Predictive Model for Early Fluid Imbalance in Severe Acute Pancreatitis

Abstract

Background: Fluid loss is a pathophysiological process in severe acute pancreatitis (SAP) that is important to control, but precise fluid therapy protocols are lacking. This study aims to build a prediction model for fluid loss in SAP by applying machine learning methods. Method: This is a retrospective observational study. It included SAP patients with appropriate fluid therapy volume and who visited the Emergency Department of Peking University Third Hospital from January 2016 to December 2020 within 48 hours after onset of the disease. The 48-hour fluid volume was used as a measure of fluid loss. The amount of 48-hour fluid rehydration was taken as the predictive end point, and associated variables were screened using the Lasso algorithm. Prediction models were established with five machine learning algorithms: Gradient Boosting Decision Tree, eXtreme Gradient Boosting, Light Gradient Boosting Machine, Catboost, and multiple linear regression. The validation was carried out with the test set, and the mean absolute error (MAE)，root mean square error (RMSE), R², and fitting curve were used to evaluate the prediction efficiency. The soft voting method was used to fuse the above five prediction models to improve the performance of model. The SHAP (SHapley Additive explanation) method was used to explain the optimal model. Finally, to facilitate practical clinical application, the model was evaluated by analyzing 10 cases from the test set. Results: A total of 308 cases were included, from which 90% of patients were randomly allocated as the training set, and the rest were included in the test set. The Lasso algorithm was used to screen the 16 variables most associated with the amount of 48-hour fluid replacement. Of the five machine learning algorithms that were used to build the prediction models, the MAE and RMSE values of the XGBoost algorithm were the smallest and the R² was the closest to 1, which indicated that the XGBoost was the optimal model. After model fusion, the model performance was further improved. In order to enhance the visualization of the model and to facilitate clinicians’ understanding of the model, we used the model interpretation tool SHAP to explain the optimal model, XGBoost. Application of the model in 10 actual cases showed that the difference between the predicted fluid loss and the actual 48-hour rehydration volume ranged from 31.07-329.80 mL, validating the model’s good predictive ability. Conclusion: In this study, we developed the Fluid Imbalance Predicting Model for SAP (FIPM-SAP), which can predict the specific amount of fluid loss in SAP patients. The predictive performance was good, demonstrating that the model has practical application for guiding clinicians in their assessment of 48-hour rehydration volume.