Predicting infected pancreatic necrosis in acute pancreatitis using machine learning models and feature selection

Area of Science:

• Medical Informatics
• Machine Learning in Medicine
• Critical Care Medicine

Background:

• Infected pancreatic necrosis (IPN) is a severe complication of acute pancreatitis (AP), posing significant mortality risks.
• Early and accurate prediction of IPN remains a critical clinical challenge.
• Existing prediction methods often lack interpretability, hindering clinical trust and application.

Purpose of the Study:

• To develop and externally validate interpretable machine learning models for predicting individualized IPN risk.
• To enhance early risk stratification for patients with acute pancreatitis.
• To identify key clinical predictors associated with IPN development.

Main Methods:

• Retrospective analysis of 728 AP patients, with external validation on 166 patients.
• Development of interpretable machine learning models using embedded feature selection (regularized linear and tree-based algorithms).
• Evaluation of model performance using nested cross-validation, AUC, precision-recall, Brier score, and SHAP for explainability.

Main Results:

• The Random Forest model demonstrated superior performance with an external AUC of 0.764.
• Key predictors for increased IPN risk included Fibrinogen, APACHE II score, D-dimer, IL-6, and C-reactive protein.
• Higher Lymphocyte count and Hematocrit were identified as protective factors against IPN.

Conclusions:

• An interpretable Random Forest model offers robust discrimination and calibration for IPN prediction.
• The model provides a transparent, data-driven framework for early IPN risk stratification in AP patients.
• Prospective multicenter validation is recommended prior to clinical implementation.