Insights into a machine learning-based prediction model for colorectal polyp recurrence after endoscopic mucosal resection

Abstract

In 2025, Shi et al constructed a model utilizing machine learning techniques to predict the one-year recurrence of colorectal polyps following endoscopic mucosal resection, showing excellent discriminatory performance with an area under the curve exceeding 0.90. However, limitations exist regarding its narrow temporal scope, potential overestimation due to feature collinearity and imputation opacity, and limited generalizability due to single-center derivation and validation. Moreover, no clear clinical implementation strategy was outlined. Prospective multicenter validation and integration of endoscopist variability, longitudinal outcome data, and deployment mechanisms are warranted to ensure broader applicability and clinical utility.

Key Words: Colorectal polyp recurrence; Endoscopic mucosal resection; Machine learning; Risk prediction; Clinical implementation; External validation

Core Tip: This letter provides a critical appraisal of a recent machine learning model designed to predict colorectal polyp recurrence after endoscopic mucosal resection. It highlights key methodological issues, such as endpoint selection, imputation transparency, and external validation, while offering constructive recommendations to enhance clinical applicability and alignment with international surveillance guidelines.

TO THE EDITOR

We read with interest the study by Shi et al[1], who introduced and assessed a machine learning (ML) model aimed at predicting one-year recurrence of colorectal polyps following endoscopic mucosal resection (EMR). While the model is an encouraging step toward precision surveillance, several important limitations and enhancement opportunities should be discussed to improve both methodological robustness and clinical relevance.

First, the study’s endpoint - one-year recurrence - is somewhat narrow in scope. International guidelines typically recommend surveillance intervals of 3 years for adenoma follow-up unless high-risk features are present[2,3]. This aligns with global practice patterns emphasizing risk-based stratification. While the authors note that recurrence risk peaks within one year in their Chinese cohort, a broader focus on multi-year outcomes (e.g., 3-year recurrence) would strengthen clinical applicability and harmonize the model’s scope with international standards[2,3].

Second, although the model's performance metrics, particularly the area under the curve of the XGBoost model (> 0.90), are impressive, they may be over-optimistic due to possible data leakage and unaddressed confounders[4,5]. The use of SHapley Additive exPlanation to interpret feature importance is commendable[6], but the authors do not discuss collinearity or redundancy across included features. Notably, some predictors (e.g., polyp number and size) are closely interrelated and may have inflated apparent model performance. Moreover, handling of missing data using multiple imputation warrants more transparency, especially for biomarkers with > 15% missingness. Proper documentation of imputation methods and sensitivity analyses would enhance the study’s reproducibility and credibility. As highlighted in recent methodological reviews, robust imputation techniques and transparent reporting are essential to reduce bias and maintain model validity in clinical prediction research[4,5].

Third, the generalizability of the model remains limited. All data were derived from a single geographic region in China, and external validation was conducted only on a modest sample from the same institution[7]. The absence of international, multi-center validation raises concerns about overfitting to local practice patterns. Additionally, the model does not account for endoscopist-related variability, which is a known determinant of recurrence[7].

Lastly, although the authors developed an online calculator based on the XGBoost model, the clinical implementation pathway remains vague. How this tool would integrate with existing surveillance guidelines and electronic health records[8,9], or whether clinicians would realistically adopt such a model, is not addressed. A decision impact analysis or user-acceptability study would be useful next steps[10].

In summary, the authors present an important advancement in post-polypectomy risk stratification using ML. However, additional methodological transparency, broader validation, and implementation planning are essential for translating this promising tool into clinical practice.

CONCLUSION

While the ML model developed by Shi et al[1] represents a promising tool for predicting colorectal polyp recurrence, its current limitations - such as narrow endpoint selection, lack of external validation, and incomplete methodological transparency - highlight the need for further refinement. Future prospective, multi-center studies and the incorporation of real-world clinical variables will be critical for improving the model’s robustness, generalizability, and clinical integration.