Machine learning-based prediction model for post-ERCP cholangitis in patients with malignant biliary obstruction: a...

Area of Science:

• Gastroenterology and Hepatology
• Medical Informatics
• Oncology

Background:

• Endoscopic retrograde cholangiopancreatography (ERCP) is a key palliative treatment for unresectable malignant biliary obstruction (MBO).
• Post-ERCP cholangitis (PEC) significantly impacts MBO patient survival.
• Accurate prediction of PEC risk is vital for individualized treatment and improved prognosis, yet no clinical predictive models currently exist.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) prediction model for post-ERCP cholangitis (PEC) risk.
• To utilize multicenter cohorts for robust model development and validation.
• To provide a tool for early identification of high-risk MBO patients.

Main Methods:

• Data from 1026 MBO patients (training/internal test) and 395 (external validation) were analyzed.
• Six ML methods were employed to construct prediction models, with XGBoost selected for its superior performance.
• The SHapley Additive exPlanation (SHAP) method was used for model interpretability.

Main Results:

• The incidence of PEC was 9.5% (135/1421) among MBO patients.
• Independent risk factors for PEC included radiofrequency ablation, elevated white blood cell count, moderate jaundice, and abnormal serum amylase.
• The XGBoost model demonstrated strong predictive performance (AUC: 0.9654 training, 0.7670 internal, 0.7270 external) with good calibration and clinical net benefit.

Conclusions:

• An interpretable XGBoost model was successfully developed and validated using multicenter data to predict PEC risk in MBO patients.
• The model enables clinicians to identify high-risk patients preoperatively.
• This facilitates individualized treatment strategies, ultimately aiming to improve patient prognosis.