Machine learning-based prediction model in anastomotic leakage in rectal cancer patients: a multicenter, prospective cohort study

Monday, May 11, 2026

4:11 PM - 4:18 PM ET

Abstract Podium Presenter(s)

Leqi Zhou

MD, PhD
Department of Colorectal Surgery, Shanghai Changhai Hospital, Naval Medical University.
Shanghai Changhai Hosputal
Shanghai, Shanghai, China (People's Republic)

All disclosed relationships have been reviewed and managed prior to the start of this activity. Any relevant relationships will be listed below. This individual has no financial relationships to disclose

Author(s)

GY

Guanyu Yu, n/a

Associate professor, MD, PhD
Department of Colorectal Surgery, Shanghai Changhai Hospital, Naval Medical University.
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.
YW

Yingjie Wu, n/a

MD
Department of colorectal surgery, Shanghai Changhai Hospital, Naval Medical University
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.
LZ

Lu Zhou, n/a

MD
Department of colorectal surgery, Shanghai Changhai Hospital, Naval Medical University
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.
XZ

Xiaoming Zhu, n/a

MD
Department of colorectal surgery, Shanghai Changhai Hospital, Naval Medical University
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.
TZ

Tianshuai Zhang, n/a

MD
Department of Colorectal Surgery, Shanghai Changhai Hospital, Naval Medical University.
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.

Author(s)

RW

Rongbo Wen, n/a

MD, PhD
Department of Colorectal Surgery, Shanghai Changhai Hospital, Naval Medical University.
Shanghai, Shanghai, China (People's Republic)

All disclosed relationships have been reviewed and managed prior to the start of this activity. Any relevant relationships will be listed below. This individual has no financial relationships to disclose

Author(s)

HG

Haifeng Gong, n/a

MD
Department of Colorectal Surgery, Shanghai Changhai Hospital, Naval Medical University.
Shanghai, Shanghai, China (People's Republic)

Disclosure information not submitted.

Senior Author(s)

WZ

Wei Zhang

Changhai Hospital

Disclosure information not submitted.

Purpose/Background: Colorectal cancer is the third most common malignant tumor in the world, but surgical resection is still the most important means of colorectal cancer treatment. As one of the most serious complications of rectal cancer after surgery, anastomotic leakage can significantly increase postoperative mortality and the risk of recurrence, and seriously affect the prognosis of patients. Relying on the advantages of machine learning algorithms, this study can provide a new way to optimize the prediction of the occurrence of postoperative anastomotic leakage, thereby improving clinical decision-making and improving public health economics.

Methods/Interventions: In this study, 536 patients who underwent low anterior resection or conformal sphincter-preservation operation for rectal cancer were prospectively included from June 2023 to January 2025, and 536 patients were retrospectively collected from January 2021 to May 2023, all of these patients were from four top-tier hospital in China. Univariate analysis and multivariate logistic regression analysis were used to screen the risk factors associated with postoperative anastomotic leakage of rectal cancer, and the predictors were screened by LASSO regression. The data is divided into prediction set and training set, modeled by ten machine learning algorithms, the optimal model is evaluated by parameters such as receiver operating characteristic curve and area under the curve. The SHAP analysis is applied to explain the model.

Results/Outcomes: The results of univariate analysis identified 14 risk factors for postoperative anastomotic leakage and the multivariate logistic regression showed that five factors were independent risk factors for postoperative anastomotic leakage (p < 0.05). The presence of a high correlation between modeling variables may make the model unstable due to multicollinearity, so we conducted a multicollinearity analysis of the 14 included factors before modeling, and the results showed that there was a certain correlation between some factors (Figure A). Through LASSO regression, eight variables were screened out as risk factors (Figure B and Figure C). In this study, 536 rectal cancer cases were prospectively included, with a probability of anastomotic leakage of 12.7% (68/536), and the training set and test set were divided into 7:3 ratio. Modeling found that the random forest (RF) model and LightGBM had the best performance in the training set (Figure D). The internal verification found that the sensitivity and the AUC of RF were better than LightGBM (Figure E). The SHAP swarm plot and SHAP waterfall chart explain how the model get the result. (Figure F and G). The external verification found that the prediction performance of the random forest model was better than that of other models with its high sensitivity, specificity, accuracy and AUC (Figure H), and its clinical decision curve showed benefits that were significantly better than other models.

Conclusion/Discussion: The prediction model based on the random forest machine learning algorithm developed in this study can well predict the occurrence of anastomotic leakage after rectal cancer, and more importantly, the prediction model can make decisions for prophylactic ileostomy that are better than the surgeon's experience, which is expected to provide auxiliary decision-making tools for the selective implementation of prophylactic ileostomy and optimize the prevention of anastomotic leakage.