Predicting postoperative malnutrition in patients with oral cancer: development of an XGBoost model with SHAP analysis and web-based application.

Background: Postoperative malnutrition, which significantly affects recovery and overall quality of life, is a critical concern for patients with oral cancer. Timely identification of patients at nutritional risk is essential for implementing appropriate interventions, thereby improving postoperative outcomes. Methods: This prospective study, which was conducted at a tertiary hospital in China between August 2023 and May 2024, included 487 postoperative oral cancer patients. The dataset was divided into a training set (70%) and a validation set (30%). Predictive models were developed via four supervised machine learning algorithms: logistic regression (LR), support vector machine (SVM), light gradient boosting machine (LGBM), and extreme gradient boosting (XGBoost). Nutritional risk was assessed via the Nutritional Risk Screening 2002 (NRS-2002) tool and diagnosed via the Global Leadership Initiative on Malnutrition (GLIM) criteria. Model performance was evaluated on the basis of discrimination, calibration, and clinical applicability, with SHAP analysis used for interpretability. Statistical analysis was conducted via R software, with appropriate tests for continuous and categorical variables. Results: Of the 487 oral cancer patients, 251 (51.54%) experienced postoperative malnutrition. The study cohort was split into a training set comprising 340 patients and a validation set comprising 147 patients. Seven key predictors were identified, including sex, T stage, repair and reconstruction, diabetes status, age, lymphocyte count, and total cholesterol (TC) level. The XGBoost model demonstrated an area under the curve (AUC) of 0.872 (95% CI: 0.836-0.909) in the training set and 0.840 (95% CI: 0.777-0.904) in the validation set. Calibration curves confirmed the model's robust fit, and decision curve analysis (DCA) indicated substantial clinical benefit. Conclusion: This study represents the first development of an XGBoost-based model for predicting postoperative malnutrition in patients with oral cancer. The integration of SHAP for model interpretability, along with the creation of an intuitive web tool, enhances the model's clinical applicability. This approach can significantly reduce malnutrition-related complications and improve recovery outcomes for oral cancer patients.