# Explainable machine learning model to predict refeeding syndrome in patients with severe acute pancreatitis

## TL;DR

This study develops a machine learning model to predict refeeding syndrome in severe pancreatitis patients, helping doctors decide when to start nutrition and intervene early.

## Contribution

The novel contribution is an explainable machine learning model for predicting refeeding syndrome risk in SAP patients using clinical features and SHAP analysis.

## Key findings

- The GBM model achieved an AUC of 0.851 in training and 0.762 in testing for RFS prediction.
- SHAP analysis identified serum potassium, sodium, and calcium as the most important predictors of RFS risk.
- The model showed good calibration and net benefit across clinical decision thresholds.

## Abstract

To construct and validate a risk prediction model for refeeding syndrome (RFS) in patients with severe acute pancreatitis (SAP), identify high-risk individuals before overt electrolyte abnormalities occur, and provide decision support for the timing of enteral nutrition initiation and early personalized intervention.

A retrospective cohort study was conducted on SAP patients admitted to Xuzhou Medical University Affiliated Hospital (XYFY) between September 2018 and September 2025. Patients were divided into RFS and Non-RFS groups based on the development of RFS after initiating enteral nutrition. Clinical data differences between groups were compared, least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, and six machine learning (ML) algorithms were applied to build prediction models. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curves. SHapley Additive exPlanations (SHAP) analysis was performed to interpret the contribution of key features.

Seven predictive features were identified for model construction. The gradient boosting machine (GBM) model exhibited good generalization ability, with area under the curve (AUC) values of 0.851 (95% CI: 0.809–0.894) in the training set and 0.762 (95% CI: 0.672–0.852) in the testing set. Calibration curves confirmed consistency between predicted probabilities and actual outcomes, while decision curves demonstrated favorable net benefits across different clinical decision thresholds. SHAP analysis ranked feature importance as follows: serum potassium (K), serum sodium (Na), serum calcium (Ca), gastrointestinal decompression, blood urea nitrogen (BUN), diabetes mellitus (DM) history, and diuretic use.

The GBM model effectively predicts RFS risk in SAP patients after initiating enteral nutrition.