Preoperative risk stratification for pathological upgrading in colorectal polyps using explainable machine learning: implications for screening optimization and resource allocation.

Background: Pathological upgrading in colorectal polyps, which occurs when resection specimens have a higher histological grade than preoperative biopsies, can lead to an underestimation of the severity of the disease and suboptimal treatment. Given the substantial global burden of colorectal cancer (CRC), improving preoperative risk stratification of colorectal polyps is essential for optimizing screening strategies and resource allocation in CRC prevention. Methods: Between December 2019 and December 2024, 593 patients with colorectal polyps were included in this retrospective study, undergoing endoscopic biopsy and subsequent complete resection. Clinical and endoscopic variables were gathered, and feature selection was conducted using LASSO regression and the Boruta algorithm. Repeated ten-fold cross-validation was used to develop and evaluate six machine learning models: RF, CART, NNet, LR, GBM, and XGBoost. AUC, accuracy, sensitivity, specificity, MCC, and Brier score were used to assess the model's performance. A web-based calculator was developed to aid in clinical implementation, using SHapley Additive exPlanations (SHAP) to interpret the optimal model. Results: Among the 593 patients, 150 (25.3%) experienced pathological upgrading. Five key predictors were identified: maximum tumor diameter, surface color, erosion, villous structure, and lesion location. With an AUC of 0.890 in the training set and 0.863 in the test set, the XGBoost model exhibited superior performance, along with strong calibration and discrimination. SHAP analysis showed that lesion location, particularly rectal location, was the most influential factor, followed by erosion and tumor size. To aid in individual risk prediction, a user-friendly online calculator was designed (available at: https://changchangzhang2001.shinyapps.io/blsj/). The calculator will remain freely accessible for at least 3 years after publication, and the source code is available from the corresponding author upon reasonable request. Conclusion: Researchers in this study devised a machine learning model that provides explanations for predicting pathological upgrading in colorectal polyps. The model serves as a useful instrument for assessing preoperative risks and aids in making decisions based on risk, potentially boosting early detection efficiency, refining endoscopic management approaches, and improving resource allocation in colorectal cancer prevention programs.