Artificial Intelligence Can Help Extract Dysplasia Diagnoses From Electronic Health Records

Natural language processing (NLP) is a branch of artificial intelligence (AI) systems that involves teaching computers the ability to understand written text like how humans do. The primary aim of this study was to create an NLP algorithm to identify dysplasia in pathology reports for patients with Barrett’s esophagus (BE).

The study randomly selected a cohort of 1000 patients with BE from the national Veterans Affairs electronic health records database. The pathology reports of these patients were manually reviewed by 2 investigators and classified for the presence of BE and grade of dysplasia. The NLP development set included 600 patients, and the validation set included the remaining 400 patients. 

Compared with human reviewers, the NLP algorithm identified dysplasia in the development set with a 98.0% accuracy and 91.7% recall (precision, 93.2%; F-measure, 92.4%). In the validation set, NLP detected dysplasia with 98.7% accuracy and 92.3% recall (precision, 100.0%; F-measure, 96.0%).  The most common reason for false negatives (4 of 8) in both sets was excessive line breaks in pathology reports.