Predicting the acute pancreatitis severity with multi-machine learning models: constructing an online prediction platform

TL;DR

This paper develops an online platform using machine learning to predict the severity of acute pancreatitis based on patient data, aiming to help doctors make timely treatment decisions.

Contribution

The novel contribution is the construction of a web-based prediction platform using multiple machine learning models to assess acute pancreatitis severity.

Findings

LightGBM achieved the highest predictive accuracy with AUC values of 0.9726 (training) and 0.9301 (test).

Calcium (Ca), white blood cell (WBC), alpha-hydroxybutyrate dehydrogenase (α-HBDH), and glucose (Glu) were identified as key predictive features for severe acute pancreatitis.

The online platform enables rapid and effective assessment of severe acute pancreatitis risk for clinical use.

Abstract

Early assessment of acute pancreatitis (AP) severity is critical. We therefore built a web-based calculator that instantly estimates the probability that a patient admitted with AP will progress to the severe form.

Clinical records for patients who were diagnosed as AP at the Second Affiliated Hospital of Guilin Medical University between the start of 2016 and May 2025 were retrospectively examined. The dataset was randomly divided into training set (70%) and test set (30%). For the traditional machine learning models, we employed 5-fold cross-validation combined with random search for hyperparameter optimization during training. Feature selection was performed using Random Forest (RF) and the Least Absolute Shrinkage and Selection Operator (LASSO) methods. Model construction included Logistic Regression (LR), Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory Network (LSTM). The area under the receiver operating characteristic curve (AUC), among other metrics, served to evaluate model efficacy. SHapley Additive exPlanations (SHAP) and Partial Dependency Plots (PDP) were employed to explain model predictions, and a clinical application risk prediction platform was further developed.

1289 patients with AP were included, with 11 variables screened to develop 10 models. Among these, the LightGBM demonstrated the highest predictive accuracy on training and test sets, with AUC (95% CI) values of 0.9726 (0.9626-0.9818) and 0.9301 (0.9113-0.9481), respectively. SHAP and PDP analyses identified Ca, WBC, α-HBDH, and Glu as key predictive features for severe acute pancreatitis (SAP). Calcium levels exerted a negative influence on SAP prediction, whereas WBC, α-HBDH, and Glu exerted positive influences, exhibiting positive synergistic effects among these three variables.

Our study highlights the substantial predictive potential of Ca, WBC, α-HBDH, and Glu for SAP. We have built a predictive online platform for clinical use, enabling healthcare professionals to rapidly and effectively assess SAP risk, thereby facilitating timely intervention and treatment.

Figures