Automatic Evaluation of Small Bowel Cleanliness in Capsule Endoscopy: A Comparative Study of Four Preparations using Artificial Intelligence

Abstract Background Diagnostic yield (DY) of small bowel capsule endoscopy (SBCE) is limited by the presence of intestinal content. An objective assessment of the cleanliness grade of the small bowel (SB) is currently lacking. We developed and validated an artificial intelligence (AI) tool to assess SB cleanliness in a multicenter observational study comparing four types of bowel preparations. Objectives Identify the best preparation for SBCE applying a system based on convolutional neural networks that assess the SB degree of cleanliness. Design We included 372 SBCE consecutive procedures with four different bowel preparations. Group A (n = 93): liquid diet (LD), Group B (n = 92): macrogol 3350 + ascorbic acid 1000 mL, Group C (n = 94): sodium picosulfate, and Group D (n = 93): macrogol 500 mL. The degree of cleanliness of the SB was evaluated with the automatic tool (0–100%). Gastric and SB transit times, DY, and degree of satisfaction with the preparation were also evaluated. Results Mean age of included patients was 58.14 years (confidence interval 95% 18–93 years); 55.9% were women. There were no significant differences in the degree of cleanliness of the SB evaluated with AI between groups (Mean Group A: 51.28%; B: 54.79%; C: 52.44%; D: 55.34%; p=0.415), nor in transit times or DY. LD was the best-tolerated preparation. Multivariable analysis identified age and diabetes as independent predictors of lower cleanliness. Conclusion We demonstrated using an AI-based tool, that LD provides adequate preparation for SBCE, questioning the routine administration of laxatives and supporting a patient-centered approach.