Development and Validation of a Colorectal Cancer Diagnostic Model

Colorectal cancer imposes a heavy disease burden in China. Early detection of colorectal cancer and advanced precancerous lesions can improve clinical outcomes and reduce colorectal cancer mortality. However, colonoscopy, the standard method for screening, is invasive and resource-intensive. A practical, accurate, and clinically applicable risk prediction model may help identify individuals with a higher probability of prevalent colorectal cancer or advanced colorectal neoplasia, support risk-adapted early detection strategies, and optimize resource utilization. Yet existing diagnostic and risk prediction tools have been developed primarily in Western populations and demonstrate limited applicability to Chinese individuals.

This study is designed as a multicenter, observational, and cross-sectional study enrolling adults aged 18-80 years who are already scheduled to undergo colonoscopy during routine clinical care at participating hospitals. The study does not assign participants to any medical intervention and does not determine whether they should undergo colonoscopy. All decisions regarding the colonoscopy, including the timing, conduct of the procedure, pathology evaluation, and subsequent clinical care, are made by the treating clinicians following the routine clinical pathway of the respective participating institution.

After providing informed consent, eligible participants will complete a standardized electronic questionnaire before colonoscopy. The questionnaire will collect candidate predictors, including demographics, lifestyle and behavioral factors (e.g., smoking, alcohol consumption, physical activity, diet), personal and family medical history, and gastrointestinal symptoms. After the index clinical procedure, clinical colonoscopy and pathology reports will be abstracted. The main outcome is prevalent colorectal advanced neoplasia at the index colonoscopy, including colorectal carcinoma, advanced adenoma (adenoma ≥10 mm, or with high-grade dysplasia, or with villous/tubulovillous histology), and advanced serrated adenoma (traditional serrated adenoma or sessile serrated adenoma ≥10 mm or with high-grade dysplasia).

At least 20,000 participants will be enrolled across multiple sites, including Peking University Cancer Hospital \& Institute, the First Affiliated Hospital of Xiamen University, Fujian Provincial Hospital, Ningxia Hui Autonomous Region People's Hospital, and other collaborating institutions. Participants are enrolled competitively until the total sample size is reached.

The sample size was determined to support the development and validation of a multivariable risk prediction model, following the framework proposed by Riley et al. We adopted a conservative set of assumptions, with an expected prevalence of advanced colorectal neoplasia of approximately 6%, an anticipated model discrimination (AUC) in the range of 0.65 to 0.70, and approximately 40 candidate predictor parameters covering demographics, lifestyle factors, personal and family medical history, prior colonoscopy history, and gastrointestinal symptoms. The minimum sample size required for model development is estimated at 9,000 to 13,500 participants. After reserving 20% to 30% of enrolled participants for internal validation, the total required sample size increases to approximately 11,000 to 19,000. To further accommodate potential incomplete data, non-evaluable outcomes, and multi-center external validation, the final enrollment target was set at a minimum of 20,000 participants.

Candidate prediction models will be developed using logistic regression and machine-learning algorithms. Internal validation will be performed using cross-validation and bootstrap methods. External validation will be performed across participating centers. Model performance will be assessed using the area under the receiver operating characteristic curve, calibration measures, and decision-curve analysis. The final model and risk stratification scheme will be selected based on predictive performance, stability, interpretability, feasibility, and data-collection burden. This scheme will then be translated into a practical tool for clinical and public health settings to guide individualized colorectal cancer screening strategies.