrate the importance of recognizing and accounting for the number of times a more frequently measured clinical variable, such as body weight, is recorded as it is critical to determine the true impact of other similar variables on disease outcomes when conducting longitudinal analysis of EHR data.