DEVELOPMENT AND INITIAL VALIDATION OF AN INTERACTIVE ARTIFICIAL INTELLIGENCE ASSESSMENT OF COLONOSCOPY QUALITY

Background and aim: Despite extensive cleaning and disinfection protocols, 5-15% of ready-to-use duodenoscopes are still contaminated with high concern (HC) microorganisms such as Klebsiella pneumoniae, Escherichia coli and Pseudomonas aeruginosa. Changes in cleaning protocols might reduce the contamination rate and hereby prevent future outbreaks. This study aims to investigate the effect of a new endoscope cleaning brush on contamination of duodenoscopes with HC microorganisms.
Methods: In this retrospective observational study, the results of the duodenoscope surveillance cultures between March 2018 and June 2022 were collected. Contamination with HC microorganisms was defined as ≥1 colony-forming units of gastrointestinal microorganisms including P. aeruginosa or Staphylococcus aureus. Cultures of quarantined duodenoscopes were not included. From December 2020, a new endoscope cleaning brush with an additional wiper (Endoss Push and Pull brush, JPP50) was introduced for the manual pre-cleaning of the Pentax ED34-i10T2 duodenoscopes instead of the Pentax Single Use Brush (CS5522A). Using a generalized mixed effect model, the effect of the introduction of the new endoscope cleaning brush on Pentax ED34-i10T2 duodenoscope contamination with HC microorganisms was assessed. Other covariates were frequency of use, repairs by the manufacturer and the result of the prior culture.
Results: Data from 195 cultures of eight duodenoscopes was collected; 122 cultures prior to the introduction of the new brush and 73 after introduction. Prior to the introduction, 51/122 (41.8%) cultures were positive with HC microorganisms, especially with P. aeruginosa (29/122, 23.8%). After introduction of the brush only 7/73 (9.6%) were positive with HC microorganisms of which only one was positive with P. aeruginosa. Duodenoscopes cleaned with the new brush had a significantly lower odds of contamination with HC microorganisms compared to duodenoscopes cleaned with the old brush (OR = 0.17, 95 % CI [0.07, 0.42], p <0.001).
Conclusions: The Endoss Push and Pull brush significantly reduced contamination of ED34-i10T2 duodenoscopes, and it is therefore promising in the prevention of healthcare-associated infections and outbreaks. In the future, these results should be confirmed in a prospective study with different types of duodenoscopes.

INTRODUCTION:
Improvements in colonoscopy quality are associated with reductions in interval colorectal cancer death. However, measurement of colonoscopy quality in practice remains challenging. We aim to describe and validate an interactive tool for artificial intelligence (AI) assessment of colonoscopy quality (AI-CQ) using recorded videos.

METHODS:
Based on quality guidelines, metrics were selected for AI development: insertion time (IT), withdrawal time (WT), retroflexion frequency, polyp detection rate (PDR), polyps per colonoscopy (PPC), and number of right colon evaluations. We also developed two novel metrics: withdrawal time excluding polypectomy time (WT-PT) and high-quality withdrawal time (HQ-WT; withdrawal time with clear colon image). The AI model was pre-trained using a self-supervised vision transformer on unlabeled colonoscopy images (n = 1x10⁷) mutually exclusive from all other datasets. The vision transformer model was finetuned for multi-label classification on another mutually exclusive colonoscopy image dataset (n = 9854) using anatomical, procedural, and pathological labels (label n = 14). During inference, colonoscopy video frame predictions were generated at a resolution of one frame per second and employed a binary threshold of ≥ 0.5 to denote presence; these predictions were subsequently used to calculate all metrics. A timeline of video predictions and metric calculations were presented to clinicians in addition to the raw video using a web-based application. The model was developed using videos at a single hospital and externally validated using 50 screening and surveillance colonoscopy videos from 6 colonoscopists at a second hospital.

RESULTS:
The interactive AI-CQ tool is presented (Figure). The cecum was reached in 48/50 cases; AI-CQ accuracy to identify cecal intubation was 88%. In 6 cases, AI-CQ did not identify the cecum due to inadequate bowel preparation obscuring landmarks (n=4) and failure to recognize landmarks (n=2).

IT (p = 0.26) and WT (p = 0.34) were similar between manual and AI-CQ measurements and significantly (p < 0.001) positively correlated (Table). On average, HQ-WT was 45.9% (IQR: 14) of, and significantly correlated with (ρ = 0.85; p < 0.001), normal WT time. Mean WT-PT was 567s, similar to mean normal colonoscopy WT (558s).

AI-CQ produced similar PDR (p = 0.66) and PPC compared to manual (p = 0.34). Rectal retroflexion was correctly identified in 95.2% of colonoscopies and the number of right colon evaluations in all colonoscopies.

DISCUSSION:
AI-CQ can be utilized to rapidly measure quality and facilitates AI-augmented review of inspection and polypectomy technique to provide actionable feedback. Further, novel inspection metrics such as WT-PT and HQ-WT, which can only be feasibly calculated by AI, may prove beneficial but require further study.