PTH-36 identification & service evaluation of a primary sclerosing cholangitis cohort using natural language processing

Livingstone, Robert, Phan, Hang, Borca, Florina, Sarkar, Srishti, Minto, Moeed, Dixey, Annie, Qaisar, Razzi, Patel, Janisha, Stammers, Matthew and Gwiggner, Markus (2021) PTH-36 identification & service evaluation of a primary sclerosing cholangitis cohort using natural language processing. Gut, 70, A188-A189. (doi:10.1136/gutjnl-2021-BSG.351).

Abstract

Introduction: primary sclerosing cholangitis (PSC) is a rare and difficult to treat condition. PSC is strongly associated with malignancy, therefore screening and surveillance are paramount. PSC however does not have a unique UK ICD-10 diagnostic code, hence reliable patient cohort identification and thorough service evaluation is challenging. We used natural language processing (NLP) to identify the PSC patient cohort at University Hospital Southampton (UHS) and audited associated outcomes against recently updated British Society of Gastroenterology (BSG) management guidelines.

Method: records of all patients with PSC at our institution between 2008-2020 were identified using our NLP methodology. We used fuzzy matching to analyse clinical records, and tokenized and lemmatized key paragraphs to identify key diagnostic patterns and exclude diagnostically uncertain or exclusive sentences. Anonymised discharge summaries, clinic letters, radiology reports, endoscopy records and histology were extracted and digitally trawled to identify the cohort characteristics.

Results: we identified 125 patients with PSC followed-up at UHS. 39.2% (49) of these patients were missed in a parallel criterion-based review of case notes.
We calculated an age-standardised point prevalence of 12.52 cases per 100,000 patients, 124% higher than typically cited UK figures. Service evaluation revealed high rates of clinic follow-up however lower than recommended rates of screening with colonoscopy and imaging (see Table 1). Introduction of a combined PSC/IBD clinic as a targeted service delivery intervention is addressing this shortfall with significant impact after 1 year. [PTH-36 Table 1 not included].

Conclusions: PSC cohorts are difficult to identify due to a lack of a UK clinical code. An NLP based methodology proved highly effective at identifying all cases within our institution, with a 64.5% increase compared to conventional methods. This allowed rapid patient cohort identification and conversion of unstructured data to clinically useful structured data and could be reproduced at other institutions and for other diseases.

This record has no associated files available for download.

Contributors

Author: Matthew Stammers

Download statistics