|
Abstract:
|
Electronic Health Records include a wide variety of clinical and administrative data that can be used to describe patient phenotypes. However, available measures vary systematically across healthcare settings, necessitating novel approaches to generating phenotypes that make best use of the available data. Although healthcare systems-based research networks have great promise as research resources, providing detailed data on large, representative patient populations, the problem of between-site data inconsistencies is particularly acute for such networks. For instance, measurement of biomarkers, medication prescribing patterns, and use of diagnosis codes may vary systematically across sites within the network. We propose a Bayesian latent variables approach, harnessing existing evidence and expert opinion on the relationships among data elements, to facilitate estimation of a common phenotype across healthcare systems. Through simulation studies we demonstrate improved efficiency associated with using this latent phenotype compared to traditional missing data approaches. We applied our new approach in a study of pediatric diabetes, using data from a network of children's hospitals.
|
