(P-554) Identification of Patients with Nonalcoholic Steatohepatitis (NASH) in an Electronic Health Record (EHR) Database

Background: Electronic health record (EHR) databases represent useful real-world data sources for studying the epidemiology of NASH. NASH is a relatively asymptomatic disease without a currently approved treatment, usually presenting as an incidental finding of abnormal liver biochemistries or abnormal liver imaging. Natural language processing (NLP) applied to full text clinical notes may improve case finding of ‘presumptive’ NASH patients above the use of ICD-10 diagnosis code K75.81 alone.

Objectives: To develop a NASH-identification algorithm within an EHR database using a combination of structured (diagnosis codes, procedure codes, labs) and NLP data.

Methods: NASH patients were identified between 2016 and 2019 within Optum’s EHR Research Database by structured and NLP data. The algorithm required patients to have either evidence of NASH or non-alcoholic fatty liver disease (NAFLD) with supporting evidence of steatohepatitis, fibrosis, cirrhosis, a positive liver imaging study, or a positive liver biopsy. To limit false positives, exclusion criteria were applied to remove patients with other causes of fatty liver disease. NASH, NAFLD, steatohepatitis, fibrosis, and cirrhosis were identified by a diagnosis code or an NLP affirmation. FibroSure results were derived from the NLP data and all other fibrosis scores were calculated from the structured data. Positive image/biopsy results were identified using NLP data.

Results: The algorithm identified 93,204 patients with NASH including 77,728 (83%) identified by structured data alone, 12,878 (14%) identified by a combination of structured and NLP data, and 2,598 (3%) identified by NLP data alone. Patients who were identified by a combination of structured and NLP data appeared to have a higher burden of some comorbidities consistent with NASH (e.g. type 2 diabetes, cirrhosis, esophageal varices, ascites, portal hypertension). Additionally, these patients had less healthy biochemistry and more healthcare utilization.

Conclusions: NASH patients may be identified using a combination of structured and NLP data in EHR databases. NLP data identifies additional patients with undiagnosed or unrecognized NASH where the sample size increased appreciably when NLP was used in combination with structured data. NLP data helped to identify NASH-related conditions including NAFLD, steatohepatitis, fibrosis, cirrhosis, and image/biopsy results. With algorithms such as this, EHR databases may be used to better understand the epidemiology of diseases that require further characterization in population-based settings.