Abstract
BACKGROUND AND AIMS: Our understanding of the epidemiology and natural history of gastroesophageal precancerous conditions is limited by a lack of robust analyses using large-scale individual-level data. We aimed to develop and validate a natural language processing (NLP) algorithm to identify and phenotype esophageal and gastric precancerous conditions and cancer and apply this to the Million Veteran Program (MVP), a uniquely powerful nationwide genomic biobank linked to individual-level electronic health record data. METHODS: We identified 121,808 individuals in MVP who underwent upper endoscopy with biopsies. From these, 426 pathology notes from 426 individuals were used to develop and manually validate an NLP rule-based algorithm identifying intestinal metaplasia, dysplasia, and tumors of the stomach or esophagus. Anatomic subsite and "qualifier" terms were also evaluated (eg, dysplasia grade). Performance metrics were calculated. RESULTS: The algorithm identified all prespecified conditions with excellent accuracy, ranging from 97.6% to 100% (Bonferroni-corrected 95% lower bound 94.5%-98.5%). For gastric intestinal metaplasia, the algorithm achieved 91.7% precision, or positive predictive value and 86.8% recall, or sensitivity (F1 score 89.2%), with 99.0% specificity and 98.3% negative predictive value; while for Barrett's esophagus, it achieved 98.9% precision and recall, or sensitivity (F1 score 98.9%), with 99.6% specificity and 99.6% negative predictive value. When applied to the full MVP cohort (N = 121,808), 13.2% had gastric intestinal metaplasia (mean age 65 years) and 14.5% had Barrett's esophagus (mean age 64 years). CONCLUSION: This study confirms the ability to use NLP on large-scale unstructured data linked to robust genetic and clinical data for future gastroesophageal precancer analyses and to inform targeted prevention and/or early detection interventions (eg, endoscopic surveillance).