Colorectal cancer (CRC) is a leading cause of cancer-related deaths globally. Colonoscopy, a key method for CRC prevention, relies heavily on optimal bowel preparation. Inadequate bowel preparation, affecting 20-25% of patients, results from poor adherence to dietary restrictions and purgative instructions. Current educational approaches, including booklets, visual aids, and some apps, have limitations; they lack personalized feedback based on real-time assessment of bowel preparation status. This study proposes using AI to evaluate bowel preparation via fecal photographs, enabling personalized guidance through a smartphone app. The AI system assesses fecal images, and the app provides tailored reminders and instructions, addressing the limitations of existing methods.

Existing literature highlights the importance of adequate bowel preparation for successful colonoscopy and the challenges in achieving it. Several studies demonstrate that enhanced patient education improves bowel preparation quality. Various methods have been employed, including patient education booklets, cartoon visual aids, and nurse-delivered education. Social media platforms and smartphone apps have also shown promise. However, a limitation of these methods is the lack of personalization; educational content is often fixed and doesn't adjust to individual bowel preparation status. The ability to provide real-time feedback and personalized recommendations based on the individual's bowel preparation status is critical. AI has the potential to address this limitation by enabling real-time assessment of bowel preparation quality.

This study comprised two parts: developing an AI-based bowel preparation prediction system and conducting a clinical trial evaluating the system's efficacy through a smartphone app. The AI system was trained on a dataset of 5362 fecal photographs from 992 patients. The images were labeled as 'adequate' or 'inadequate' based on the Boston Bowel Preparation Scale (BBPS) scores from corresponding colonoscopies. The ShuffleNet V2 neural network was used due to its efficiency on smartphones. The developed AI system was integrated into a smartphone app ('Qing Chang') that provides personalized education and real-time bowel preparation assessments. A prospective, multicenter, single-masked, randomized controlled trial enrolled 524 smartphone-owning patients scheduled for colonoscopy. Patients were randomized 1:1 to either the control group (standard bowel preparation instructions) or the AI-driven app group. The primary outcome was the percentage of patients with adequate bowel preparation (BBPS score ≥6). Secondary outcomes included BBPS scores, compliance with instructions, polyp and adenoma detection rates, and various procedure-related parameters. Statistical analysis was performed using appropriate tests for continuous and categorical data.

The AI-based bowel preparation prediction system demonstrated high accuracy (95.15%), specificity (97.25%), and AUC (0.98) in the test dataset. In the full analysis set (n=500), the AI-driven app group showed a significantly higher rate of adequate bowel preparation (88.54% vs 65.59%, P<0.001) compared to the control group. The mean BBPS score was also significantly higher in the app group (6.74 ± 1.25 vs 5.97 ± 1.81, P<0.001). Compliance with dietary restrictions (93.68% vs 83.81%, P=0.001) and purgative instructions (96.05% vs 84.62%, P<0.001) were significantly better in the app group. The rate of additional purgative intake was also significantly higher in the app group (26.88% vs 17.41%, P=0.011). Subgroup analysis revealed that the app's benefit was consistent across most subgroups, with larger improvements observed in older adults and patients without prior colonoscopy experience. There were no significant differences in polyp or adenoma detection rates between the groups.

This study demonstrates that an AI-driven smartphone app significantly improves the quality of bowel preparation for colonoscopy by providing personalized, real-time feedback and enhanced patient education. The higher rate of adequate bowel preparation in the app group, coupled with improved compliance and increased additional purgative use, underscores the app's effectiveness. While no significant difference was found in polyp or adenoma detection rates, this may be due to the sample size. The app offers several advantages over traditional methods: real-time feedback, personalized recommendations, and increased patient engagement. The lightweight AI model ensures compatibility with most smartphones, and the app design protects patient privacy. The results suggest that this AI-powered approach could be a valuable tool in improving colonoscopy outcomes and patient experience.

This study successfully demonstrates that an AI-driven smartphone application can significantly improve the quality of bowel preparation for colonoscopy. The app's personalized feedback and enhanced education lead to higher rates of adequate preparation and better patient compliance. Future research could explore the app's efficacy in diverse populations and investigate the long-term impact on colonoscopy outcomes and CRC screening rates. Further development could integrate other factors influencing bowel preparation, such as patient comorbidities and medication use, for even more personalized recommendations.

This study's limitations include the inclusion of only smartphone users, potentially limiting generalizability. The sample size might have been insufficient to detect significant differences in polyp and adenoma detection rates. Future studies should address these limitations by including a broader population and increasing the sample size to investigate the effect on detection rates more thoroughly.