Shrinking the Haystack: An Approach to Identifying Idiopathic Pulmonary Fibrosis in the Electronic Health Record using a Computable Phenotype

Abstract

Abstract Background: Computable phenotypes are computerized search queries that allow efficient identification of specific groups of individuals (e.g., that may meet eligibility criteria for a clinical trial). Heterogeneous clinical syndromes challenge this approach because disease definitions and sub-phenotypes evolve, and diverse phenotypes may be needed for various applications (“use cases”) for diverse research aims. Herein we describe the development and validation of a computable phenotype for the rare disease idiopathic pulmonary fibrosis (IPF), that addresses its evolving terminology and variable use cases. The goal of this study was to develop and execute a single computable phenotype for IPF using standard data architecture, and to evaluate it for different use cases, each with its own gold standard for validation. Methods: The PaTH PCORnet Clinical Research Network (PaTH) IPF Working Group developed the candidate IPF computable phenotype and executed it against the Penn State PaTH to Health source population of 588,000 patients with an electronic medical record at Penn State Hershey Medical Center between January 1, 2011 and December 31, 2015. We established a consensus clinician diagnosis and performed duplicate (2-person parallel) chart review on a 100% sample with discrepancy adjudication. We evaluated the computable phenotype performance for two use cases, each with a separate gold standard: the Inclusive Use Case [gold standard defined as IPF, familial pulmonary fibrosis (FPF), or combined pulmonary fibrosis and emphysema (CPFE)] and the Restrictive Use Case (gold standard defined as IPF, but not FPF nor CPFE). Results: The IPF computable phenotype yielded an IPF Cohort (N=157) and an estimated population prevalence of 26.7/100,000. The computable phenotype had positive predictive values (PPV) for the Inclusive Use Case and Restrictive Use Case of 57% (89/157) and 47% (74/157), respectively, and an estimated population prevalence of 15.1 and 12.6/100,000, respectively. Conclusions: These findings demonstrate the utility of a single computable phenotype that can be validated against different gold standards depending on the intended health care or research use case. In a disease where there is no discrete biomarker, this provides a flexible approach to meet diverse clinical research needs. Trial registration: N/A