Prediction of Early Clinical Deterioration in Acute Pancreatitis: A Machine Learning-based Approach.

INTRODUCTION

Early identification of clinical deterioration in acute pancreatitis patients is critical for improving outcomes in emergency settings. This study aimed to develop a machine learning model using only admission data to predict intensive care unit (ICU) needs within the first 72 hours and to compare its diagnostic performance with conventional scoring systems.

METHODS

This retrospective, single-center study included 448 patients with acute pancreatitis admitted to the emergency department. The dataset was randomly split into 70% training and 30% test sets. After SMOTE-based class balancing and 5-fold cross-validation, a Random Forest model was developed using 35 routinely collected variables. Model performance was evaluated using AUC, sensitivity, specificity, F1 score, and calibration. . The DeLong test compared diagnostic performance with CTSI, HAPS, Ranson, and Glasgow-Imrie scores.

RESULTS

The Random Forest model achieved an AUC of 0.974 (95% CI: 0.940-1.000), sensitivity of 92.9%, and specificity of 94.2% in the test set. It significantly outperformed all conventional scoring systems (P<0.05). Permutation-based feature importance analysis revealed calcium, Delta Neutrophil Index, urea, glucose, and acute peripancreatic fluid collection as the most influential variables.

CONCLUSIONS

Based solely on emergency admission data, our model demonstrated superior diagnostic accuracy compared with established scores and may serve as a practical early warning tool. Prospective multicenter validation is recommended.