An interpretable XGboost algorithm for predicting 30-day mortality in acute pancreatitis using routine biomarkers

To develop and validate a machine learning (ML) algorithm for predicting 30-day mortality in adult patients with acute pancreatitis (AP) using routinely readily available laboratory parameters.

This retrospective cohort study analyzed 965 consecutive AP patients hospitalized between January 2017 and December 2019. Feature selection was performed using Least Absolute Shrinkage and Selection Operator (LASSO) regression, followed by the evaluation of eleven machine learning algorithms. Model performance was assessed using via the area under the receiver operating characteristic curve (AUC). The optimal model was further interpreted using SHapley additive explanation (SHAP).

The overall 30-day mortality rate was 4.5% (43/965). Among the evaluated algorithms, extreme gradient boosting (XGboost) demonstrated the best performance in both the development and validation cohorts. SHAP analysis identified creatine kinase (CK) as the most influential predictor, followed by lactate dehydrogenase (LDH), age, prothrombin time (PT), and carbohydrate antigen 19 - 9 (CA19-9). A simplified algorithm incorporating only the top five features achieved near-perfect discrimination in the development cohort (AUC = 1.000) and maintained robust performance in the validation cohort (AUC = 0.847). This five-feature XGboost algorithm showed superior performance compared to conventional prognostic scores in the development cohort.

An interpretable XGboost algorithm incorporating five key parameters (CK, LDH, age, PT, CA19-9) shows potential to predict predicts 30-day mortality in AP patients. This single-center retrospective model demonstrates promising internal performance but requires prospective multicenter external validation before clinical implementation. The web-based platform is provided as a research prototype for proof-of-concept and must not be used for real-time clinical decision-making at this stage.