Machine Learning-Based Investigation of Factors Influencing Recurrence of Colorectal Adenomatous Polyps: A...

Colorectal adenomatous polyps are well-recognized precancerous lesions of colorectal cancer. Patients who undergo endoscopic polypectomy still face a high risk of polyp recurrence, and individualized risk stratification remains challenging in routine clinical practice. Dyslipidemia has been implicated in adenoma development, but its role in recurrence and the predictive value of machine learning tools are understudied in Chinese populations. This study aimed to identify independent risk factors for adenoma recurrence and compare the performance of eight machine learning prediction models.

This single-center retrospective cohort study included 769 patients who underwent colonoscopic polypectomy and completed at least one surveillance colonoscopy. A non-random site-based split was used to derive a training cohort (n = 539, Endoscopy Center) and an independent internal test cohort (n = 230, Colorectal Center). Univariate and multivariate Cox proportional hazards regression were applied to identify independent predictors of recurrence. Eight machine learning models were constructed using the selected predictors, and their discriminative performance, calibration, and clinical net benefit were comprehensively evaluated.

Abnormal high-density lipoprotein cholesterol (HDL-C), higher baseline polyp count, and larger total polyp volume were independent risk factors for adenoma recurrence. In the training set, the gradient boosting machine (GBM) achieved the highest AUC of 0.874, followed by XGBoost (AUC = 0.866); in the test set, GBM and XGBoost maintained favorable discriminative performance with AUCs of 0.863 and 0.849, respectively. The two models delivered comparable clinical net benefit across clinically relevant probability thresholds. XGBoost demonstrated acceptable calibration (Hosmer-Lemeshow p = 0.065), whereas GBM showed statistically significant miscalibration (p = 0.043).

Abnormal HDL-C and greater baseline polyp burden are independent predictors of earlier colorectal adenoma recurrence. Machine learning models, particularly gradient boosting algorithms, achieve favorable discriminative performance and may serve as complementary tools for post-polypectomy risk stratification, though further calibration optimization is warranted prior to clinical application.