Phenotypic characterisation of suspected small bowel Crohn’s disease with natural language processing of MRE reports

Introduction Location and Behaviour are important considerations in the phenotypic characterisation of Crohn’s disease. Although enteroscopy and video capsule endoscopy allow direct visualisation of the small bowel mucosa, radiographic examinations (CT and MRI) are less invasive and provide additional information in stricturing and penetrating disease. Such studies tend to be reported in semi-structured free-text which presents a challenge for automated classification of large cohorts. We set out to write a natural language processing (NLP) algorithm to extract phenotypic characteristics from the magnetic resonance enterography (MRE) reports of patients with suspected Crohn’s disease to help define our local IBD population. Methods Reports from 904 consecutive MRE scans in our hospital trust were anonymised and imported into a Microsoft Excel datasheet. A senior gastroenterology trainee encoded phenotypic characteristics of Crohn’s disease into new attribute columns to provide a ‘ground truth’ reference. The anonymised raw dataset was also imported into a Python pandas dataframe for NLP. The NLP algorithm: Separated all words by a single space, converted to lower case and removed superfluous punctuation Corrected spelling of key words using a Levenshtein threshold Identified disease phenotypes by matching regular expressions Excluded negating phrases with pre– and post-concept regular expressions Excluded positive diagnoses documented within a family history Identified the location of Crohn’s pathologies from intersection with a set of GI anatomical locations Linked the outputs to the ground truth reference data to determine sensitivity, specificity, positive and negative predictive values

Citation

Materacki, L., Collins, D., Zeki, S., et al. (2021). Phenotypic characterisation of suspected small bowel Crohn’s disease with natural language processing of MRE reports (P126). Gut, 70(Suppl 1), A107–A108. DOI: 10.1136/gutjnl-2020-bsgcampus.201

DOI

10.1136/gutjnl-2020-bsgcampus.201

URI

https://hdl.handle.net/20.500.14709/850

License

Available on journal website. Click doi.

Usage rights

Available on journal website. Click doi.