Development and external assessment of a risk prediction model for colorectal adenoma after endoscopic resection

BACKGROUND Colorectal adenoma is the most important precancerous lesion of colorectal cancer.However, large-scale evidence supporting risk-stratified surveillance after endoscopic resection remains limited in China.AIM To identify independent risk factors for adenoma recurrence and develop a 3-year recurrence prediction model using Cox proportional hazards regression.Two external cohorts with shorter follow-up were used to assess model discrimination at intermediate follow-up time points; these do not constitute formal external validation of the 3-year prediction horizon.METHODS A total of 5161 patients who underwent at least two colonoscopies were enrolled from the Affiliated Hospital of Chengdu University and constituted the development cohort (median follow-up 3.2 years).Cox proportional hazards regression was used as the primary analysis with a prespecified 3-year prediction horizon.Internal validation was performed using ten-fold cross-validation.Two independent external cohorts from Yibin (n = 1119, median follow-up 13.1 months) and Zigong (n = 4673, median follow-up 14.1 months) were evaluated only at their respective intermediate follow-up time points.A secondary, exploratory multivariable logistic regression was performed to derive a simplified points-based risk-scoring chart; however, this scoring tool is subject to surveillance-time bias because follow-up durations differed substantially across cohorts, and therefore it should not be used for clinical bedside risk stratification without further methodologic refinement.The study followed a predefined statistical analysis plan and was reported in accordance with the TRIPOD guidelines.4 / 44 RESULTS In the multivariable Cox regression, older age, male sex, polyp size ≥ 1 cm, ≥ 3 polyps, high-risk adenoma, and non-adenomatous polyps were identified as independent risk factors for recurrence (all P < 0.05).The Cox model demonstrated moderate discriminatory ability, with a time-dependent area under the receiver operating characteristic curve (AUC) of 0.789 (95% confidence interval: 0.772-0.806)at 3 years in the development cohort and 0.756 (95% confidence interval: 0.728-0.781) in internal validation.The exploratory logistic-derived risk score stratified patients into low-, intermediate-, and high-risk groups with recurrence rates of 26.4%, 47.9%, and 62.5%, respectively.In the two external cohorts, discrimination assessed at intermediate followup (13.1 and 14.1 months) yielded AUCs of 0.661 and 0.639; these values cannot be directly compared with the 3-year development-cohort AUC because they correspond to different prediction time points.Formal external validation at the 3-year horizon could not be performed due to insufficient long-term follow-up in the external cohorts.CONCLUSION The 3-year Cox model demonstrated moderate discrimination in internal validation.External cohorts only permitted evaluation at intermediate follow-up and do not constitute formal external validation for the 3-year prediction horizon.The logisticderived risk score is exploratory and hampered by surveillance-time bias; prospective studies with uniform long-term follow-up are needed for true external validation and recalibration before any clinical application.This model can serve only as a preliminary research-oriented risk-triage aid.