P49 Systematic review of natural language processing applied to gastroenterology & hepatology

Abstract

<h3>Introduction</h3> Unstructured clinical free text contains invaluable information that computers can only readily understand using Natural language processing (NLP). The purpose of this systematic review is to provide clinicians with an accessible understanding of NLP as it has been applied within gastroenterology up to the present to lay the foundations for future study in this new and promising field. <h3>Methods</h3> Articles from seven scholarly databases (ACM Digital Library, Arxiv, Embase, IEEE Explore, Pubmed, Scopus and Google Scholar) were searched for studies published 2015–2023 describing the development or application of NLP methods within gastroenterology and hepatology. Studies lacking a description of appropriate validation or NLP methods were excluded, as were studies unavailable in English, focused on non-gastrointestinal diseases and duplicates. Two independent reviewers extracted study information, clinical/algorithm details, and relevant outcome data. Quality was appraised using a checklist of quality indicators for NLP studies. <h3>Results</h3> 54 studies were identified utilising NLP in Endoscopy, Inflammatory Bowel Disease, Gastrointestinal Bleeding, Liver and Pancreatic Disease. n=21 (38.9%) of studies focused on colonoscopy, n=13 (24.1%) on liver disease, n=7 (13.0%) on inflammatory bowel disease, n=4 (7.4%) on gastroscopy, n=4 (7.4%) on pancreatic disease and 2 (3.7%) studies focused on endoscopic sedation/ERCP. Classification tasks accounted for n=32 (59.2%) of studies, followed by task automation n=15 (27.8%) and prediction tasks n=6 (11.1%). Algorithm precision ranged from 28.5–100% and recall 25–99.9% across various tasks. However, only n=3 (5.7%) of studies scored a low risk of bias across all assessed quality domains, and only n=13 (n=24%) of studies featured robust external validation. Open-source code was only available for n=5 (9.3%) of studies. Only one (1.9%) study was deployed in production as part of a human-in-the-loop system. <h3>Conclusions</h3> Opportunities for future progress in this novel field abound. However, they require clinician engagement, collaboration and sharing of appropriate datasets and code. Despite some progress, particularly in colonoscopy, we will need to see validated, trusted, semi-autonomous NLP systems deployed widely before significant clinical benefits are realised.

Ähnliche Arbeiten