Constructing a prediction model for acute pancreatitis severity based on liquid neural network

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Severe acute pancreatitis (SAP) presents significant morbidity and mortality risks.
• Early and accurate prediction of SAP severity is critical for effective patient management.
• Current diagnostic methods may lack the precision needed for timely intervention.

Purpose of the Study:

• To develop and evaluate a novel Liquid Neural Network (LNN) model for predicting SAP severity.
• To compare the LNN model's performance against traditional machine learning models.
• To identify key clinical predictors of SAP severity using explainable AI techniques.

Main Methods:

• Retrospective analysis of acute pancreatitis (AP) patient data from January 2020 to June 2024.
• Data preprocessing and Synthetic Minority Oversampling Technique (SMOTE) for imbalance.
• Development and comparison of Logistic Regression, Decision Tree, Random Forest, XGBoost, and LNN models.
• Performance evaluation using Area Under the ROC Curve (AUC) and other metrics.
• SHapley Additive exPlanations (SHAP) for model interpretability.

Main Results:

• The LNN model achieved the highest performance in predicting AP severity, with an AUC of 0.9659.
• LNN demonstrated superior accuracy, precision, recall, F1 score, and specificity (>0.90) compared to other models.
• SHAP analysis identified calcium levels, amylase activity, and basophil percentage as significant predictors of SAP.

Conclusions:

• The Liquid Neural Network (LNN) model shows exceptional potential for the early assessment of AP severity.
• LNN offers a promising tool for clinical decision support in managing AP patients.
• Improved early severity prediction can lead to optimized treatment strategies and better patient prognoses.