Machine learning–based dynamic prediction model for postoperative malnutrition risk in esophageal cancer patients A retrospective study

Abstract Background Malnutrition is the most common postoperative complication associated with esophageal cancer (EC), which severely impacts patient survival outcomes. However, early predictive strategies for postoperative malnutrition in EC patients remain underdeveloped. Objective This study aimed to develop a dynamic predictive model for postoperative malnutrition in EC patients, to facilitate early clinical identification and timely intervention of at-risk individuals. Methods In this study, 805 EC patients were enrolled as the training and validation set. Additionally, 101 EC patients were enrolled as the external validation set. Baseline data, pathological features, and biochemical indicators were collected at 1, 3, and 6 months postoperatively. LASSO regression was used to screen baseline characteristics at 1 month postoperatively. Four machine learning models (Logistic Regression, Support Vector Machine, Decision Tree, and Random Forest) were constructed and compared via 10-fold cross-validation to select the optimal one, which was then externally validated at 3 and 6 months. Model performance was evaluated using accuracy, sensitivity, specificity, and AUC, and SHAP was applied to interpret feature importance. Results The prediction model constructed using Random Forest demonstrated favorable predictive performance. The AUC value of the model at 1 month postoperatively was 0.974 (0.9491-1), and the AUC values with 95% confidence intervals (CI) for external validation at 3 and 6 months postoperatively were 0.814 (0.726-0.900) and 0.801 (0.714–0.889). Additionally, feature importance analysis revealed that BMI below 18.5 kg/m², weight loss > 3 kg within 3 months, and age > 70 years were associated with a higher risk of malnutrition. Conclusions This study developed a postoperative malnutrition risk prediction model for EC patients based on Random Forest, which accurately predicts malnutrition risk at 3 and 6 months postoperatively, thereby providing auxiliary decision-making support for early clinical identification.