A Machine Learning-Based Clinical Tool for Predicting Inadequate Bowel Preparation: Development and Validation

Colonoscopy is an essential tool for diagnosing and treating colorectal diseases, with bowel preparation serving as a key determinant of its quality. This study aimed to develop and validate a machine learning model using nonpharmacological parameters to predict the risk of inadequate bowel preparation and develop a clinical risk assessment tool.

Clinical data were prospectively collected from patients who underwent colonoscopy with anesthesia at Beijing Tiantan Hospital from November 2024 to June 2025. Bowel preparation quality was evaluated using the Boston Bowel Preparation Scale. Three feature selection methods and 5 machine learning algorithms were used to develop models and evaluate their predictive performance.

Six factors were identified as predictive of inadequate bowel preparation. Cross-validation revealed that the Firth regression-based algorithm yielded an area under the operating characteristic curve (AUC) values of 0.718 (95% confidence interval: 0.647-0.789) for the training set and 0.715 (95% CI: 0.611-0.818) for the validation set, demonstrating balanced performance, calibration, and stability. The corresponding clinical prediction tool exhibited good discrimination [AUC, 0.709; 95% CI: 0.605-0.813] and calibration in the validation set, indicating its potential clinical value.

Our findings indicate that body mass index, waist-to-hip ratio, a composite score of lower gastrointestinal symptoms, history of diabetes, and a composite smoking and alcohol intake score are significant risk factors of inadequate bowel preparation, and hematochezia exhibited the opposite effect. Overall, the Firth regression-based model and its associated risk prediction tool demonstrated good performance in identifying patients with inadequate bowel preparation.