Usefulness of Random Forest Algorithm in Predicting Severe Acute Pancreatitis

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Modeling

Background:

• Severe acute pancreatitis (SAP) prediction remains a clinical challenge.
• Developing accurate and interpretable models is crucial for timely intervention.

Purpose of the Study:

• To develop an interpretable random forest (RF) model for predicting severe acute pancreatitis (SAP).
• To compare the performance of the RF model against logistic regression (LR) and BISAP score.

Main Methods:

• Retrospective review of clinical and laboratory data from 648 acute pancreatitis patients.
• Development of RF and LR models using a 3:1 training/test split.
• Validation using tenfold cross-validation, ROC curves, and precision-recall curves.
• Utilized Local Interpretable Model-agnostic Explanations (LIME) for model interpretability.

Main Results:

• The RF model achieved high discriminatory performance with AUCs of 0.89 (training) and 0.96 (test).
• Key predictors identified by RF include BUN, serum creatinine, albumin, HDL, LDL, calcium, and glucose.
• The RF model outperformed both LR and BISAP score in AUC and diagnostic accuracy.
• LIME plots provided visualized explanations for individual predictions.

Conclusions:

• An interpretable RF model demonstrates superior performance in predicting SAP.
• LIME-based interpretation facilitates individualized risk assessment in clinical practice.
• A nomogram incorporating albumin, creatinine, glucose, and pleural effusion also shows predictive utility.