Development and validation of the Open-Source Automatic Bowel Preparation Scale.

Abstract

Background and aims: Insufficient bowel preparation accounts for up to 42% of missed adenomas in colonoscopy. However, major analysis programs found no correlation between adenoma detection rate and the human-rated Boston Bowel Preparation Scale (BBPS), indicating limitations of the scale. We therefore aimed to develop an open-source automatic bowel preparation scale (OSABPS) based on artificial intelligence that is correlated to the polyp detection rate (PDR).

Methods: OSABPS was trained on 50,000 frames from 20 colonoscopies from three hospitals. It involved quantifying the presence of fecal matter within the colonoscopy frames, using an approach termed the fecal ratio - the proportion of pixels identified as feces (F) relative to those identified as mucosal tissue (M) (OSABPS = F/M) - thereby making 0 the optimal score indicating a perfect cleansing. Youden's J was used to set the threshold, as it determines the optimal balance between sensitivity and specificity. The algorithm was then tested on 1,405 colonoscopies from three hospitals (internal validation), and 5,525 frames from a public colonoscopy database (Nerthus, external validation).

Results: Internal validation: OSABPS correlated significantly with BBPS (Pearson's r = -.42, P<.001). A threshold of .09 OSABPS was determined using Youden's J. PDR was higher for colonoscopies below the threshold of Youden's J (Two proportion z-test, P<.001). External validation: OSABPS correlated significantly with BBPS (Pearson's r = -.70, P<.001).

Conclusions: OSABPS can automatically, instantly and without human bias assess bowel preparation quality. Colonoscopies with an OSABPS > .09 should be considered for reexamination. OSABPS' open-source nature allows free implementation.