Design and Validation of an Artificial Intelligence System to Detect the Quality of Colon Cleansing Before Colonoscopy

The patient perception of the last bowel movement before the colonoscopy has been shown a powerful predictor of bowel cleansing rated during colonoscopy. A large study involving 1011 patients distributed in a derivation cohort (633 patients) and a validation cohort (378 patients) using a set of 4 pictures resembling bowel cleansing qualities showed a moderate agreement with the BBPS. In addition, a good agreement was found when the staff perception and patient perception of the last bowel movement were compared. These findings offer an excellent opportunity to test rescue cleansing interventions the same day of the examination, before colonoscopy.

Over the last two years, artificial intelligence applications have wrought a substantial breakthrough in several disciplines, including endoscopy. Machine learning and its more advanced form deep learning, refers to the development of algorithms (convolutional neural networks) with the ability to learn and perform certain tasks. In the endoscopy setting, computer vision applications have been stated as research priority field. Based on all this experience, the aim of this study was to design and to validate a convolutional neural network capable of automatically predicting the quality of the patient cleansing at home after the intake of the bowel cleansing solution and before attending the colonoscopy. The other aim was to prospectively assess in a cohort of patients the agreement between the result of the last rectal effluent quality assessed by the convolutional neural network and the cleansing quality assessed during the colonoscopy assessed by a validated scale (Boston Bowel Preparation Scale, BBPS) This study is nested in an observational prospective study conducted at the Open Access Endoscopy Unit of the Hospital Universitario de Canarias between February 2021 and May 2021 (NCT04702646). A total of 633 consecutive outpatients with a scheduled colonoscopy participated in this study (a total of 266 patients (42%) sent at least one picture). After this study, patients in whom an outpatient colonoscopy was requested, were asked to provide pictures of their effluents during bowel preparation intake. A subgroup of these images will be classified by the personal of our unit in adequate and inadequate and will be used to train the convolutional neural network. Another set of images will be used to validate the convolutional neural network. Additionally, the investigators will validate in-vivo the convolutional neural network comparing its classification of the effluent quality with a validated colon cleansing scale during the colonoscopy.