AI-assisted assessment of bowel preparation from patient-generated images for pre-procedure triage.

Inadequate bowel preparation before colonoscopy or colorectal surgery can delay procedures, require repeat preparation, and increase clinical workload. Current assessment depends on manual review of patient-reported rectal effluent images by nursing staff, which is time-consuming and subject to inter-observer variation. To develop and explore the feasibility of an AI-assisted pre-procedure triage model that classifies bowel-preparation adequacy from patient-generated toilet images to support pre-colonoscopy assessment. A total of 1,508 patient-generated toilet images were retrospectively collected from colorectal surgery patients at a tertiary hospital (Aug 2019–Feb 2023). To avoid within-patient correlation, we retained one image per patient, excluding 490 images from patients with multiple submissions, resulting in 1,018 images from 1,018 unique patients. Three raters (resident, faculty surgeon, graduate student) independently labeled each image; disagreements were resolved by a two-of-three consensus rule. A DenseNet-201 model incorporating a Feature Pyramid Network (FPN) for multi-scale feature representation was trained to classify bowel preparation as clean or not clean. Model performance was assessed using stratified five-fold cross-validation on a development set (n = 778) and a held-out test set drawn from high-confidence, three-rater consensus cases (n = 240), using clinically relevant metrics including AUROC, F1-score, sensitivity, and specificity. The model achieved a mean AUROC of 0.886, F1-score of 0.882, and sensitivity of 0.931 for detecting inadequate preparation (10 repeated five-fold cross-validation; development set, n = 778). Grad-CAM visualization showed attention focused on residual stool and water turbidity, consistent with clinician interpretation. Ablation analysis indicated incremental improvements from random oversampling, smoothing weight decay, and FPN integration. The AI-assisted bowel-preparation triage model showed robust discrimination of inadequate preparation using real-world patient images. This approach has the potential to streamline manual review and enhance efficiency in pre-procedure assessment. Further studies should validate the model prospectively and evaluate its integration into clinical workflow.