Improved early prediction of acute pancreatitis severity using SHAP-based XGBoost model: Beyond traditional scoring systems

Abstract

Background: Acute pancreatitis (AP) progresses to severe forms in about 20 % of cases, leading to high morbidity and mortality. Traditional clinical scoring systems for severity prediction (e.g., Ranson, BISAP), are limited by delayed applicability, and suboptimal diagnostic accuracy.

Aims: To develop and validate machine learning (ML) models for early prediction of moderately severe and severe acute pancreatitis (MSAP-SAP), and to compare them with conventional scores.

Methods: A retrospective cohort of 816 patients (2014-2023) was analyzed. ML models were developed using admission (24-hour) and early (48-hour) data. Models were trained and tested using an 80:20 stratified split and evaluated based on ROC-AUC. F-Anova, Mutual Information and SHapley Additive exPlanations (SHAP) were used for feature selection. SHAP was also used for model interpretability.

Results: The XGBoost model with SHAP-based feature selection (XGB^SH) achieved the highest predictive performance with ROC-AUCs of 0.89 (24-hour) and 0.94 (48-hour) on the test cohort. Key predictive features included SIRS, BUN, CRP, creatinine, and pleural effusion. Compared to Ranson and BISAP (both ROC-AUC = 0.72), the XGB^SH models demonstrated superior accuracy and allowed flexible, threshold-based classification.

Conclusion: The proposed SHAP-enhanced XGBoost model offers a reliable and interpretable tool for early prediction of AP severity, improving clinical decision-making and patient management.

Keywords: Acute pancreatitis; Clinical decision support; Machine learning; Severity prediction.