Predicting High-Risk Hospitalization in Diabetic Outpatients Using Machine Learning: A Real-World Study in a Tertiary Hospital in Brazil

Hospital admissions among diabetic patients pose a major clinical and economic challenge, especially in middle-income countries. This study evaluated the performance of machine learning models in predicting hospitalizations within 12 months, using electronic health record data from 5,964 diabetic outpatients treated at a Brazilian tertiary hospital. Four algorithms were tested: logistic regression, random forest, multilayer perceptron, and gradient boosting. Gradient boosting achieved the best performance (AUC = 0.83; Brier score = 0.165). The most predictive variables included prior hospitalizations, glycated hemoglobin, creatinine, and insulin use. Results suggest that real-world data–driven approaches can help identify high-risk patients and support proactive care strategies in outpatient settings. This study contributes to the development of AI-based clinical tools in the context of Brazilian public healthcare.