Related Experiment Video
Updated: Sep 15, 2025
An Orthotopic Resectional Mouse Model of Pancreatic Cancer
Published on: September 24, 2020
Machine learning model for postpancreaticoduodenectomy haemorrhage prediction: an international multicentre cohort
Zhe Zhang1, Xueping Zhao2, Minjie Shang3
1The Second Clinical Medical College, Zhejiang Chinese Medical University, Hangzhou, China.
Objectives:
To develop and validate a machine learning model for precise risk stratification of postpancreaticoduodenectomy haemorrhage (PPH), enabling early identification of high-risk patients to guide clinical intervention.
Design:
Retrospective international multicentre cohort study with model development and external validation.
Setting:
Training data from the American College of Surgeons-National Surgical Quality Improvement Program database (USA, 2014-2017) and external validation data from the National Cancer Center (China, 2014-2019).
Participants:
3609 patients in the training cohort and 1347 in the validation cohort undergoing pancreaticoduodenectomy. Patients with missing data or non-relevant variables were excluded.
Primary And Secondary Outcome Measures:
Primary outcome: clinically relevant PPH (International Study Group of Pancreatic Surgery grades B/C).
Secondary Outcomes:
model discrimination (area under the curve (AUC)), calibration (Hosmer-Lemeshow test), clinical utility (decision curve analysis) and risk stratification performance.
Results:
The least absolute shrinkage and selection operator (Lasso)-gradient boosting machine model identified eight predictors: albumin, haematocrit (HCT), American Society of Anesthesiologists (ASA) class, operative time, vascular resection, sepsis, reoperation and pancreatic fistula. It achieved AUCs of 0.84 (95% CI 0.82 to 0.86) in training and 0.82 (95% CI 0.78 to 0.85) in validation, outperforming logistic regression and other machine learning models. Risk stratification into low-risk, medium-risk and high-risk groups showed strong discriminatory power (AUCs: 0.72-0.75). Decision curve analysis confirmed net clinical benefit, and SHapley Additive exPlanations values highlighted HCT and operative time as top contributors. The model was deployed as an interactive application for real-time risk assessment.
Conclusions:
This novel machine learning model for PPH prediction integrates interpretable risk stratification and demonstrates robust performance across international cohorts. Its deployment as a clinical tool may facilitate proactive management of high-risk patients. Prospective validation is warranted prior to broad implementation.
