Integrating Feature Selection, Machine Learning, and SHAP Explainability to Predict Severe Acute Pancreatitis

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Machine Learning in Healthcare

Background:

• Severe acute pancreatitis (SAP) presents significant morbidity and challenges in early risk stratification.
• Conventional scoring systems for SAP require serial observations, limiting their utility for immediate assessment.
• Developing accurate early prediction models for SAP at emergency department (ED) presentation is crucial.

Purpose of the Study:

• To develop and compare supervised machine learning (ML) pipelines for early SAP prediction.
• To integrate feature selection and SHAP-based explainability into ML models for SAP prediction.
• To assess the performance and interpretability of ML models using data available upon ED arrival.

Main Methods:

• A retrospective, single-center cohort study of adult patients with acute pancreatitis was conducted.
• Thirty-six ML pipelines were created by pairing six feature-selection methods with six classifiers.
• SHAP (SHapley Additive exPlanations) was employed for model interpretability, and performance was evaluated using bootstrapping.

Main Results:

• The best-performing ML pipelines achieved an area under the receiver operating characteristic curve (AUROC) between 0.750 and 0.826.
• The top pipeline (Recursive Feature Elimination with Random Forest features + k-Nearest Neighbors classifier) reached an AUROC of 0.826.
• SHAP analysis confirmed that routinely available variables contributed plausibly to SAP prediction.

Conclusions:

• Integrating feature selection with ML enables accurate and interpretable early prediction of SAP using ED presentation data.
• The developed ML approach identifies actionable predictors, potentially supporting earlier triage and resource allocation.
• External validation of these ML models is recommended to confirm their generalizability.