EP543 - Machine learning-based risk scoring using clinical and surgical variables for severe complications after rectal cancer surgery

ePoster Presenter(s)

• 

  Seungho Song

  Assistant professor
  Kyungpook National University Hospital
  Kyungpook National University
  Daegu, Taegu-jikhalsi, Republic of Korea

  Disclosure information not submitted.

Author(s)

• HK

  Hyeok Kyu Kwon, n/a

  Ph.D Student
  Pohang University of Science and Technology
  Pohang-si, Kyongsang-bukto, Republic of Korea

  Disclosure information not submitted.
• 

  Seungho Song

  Assistant professor
  Kyungpook National University Hospital
  Kyungpook National University
  Daegu, Taegu-jikhalsi, Republic of Korea

  Disclosure information not submitted.
• SP

  Soo Yeun Park, M.D.

  Professor
  Kyungpook National University
  Daegu, Taegu-jikhalsi, Republic of Korea

  Disclosure information not submitted.
• 

  Hye-Jin Kim, MD

  Associate professor
  Kyungpook National University Chilgok Hospital
  Kyungpook National University
  Daegu, Kyongsang-bukto, Republic of Korea

  Disclosure information not submitted.
• 

  Hoyeon Kwon, MD

  Surgical Resident, Department of Surgery
  Department of Surgery, Kyungpook National University Hospital, Daegu, Republic of Korea
  Kyungpook National University School of Medicine
  Daegu, Kyongsang-bukto, Republic of Korea

  Disclosure information not submitted.
• JL

  Jeongjik Lee, PhD Student

  Graduate Student Researcher
  POSTECH (Pohang University of Science and Technology)
  Pohang, Kyongsang-bukto, Republic of Korea

  Disclosure information not submitted.
• MC

  Minwoo Chae, n/a

  Associate Professor
  Department of Industrial and Management Engineering, POSTECH
  Pohang, Kyongsang-bukto, Republic of Korea

  Disclosure information not submitted.

Senior Author(s)

• GC

  Gyu-Seog Choi, Colorectal surgeon

  Professor
  Kyungpook national university chilgok hospital
  Daegu, Taegu-jikhalsi, Republic of Korea

  Disclosure information not submitted.

Purpose/Background: Postoperative complications remain a major concern after rectal cancer surgery, often leading to prolonged recovery, increased costs, and adverse oncologic outcomes. Machine learning (ML) approaches may offer superior accuracy by incorporating multidimensional clinical and surgical variables.

Methods/Interventions: We retrospectively analyzed data from 3,402 patients who underwent rectal cancer surgery with anastomosis between 2011 and 2024 and prospectively collected 101 additional patients from 2024 to 2025 for prospective validation in a temporally independent cohort. Four ML algorithms-logistic regression, random forest, artificial neural network, and gradient boosting-were developed to predict severe postoperative complications, defined as Clavien-Dindo grade ≥3. Model training and hyperparameter optimization were performed with five-fold cross-validation. Sensitivity was prioritized as the primary evaluation metric, and feature importance was assessed using gradient boosting.

Results/Outcomes: Severe complications occurred in 9.8% of the retrospective cohort and 13.9% of the prospective cohort. In internal validation, the random forest model achieved the best performance (AUROC 0.898; sensitivity 0.685), while gradient boosting performed best in prospective validation (AUROC 0.910; sensitivity 0.707). Key predictors were perforation, intraoperative transfusion, protective stoma, bowel obstruction, and cerebrovascular accident. Partial dependence plots confirmed their consistent impact on complication risk.

Conclusion/Discussion: We developed and prospectively validated an ML model with robust discriminatory performance in predicting severe complications after rectal cancer surgery. By integrating preoperative and intraoperative variables, the model enables individualized perioperative risk stratification. Further multicenter validation is warranted to confirm generalizability and support implementation into clinical practice.

Receiver operating characteristic curves for the internal and prospective validation sets
ROC = Receiver operating characteristic
Partial dependence plot from the Gradient Boosting model
CVA = Cerebrovascular accident; HTN = Hypertension; AR = Anterior resection