Precision prevention of colorectal neoplasia in patients undergoing hemorrhoid surgery: an explainable machine learning model for identifying the risk of precancerous neoplastic polyps.

Background: Hemorrhoids are common and frequently present with hematochezia. However, these symptoms can overlap with manifestations of precancerous neoplastic polyps (PNP) and early colorectal cancer. Although colonoscopy is recommended for patients with hematochezia, it is often not completed in routine practice for various reasons, which may lead to missed diagnoses in some patients. We built and validated a machine learning-based model to identify patients undergoing surgery for severe hemorrhoids who are at elevated risk of PNP, with the aim of supporting clinical decision-making and prevent colorectal neoplasms. Methodology: We enrolled 589 patients who underwent hemorrhoid surgery at Wujin People's Hospital, between January 2021 and January 2025 and completed colonoscopy within 1 year before or after surgery. Demographic and clinical data recorded at initial admission comprised sex, age, body mass index (BMI), smoking, drinking, diabetes, and hypertension. Feature selection was conducted using least absolute shrinkage and selection operator (LASSO) regression, after which the dataset was randomly partitioned into training and testing cohorts at an 8:2 ratio. Our study constructed Seven machine learning models. The area under the receiver operating characteristic curve (AUROC) and standard classification metrics were used to assess discriminative performance; calibration was assessed with the Brier score. Shapley additive explanations (SHAP) -based attribution analysis was performed for the best-performing model. Finally, decision curve analysis (DCA) and risk stratification were used to assess clinical utility. Results: LASSO identified age, sex, BMI, smoking, drinking, and hypertension as key predictors. Among the evaluated models, RF achieved the highest training-set AUROC (0.892). In the testing set, the AUROC was 0.738 with a Brier score of 0.172, suggesting acceptable calibration and reasonable overall stability. Risk stratification according to tertiles of predicted probability demonstrated a distinct gradient in the prevalence of PNP, increasing from 7.5% in the low-risk group to 23.1% in the intermediate-risk group and 53.8% in the high-risk group. SHAP analysis showed that age contributed most to the predictions, followed by drinking, BMI, smoking, sex, and hypertension. Conclusion: This model may help clinicians more accurately identify high-risk individuals among patients undergoing hemorrhoid surgery, potentially reducing missed PNP and strengthening colorectal cancer prevention.