Development and validation of a machine learning-based model for diagnosing perioperative malnutrition in older adults with hip fracture.

Background: The prevalence of malnutrition is significant among older adults with hip fractures, while existing screening tools face challenges such as complex procedures and a limited ability to objectively classify malnutrition status. This study aimed to develop and test a machine learning-based diagnostic model for identifying malnutrition guided by the Global Leadership Initiative on Malnutrition (GLIM) criteria. Methods: A cross-sectional study was conducted, enrolling patients from four tertiary hospitals in Xi'an between January and September 2024. Feature selection was performed using the Boruta and least absolute shrinkage and selection operator (LASSO) methods. Diagnostic classification models were constructed using five machine learning (ML) algorithms: logistic regression (LR), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and artificial neural network (ANN). Model performance was evaluated through receiver operating characteristic (ROC) analysis, decision curve analysis (DCA), and calibration curves. Shapley additive explanation (SHAP) values were applied for model interpretation. A total of 526 patients were ultimately included (385 in the training set and 141 in the external validation set), meeting the sample size requirements. Results: The prevalence of malnutrition among older adults with hip fractures was 38.78%. The key factors associated with malnutrition were age, body mass index (BMI), decreased appetite, chronic obstructive pulmonary disease (COPD), age-adjusted Charlson Comorbidity Index (aCCI), depression, albumin (ALB), and American Society of Anesthesiologists (ASA) classification, with the aCCI showing the greatest feature importance. The areas under the ROC curve (AUCs) for internal and external validation ranged from 0.8605 to 0.9424 and 0.8353 to 0.8565, respectively. All models except the ANN demonstrated good calibration, and DCA confirmed clinical usefulness across all models. The LR and XGBoost models demonstrated the best overall discriminative performance. Based on the LR model, an online calculator was developed and is accessible at: https://hip-fracture-malnutrition-test.shinyapps.io/dynnomapp/. Conclusion: This study employed ML to systematically assess the status and associated factors of perioperative malnutrition in elderly patients with hip fractures and developed quantifiable prediction models. The models demonstrated robust performance in both internal and external validation and were visualized