Scholarlink

Background <p>Endoscopic mucosal resection (EMR) is a minimally invasive treatment for early colorectal lesions. However, post-EMR clinically significant delayed bleeding (CSPEB) is a common complication affecting patient outcomes. Accurate risk prediction is essential for optimizing management and reducing complications.</p> Methods <p>We conducted a retrospective study of 3888 patients who underwent colorectal EMR at Sun Yat-sen University’s Sixth Affiliated Hospital from January 2018 to September 2024. External validation was performed using data from 1000 patients at Shenzhen Hospital of Southern Medical University (2022–2024). CSPEB was defined as postoperative lower gastrointestinal bleeding requiring endoscopic intervention within 30&#xa0;days. Risk factors were identified using logistic regression. A random forest model with weighted sampling was developed and evaluated using ROC curves, calibration plots, and decision curve analysis (DCA). The model was implemented in a web-based application.</p> Results <p>CSPEB occurred in 1.4% of patients. Six independent risk factors were identified: sigmoid location, lesion size, number of lesions, APTT, fibrinogen, and hemoclip count. The model achieved an AUC of 0.87 ± 0.04 (sensitivity 71%, specificity 86%). External validation showed an AUC of 0.80 (sensitivity 61.5%, specificity 89.0%). SHAP methods enhanced model interpretability. Calibration and DCA confirmed strong predictive performance and clinical utility.</p> Conclusion <p>We developed and validated the first machine learning model for predicting post-EMR bleeding in the Chinese population. The model provides accurate, interpretable risk predictions and has been integrated into a user-friendly web tool to support clinical decision-making and improve patient outcomes.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Constructing and validating a risk prediction model for postoperative bleeding after colorectal EMR in the Chinese population: a machine learning-based study

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Bingfeng He,
  • Jiawei Zhang,
  • Mingli Su,
  • Wen Xu,
  • Runhua Li,
  • Dezheng Lin,
  • Juan Li,
  • Jiaxin Deng,
  • Yongcheng Chen,
  • Han Wang,
  • Junhao Wang,
  • Yuying Wang,
  • Ying Zhu,
  • Qinghua Zhong,
  • Xuefeng Guo

摘要

Background

Endoscopic mucosal resection (EMR) is a minimally invasive treatment for early colorectal lesions. However, post-EMR clinically significant delayed bleeding (CSPEB) is a common complication affecting patient outcomes. Accurate risk prediction is essential for optimizing management and reducing complications.

Methods

We conducted a retrospective study of 3888 patients who underwent colorectal EMR at Sun Yat-sen University’s Sixth Affiliated Hospital from January 2018 to September 2024. External validation was performed using data from 1000 patients at Shenzhen Hospital of Southern Medical University (2022–2024). CSPEB was defined as postoperative lower gastrointestinal bleeding requiring endoscopic intervention within 30 days. Risk factors were identified using logistic regression. A random forest model with weighted sampling was developed and evaluated using ROC curves, calibration plots, and decision curve analysis (DCA). The model was implemented in a web-based application.

Results

CSPEB occurred in 1.4% of patients. Six independent risk factors were identified: sigmoid location, lesion size, number of lesions, APTT, fibrinogen, and hemoclip count. The model achieved an AUC of 0.87 ± 0.04 (sensitivity 71%, specificity 86%). External validation showed an AUC of 0.80 (sensitivity 61.5%, specificity 89.0%). SHAP methods enhanced model interpretability. Calibration and DCA confirmed strong predictive performance and clinical utility.

Conclusion

We developed and validated the first machine learning model for predicting post-EMR bleeding in the Chinese population. The model provides accurate, interpretable risk predictions and has been integrated into a user-friendly web tool to support clinical decision-making and improve patient outcomes.

Graphical abstract