Improving bowel preparation for colonoscopy with a smartphone application driven by artificial intelligence

Colorectal cancer is highly prevalent and lethal worldwide, and colonoscopy is central to prevention through detection and removal of precursor lesions. The quality of bowel preparation critically affects colonoscopy performance, yet inadequate preparation occurs in about 20–25% of procedures, often due to poor adherence to instructions, dietary challenges, or intolerance of purgatives. Reinforced patient education using visual aids and reminders can improve preparation quality, but existing approaches generally deliver fixed, non-personalized content that does not reflect an individual’s real-time preparation status. Patients also struggle to self-assess adequacy from toilet appearances. Artificial intelligence, increasingly applied in endoscopy, could enable real-time evaluation of bowel preparation based on photographs of feces in the toilet, thereby driving personalized guidance. This study developed an AI-based prediction model and integrated it into a smartphone app to deliver individualized education and recommendations, and then evaluated its impact on bowel preparation quality in a multicenter randomized trial.

Prior work shows enhanced education improves bowel preparation quality and patient willingness to prepare adequately. Reinforcement methods include educational booklets, cartoon visual aids, nurse-delivered brochures, telephone re-education, social media messaging, and smartphone applications. While effective, these methods typically follow fixed schedules and cannot tailor content to a patient’s current bowel status, limiting personalization and real-time feedback. Visual references of clean versus dirty stool are helpful but may not cover all scenarios and can be difficult for older adults to use. AI has shown promise across gastrointestinal endoscopy tasks, suggesting potential for real-time, patient-specific evaluation of bowel cleanliness to guide adaptive instructions.

Study design: Two components were conducted: (1) development of an AI-based bowel preparation prediction system using toilet photographs; and (2) a prospective, multicenter, endoscopist-masked randomized controlled trial (RCT) to evaluate an AI-driven smartphone app integrating the model. The protocol was IRB-approved (Zhongshan Hospital B2020-297R) and registered (ChiCTR2000040360). AI model development: Patients scheduled for colonoscopy (Nov 2020–Jan 2021) using a 3-L split-dose PEG regimen were asked to photograph toilet contents after purgative doses. Colonoscopies were video-recorded and scored using the Boston Bowel Preparation Scale (BBPS). Adequate preparation was defined as total BBPS ≥6 with all segments ≥2. If a patient’s overall preparation was inadequate, all their photos were labeled “inadequate.” If adequate, only the first photo was labeled “adequate,” and intermediate photos were excluded due to uncertain status. A total of 5,362 photographs from 992 patients were randomly split at the patient level into training (~80%) and test (~20%) sets. A lightweight CNN (ShuffleNet V2) with data augmentation (random cropping, rotation) and early stopping was trained. Performance on the test set was evaluated by accuracy, sensitivity, specificity, and AUC. AI-driven app: The prediction system was integrated into the Qing Chang (v2.0) smartphone app (Henan Xuanyemed Medical Information Technology). App functions: (1) collect patient and scheduling information to coordinate preparation steps; (2) deliver personalized reinforced education (videos/articles vetted by senior endoscopists) before and during preparation; (3) evaluate bowel preparation via AI predictions on uploaded photos and provide personalized improvement suggestions, including whether to take additional PEG. Clinical trial: A multicenter, endoscopist-masked RCT (Sept 2021–Jan 2022) at four centers compared standard education (control) versus AI-driven app guidance. Eligibility: outpatients aged 18–75 scheduled for diagnostic colonoscopy, owning a compatible smartphone. Key exclusions included prior bowel surgery, severe GI motility disorders or obstruction, severe renal/cardiac disease, pregnancy/breastfeeding, toxic colitis/megalocolon, poorly controlled hypertension, significant GI bleeding, major psychiatric illness, allergy to thiosulfate purgatives, inability to use the app, or to consent. Randomization used center-stratified block methods (SAS 9.4) with opaque envelopes; endoscopists were masked to group allocation. Preparation protocol: All participants received standard education, a 3-day low-residue diet, and 4 L PEG (Heshuang). Dosing: 1 L at 20:00 the evening before (250 mL every 15 min), and 2 L 4–6 h pre-procedure. An additional 1 L PEG was reserved as a remedial measure for inadequate preparation. Control participants decided on additional PEG based on self-judgment with reference photos. App-group participants received app reminders and AI-based recommendations, including whether to take additional PEG. Colonoscopy procedures occurred 8:30–11:30 or 13:30–16:30 by experienced endoscopists (≥3,000 procedures). Procedures were video-recorded. Polypectomy/biopsy occurred during withdrawal. Outcomes: Primary outcome: proportion with adequate preparation (BBPS total ≥6 and all segments ≥2), adjudicated by two masked endoscopists with tie-break by a senior endoscopist. Secondary outcomes: total and segmental BBPS, excellent preparation (BBPS ≥8), compliance with diet and purgative instructions, cecal intubation time, withdrawal time, total procedure time, polyp detection rate (PDR), adenoma detection rate (ADR), advanced ADR (aADR), sleep quality during preparation, and willingness to repeat preparation. Analysis sets and statistics: The full analysis set (FAS) included all randomized patients who underwent colonoscopy (n=500). The per-protocol set (PPS) excluded app-group patients who did not use core app functions (insufficient photo uploads or content browsing). Safety set included all with safety assessment. Continuous variables were analyzed with t-tests; categorical variables with chi-square or Fisher’s exact test. Rates with 95% CIs were reported; alpha 0.05; secondary outcomes considered exploratory without multiple-testing correction. Sample size targeted detection of an increase in adequate preparation from 80% to 90% with 80% power, requiring ~394 analyzable patients; allowing for cancellations, 500 were planned.

Model performance: On the test dataset, the AI classifier achieved accuracy 95.15%, specificity 97.25%, and AUC 0.98. Participants: Of 578 screened, 524 were randomized (1:1). Twenty-four canceled, yielding 500 in the FAS (AI-app n=253; control n=247). Primary outcome: Adequate preparation was significantly higher with the AI-driven app (FAS): 88.54% vs 65.59%; P<0.001. Bowel cleanliness: Mean total BBPS was higher with the app: 6.74 ± 1.25 vs 5.97 ± 1.81; difference 0.77 (95% CI 0.49–1.04); P<0.001. Segmental BBPS improved for right colon (2.07 ± 0.65 vs 1.70 ± 0.84; diff 0.37; P<0.001), transverse (2.36 ± 0.57 vs 2.15 ± 0.72; diff 0.21; P<0.001), and left (2.30 ± 0.56 vs 2.12 ± 0.66; diff 0.18; P<0.001). Rates of excellent preparation were also higher in the app group (significant in FAS and PPS). Compliance and app-driven behaviors: Compliance with dietary restrictions was higher (93.68% vs 83.81%; P=0.001) and with purgative instructions (96.05% vs 84.62%; P<0.001). Additional purgative intake was more frequent (26.88% vs 17.41%; P=0.011). Among app-group patients who took additional purgatives, 95.59% achieved adequate and 27.94% excellent preparation. Colonoscopy performance: Cecal intubation time was shorter with the app (5.06 ± 2.07 vs 5.86 ± 2.85 min; diff -0.80; P<0.001). Withdrawal time and total procedure time were not significantly different. PDR (45.06% vs 39.68%; P=0.223) and ADR (27.67% vs 22.67%; P=0.198) were not significantly different. Safety: Adverse event rates did not differ significantly (28.06% vs 22.37%; P=0.136), with no between-group differences in abdominal pain, distension, or nausea/vomiting. Subgroups: Higher adequate-preparation rates with the app were observed across most subgroups, with smaller or nonsignificant differences among those scheduled for afternoon colonoscopy and among patients noncompliant with diet or purgative instructions. Older adults, those with lower educational levels, and patients without prior colonoscopy appeared to derive larger benefits.

The study addressed whether AI-enabled, real-time evaluation of bowel preparation, coupled with personalized guidance via a smartphone app, improves colonoscopy preparation quality. The app substantially increased the proportion of adequate preparations and improved overall and segmental BBPS scores. Mechanistically, benefits likely arose from tailored recommendations—particularly prompting additional purgative use when needed—and from enhanced adherence to diet and dosing schedules driven by timely reminders. Reduced cecal intubation time reflects the practical procedural advantage of better cleanliness. Although PDR and ADR were numerically higher in the app group, differences were not statistically significant, plausibly due to sample size and event rates. The approach demonstrates advantages over static educational tools by engaging patients in real time, improving self-awareness of cleanliness status, and preserving privacy while running an efficient, lightweight AI model on common smartphones. Subgroup patterns suggest the method is especially helpful for older adults, individuals with lower educational attainment, and first-time colonoscopy patients, underscoring its accessibility and potential to reduce disparities in preparation quality.

An AI-driven smartphone app that evaluates toilet photographs to assess bowel cleanliness and delivers personalized instructions significantly improves bowel preparation quality and patient compliance for outpatient colonoscopy. The intervention increased adequate and excellent preparation rates, enhanced adherence to dietary and purgative protocols, and shortened cecal intubation time without increasing adverse events. Future work should include larger, diverse populations to assess impacts on PDR/ADR, evaluate implementation in settings with lower smartphone penetration, extend support for non-smartphone users, and refine AI and user experience to further personalize dosing and timing recommendations.

Generalizability is limited to smartphone users able to operate the app; results may not apply to patients without regular smartphone access or digital literacy. The study was conducted in selected centers and outpatient settings, which may introduce selection bias. Secondary outcomes were exploratory without adjustment for multiple comparisons. The sample size may have been underpowered to detect differences in PDR/ADR. Afternoon scheduling and noncompliance subgroups showed attenuated benefits, indicating context-dependent effectiveness.