B0105 - Real World Validation of an Artificial Intelligence Characterization Support System for Prediction of Polyp Histology in Colonoscopy: Interim Analysis of a Prospective Multicenter Study

Award: Presidential Poster Award

Weiquan James Li, MBBS, MMed¹, Clement Wu, MBBS, MRCP², Christopher Jen Lock. Khor, MBBS², Raymond FH. Liang, MBBS, MRCP³, Jonathan Wei Jie Lee, MBBS, MRCP⁴, Jimmy So, MBChB, FRCSEd, FRCSG, FAMS, MPH⁴, Weida Chew, MBBS, MRCP^3
1Changi General Hospital, Singapore, Singapore; ²Singapore General Hospital, Singapore, Singapore; ³Tan Tock Seng Hospital, Singapore, Singapore; ⁴National University Hospital, Singapore, Singapore

Introduction: Colonoscopy is the gold standard for screening of colorectal cancer. Diminutive and hyperplastic polyps do not increase a patient’s risk for colorectal cancer. Prediction of hyperplastic polyp histology is crucial for the resect and discard strategy, which saves cost and decreases procedure time in colonoscopy. This study aims to validate the performance of computer-aided diagnosis (CADx) in distinguishing between hyperplastic and neoplastic polyps during colonoscopy in a real-world setting.

Methods: We conducted a prospective multicentre study comparing CADx (Fujifilm Corp., Tokyo) with endocopist optical prediction of polyp histology. We first recorded the optical diagnosis according to the NICE classification by the endoscopist. Following this, the CADx tool was switched on and its prediction recorded. Imaged polyps were resected for histological analysis, which formed the gold standard. Primary outcome was diagnostic accuracy (defined by sensitivity for diagnosis of hyperplastic polyps and concordance rate). Bowel preparation, polyp size and difficulty in location were recorded for subgroup analysis.

Results: 414 patients were assessed for eligibility across 4 large tertiary institutions in Singapore between February 2021 and June 2022. 625 polyps (303 hyperplastic, 322 neoplastic) were detected in 257 patients. Concordance rates for CADx and endoscopist predictions were 74.1% [95% confidence interval (CI) 70.5%-77.5%] and 73.1% (95% CI 69.5%-76.6%%), respectively (p=NS). Sensitivity for diagnosis of hyperplastic polyps was 84.2% (95% CI 79.6%-88.1%) and 77.6% (95% CI 72.4%-82.1%) for CADx and endoscopists, respectively (p=NS). CADx also showed superior performance in predicting hyperplastic histology in diminutive polyps compared to endoscopist optical prediction using the NICE classification (sensitivity 81.7%; 95% CI 76.2%-86.4%, versus 76.3%; 95% CI 70.4%-81.5%, respectively). Diagnostic accuracy was similar when analysed according to bowel preparation and difficulty in polyp location during colonoscopy (defined as polyp location behind fold, around bend, or unable to position at 6 o’clock).

Discussion: CADx showed a trend towards increased diagnostic accuracy in prediction of hyperplastic polyp histology during colonoscopy compared to endoscopist prediction in this interim analysis. The superior performance in sensitivity for disgnosis of hyperplastic polyps was also seen in diminutive polyps, but not with poor bowel preparation and difficult polyp location.


Disclosures:


Weiquan James Li, MBBS, MMed¹, Clement Wu, MBBS, MRCP², Christopher Jen Lock. Khor, MBBS², Raymond FH. Liang, MBBS, MRCP³, Jonathan Wei Jie Lee, MBBS, MRCP⁴, Jimmy So, MBChB, FRCSEd, FRCSG, FAMS, MPH⁴, Weida Chew, MBBS, MRCP³. B0105 - Real World Validation of an Artificial Intelligence Characterization Support System for Prediction of Polyp Histology in Colonoscopy: Interim Analysis of a Prospective Multicenter Study, ACG 2022 Annual Scientific Meeting Abstracts. Charlotte, NC: American College of Gastroenterology.