Handling missingness in electronic health record laboratory data for secondary use: a decision-making process

Abstract

Abstract Electronic health record (EHR) systems have become widely adopted, prompting a growing interest in leveraging EHR data for secondary use to improve clinical knowledge and enhance patient care, for example, by prediction modelling. A common challenge in EHR-derived healthcare datasets is missing data, which is frequently insufficiently addressed in publications. Missing data must be handled appropriately to ensure the quality of prediction models. We propose that combined data-centred and clinical domain expertise is needed for this. This paper demonstrates a stepwise decision-making process for handling missing data in EHR-derived laboratory measurements using our case study on predicting frailty status as an example. The stepwise process consisted of variable selection, evaluating different imputation methods (including the creation of missing indicator variables) and determining the preferred method per variable. Decisions were made by the data and clinical domain experts together, based on a combination of data-centric, biological and clinical considerations. In our example, the stepwise process resulted in three preferred methods to handle missing data: imputation with the median based on gender and age, random forest imputation and the missing indicator method. In this paper, we illustrate the importance of the stepwise, multidisciplinary decision-making process to identify methods for handling missing data within the context of structured laboratory test results. This process is important for making informed, context-specific choices about imputation. An essential next step, beyond the scope of this paper, is to evaluate the impact of different imputation strategies on the downstream performance of prediction models.