ABSTRACT Background Adequate bowel preparation is crucial for high‐quality colonoscopy; however, assessing preparation adequacy can be burdensome for both healthcare providers and patients. In this study, we aimed to develop artificial intelligence (AI) models for the automated identification of bowel PREParation for cOlonoscOpy (AI‐PREPOO). Methods On the day of colonoscopy, participants were instructed to use smartphones to photograph their stool in the toilet after each bowel movement following initiation of polyethylene glycol solution and upload the images to a secure web server. All images were labeled as “ready” or “not ready” for colonoscopy based on clarity and the absence of solid content. Using these labeled images, four image‐recognition models based on different deep learning architectures (AI‐PREPOO 1–4) were developed using transfer learning to classify stool status as “ready” or “not ready.” Results A total of 282 stool images were collected from 37 patients, with 141 images labeled as “ready” and 141 as “not ready.” These images were divided into training (224 images) and test (58 images) sets, and the training set was augmented to 2240 images. All models trained on the augmented dataset achieved high performance, with area under the curve (AUC) values exceeding 0.90. AI‐PREPOO 1, based on MobileNetV3‐Small, demonstrated the most balanced sensitivity–specificity profile (AUC, 0.95; sensitivity, 0.93; specificity, 0.86). Conclusions We developed AI‐based models capable of accurately assessing bowel preparation adequacy. AI‐PREPOO 1 showed a well‐balanced diagnostic performance, suggesting its potential to facilitate bowel preparation assessment and reduce the burden on healthcare providers and patients.
Kojima et al. (Mon,) studied this question.