PTU-32 Natural Language Processing Driven Comparison of Small Bowel MRI and Capsule Endoscopy Reporting for Crohn’s

IntroductionSmall Bowel Capsule Endoscopy (SBCE) has an established role in the diagnosis/management of small bowel Crohn’s disease (CD). Previous work has suggested its diagnostic yield is comparable to that of Small Bowel MRI (SBMR), but the narrative nature of reporting makes this challenging to evaluate formally.MethodsAnonymised data from the electronic health record were systematically extracted from patients undergoing SBCE at our hospital between Jan 2016- Jan 2020 and categorised according to indication. A subgroup with suspected or established Crohn’s disease was identified, and corresponding SBMR and calprotectin results were obtained. Natural language processing (NLP) techniques were used to compare the content and diagnostic accuracy of the reports.ResultsOut of 1016 patients undergoing SBCE, 494 patients were suspected as having small bowel CD. Of this cohort, 133 underwent SBCE within 180 days of SBMR. 59 patients had corresponding faecal calprotectin (FC) measured (Mean: 318.12+/-138.39).Tokenisation demonstrated that the style and content of the SBCE vs SBMR reports were very different in form and structure. The resultant ‘word-bags’ revealed that SBCE reports had significantly more ‘diagnostic’ weight, as described in Table 1:Abstract PTU-32 Table 1 Small Bowel Test ‘Crohn’s’ mentioned as the potential diagnosis A positive finding of small bowel ‘inflammation/ulceration’ described A positive finding of a ‘stricture’ described Terms of uncertainty used in report conclusion A positive finding of ‘thickening’ described SBMR 3.75%(n=5) 6.02%(n=8) 3.01%(n=4) 11.28%(n=15) 6.77%(n=9) SBCE 14.29%(n=19) 17.29%(n=23) 3.76%(n=5) 2.26%(n=3) 0.75%(n=1) χ 2 & p 8.98(p=0.003) 8.22(p=0.004) 0.12(p=0.115) 8.58(p=0.003) 6.65(p=0.009) Mean FC was significantly higher in patients with the term ‘ulceration’ reported on SBCE (703.5+/-353) compared to ‘no ulceration’ (265.2+/-148.38) (p=0.044) linking the word ‘ulceration’ with more severe disease descriptions.ConclusionsNLP is a powerful novel tool to compare narrative diagnostic accuracy between SBCE and SBMR not previously described. FC is positively correlated with small bowel ‘ulceration’ linking descriptive, diagnostic terms to biomarkers. According to our results, SBCE alone rather than SBMR is necessary to investigate small bowel Crohn’s disease unless full-thickness stricturing is suspected from clinical symptoms.