Prediction of severe acute pancreatitis using classification and regression tree analysis

Area of Science:

• Gastroenterology and Hepatology
• Medical Informatics
• Clinical Prediction Modeling

Background:

• Existing prognostic scoring systems for acute pancreatitis have limited clinical utility.
• There is a need for more accurate methods to predict severe acute pancreatitis (SAP).

Purpose of the Study:

• To develop and validate a decision tree model for predicting SAP using Classification and Regression Trees (CART) analysis.
• To identify key predictors for early risk stratification of SAP.

Main Methods:

• A cohort of 420 acute pancreatitis patients was randomly divided into training (2:1) and testing sets.
• Logistic regression identified initial predictors; CART analysis built the predictive tree model.
• Model performance was evaluated using Receiver Operating Characteristic (ROC) curves and compared against APACHE II score.

Main Results:

• Four variables (systemic inflammatory response syndrome, pleural effusion, serum calcium, BUN) were significant predictors.
• The CART-derived tree model, using pleural effusion, serum calcium, and BUN, effectively stratified patients into high (79.03%) and low (7.80%) risk groups.
• The tree model demonstrated superior predictive accuracy (AUC 0.84) over APACHE II (AUC 0.68) in the training set, validated with AUC 0.86 in the test set.

Conclusions:

• A decision tree model incorporating pleural effusion, serum calcium, and BUN shows significant promise for predicting SAP.
• This model offers a potentially valuable tool for early risk stratification and clinical decision-making in acute pancreatitis management.