Explainable Prediction of ICU Transfer in Acute Pancreatitis: A Neural Network Model With SHAP Interpretability for Clinical Decision Support

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: AUC 0.891 (excellent discrimination)

• Primary Outcome: ICU transfer requirement in acute pancreatitis patients

• Nnt Or Sensitivity: Sensitivity 0.852, Specificity 0.767, indicating good diagnostic performance

• Confidence Interval: Not reported

Clinical Application

High feasibility with web-based tool deployment, though requires validation in Australian healthcare settings Relevant for Australian EDs and gastroenterology services. Would require validation against local populations and integration with existing clinical pathways. TGA approval not required for decision support tools, but institutional governance needed. Adult patients presenting with acute pancreatitis in emergency departments and general medical wards

Abstract

OBJECTIVE: Acute pancreatitis (AP) is a life-threatening disorder commonly observed in emergency departments. Patients with acute pancreatitis may necessitate transfer to the intensive care unit (ICU) if standard treatments prove ineffective. The development of a machine learning (ML) model that can precisely forecast the necessity for ICU admission in patients with acute pancreatitis might dramatically enhance in-hospital healthcare for patients. METHODS: For this study, data were extracted from the MIMIC-III Clinical Database Care-Vue subset (MIMIC-III v 1.4), Medical Information Mart for Intensive Care-IV (MIMIC-IV v 3.1), and eICU Collaborative Research Database (v 2.0). Fifteen features were identified through Boruta's algorithm, Least Absolute Shrinkage and Selection Operator (Lasso), and Recursive Feature Elimination (RFE) utilizing the XGBoost. RESULTS: Seven machine learning models, including Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Neural Network (NNET), etc. were developed based on these features. The neural network approach demonstrated the highest reliability, attaining an AUC of 0.891, sensitivity of 0.852, and specificity of 0.767 in the test set. Meanwhile, an interpretability assessment of the most robust models was performed utilizing SHapley Additive exPlanations (SHAP) to prioritize feature importance and elucidate the final model. This model has been deployed on a web platform: https://icupredict.shinyapps.io/shiny6/. CONCLUSIONS: The neural network model effectively predicted whether patients with AP would be transferred to the ICU, demonstrating optimal performance. SIGNIFICANCE: We developed an interpretable ML model for prediction of ICU admission in AP patients. It provides clinicians with a reliable tool for early identification and treatment.