Interpretable machine learning model for predicting refeeding syndrome after colorectal cancer surgery

Refeeding syndrome (RFS) is a common yet frequently overlooked complication during postoperative nutritional support in patients undergoing colorectal cancer surgery. This study aimed to develop an explainable machine learning model for early risk prediction of RFS and to evaluate its predictive performance and clinical utility.

A total of 446 hospitalized patients who underwent curative colorectal cancer surgery were retrospectively included and randomly divided into a training set (n = 312) and a validation set (n = 134) in a 7:3 ratio. Based on clinical variables including preoperative nutritional status, electrolyte levels, and postoperative recovery indicators, four predictive models were constructed: logistic regression, random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost). Their predictive performance in the validation set was compared using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). The XGBoost model was further interpreted using SHapley Additive Explanations (SHAP) for both global and individual-level explanations.

In both the training and validation sets, the RFS group had significantly higher proportions of preoperative weight loss and comorbid diabetes than the non-RFS group (both P < 0.05). They also exhibited significantly lower preoperative serum phosphate and albumin levels, and longer postoperative recovery times for bowel sounds and first flatus (all P < 0.05). Among the models, XGBoost demonstrated the best performance in the validation set with an AUC of 0.872 (95 % CI: 0.805-0.925) and the lowest Brier score (0.113), offering the greatest net clinical benefit within the risk threshold range of 0.15-0.60. SHAP global interpretation revealed that preoperative serum phosphate, time to bowel sound recovery, preoperative albumin level, and time to first flatus were the most influential features. Low preoperative phosphate, prolonged bowel sound recovery, and low albumin levels substantially increased RFS risk. At the individual level, SHAP force plots visualized the personalized contribution paths of each feature, aiding in the identification of high-risk patients.

The XGBoost model combined with SHAP interpretation enables accurate and interpretable prediction of postoperative RFS risk. This approach may support individualized nutritional management strategies in patients following colorectal cancer surgery.