Machine learning-based prediction of hospital-acquired malnutrition at discharge using the GLIM criteria: A multicenter real-world study.

OBJECTIVE: This study aimed to investigate the incidence and clinical impact of hospital-acquired malnutrition (HAM) and develop a machine learning model for early risk prediction. METHODS: This secondary analysis included prospectively collected data from a multicenter observational study involving 34 tertiary hospitals in China. Adult patients hospitalized for 7 to 30 days were enrolled. Patients without malnutrition at admission were evaluated for HAM at discharge according to the Global Leadership Initiative on Malnutrition (GLIM) criteria. Machine learning models were developed using clinical and nutritional variables available at admission, and model performance was evaluated using discrimination, calibration, and clinical utility analyses. RESULTS: Among 4746 hospitalized patients, 1364 (28.7%) had malnutrition at admission. Of the 3382 patients without baseline malnutrition, 731 (21.6%) developed HAM during hospitalization. Patients with HAM had significantly higher rates of hospital-acquired complications (9.0% versus 2.9%), longer hospital stays (14.93 versus 13.53 days), and higher healthcare costs (¥48 980 versus ¥20 640) compared with those without HAM (all P < 0.001). Male sex, older age, pre-admission weight loss, reduced calf circumference, decreased handgrip strength, reduced muscle mass phenotype according to GLIM, and malignant tumors were associated with HAM. The XGBoost model demonstrated the best predictive performance, achieving an area under the receiver operating characteristic curve of 0.882 (95% CI: 0.857-0.906) in the independent test set. CONCLUSIONS: HAM is common among hospitalized patients without malnutrition at admission, and the developed machine learning model may enable early identification of high-risk patients for timely nutritional intervention.