Natural Language Processing Accurately Identifies Dysphagia Indications for Esophagogastroduodenoscopy Procedures in a Large US Integrated Healthcare System: Implications for Classifying Overuse and Quality Measurement

摘要

Recent evidence suggests almost half of repeat esophagogastroduodenoscopy procedures (EGDs) are overused; this prior research relied on administrative data that are often inaccurate. Our primary objective was to determine and compare the accuracy of natural language processing and administrative data to manual chart review to identify dysphagia indications for EGD procedures within the national VA healthcare system. From 396,856 EGD notes identified from 2008-2014, we classified 119,920 as “index” procedures in 2010-2012. We compared the performance of our NLP to ICD codes to correctly identify dysphagia indications in the index EGD procedures and in repeat EGD procedures. We used linked pathology data to describe esophageal biopsies performed during these EGDs. NLP performed significantly better and identified significantly more index and repeat EGD procedures with dysphagia indications than ICD codes, which has critical implications for determining appropriateness of EGD procedures.