2045-P: Predicting Acute Care Utilization by People with Diabetes Using EHR Data

Introduction and Objective: Diabetes is associated with high rates of acute care utilization (ACU), posing a substantial burden on healthcare systems. Timely identification of individuals at higher risk for ACU can enable proactive interventions and improve resource allocation. The objective of this study was to develop and evaluate machine learning models to predict ACU within 365 days by people with diabetes. Methods: We conducted a retrospective cohort study using electronic health record (EHR) data from 23,052 patients with diabetes in an urban academic health system. EHR data included encounter history, demographics, vital signs, diagnoses, medications, and laboratory results. ACU during the most recent 365-day period was encoded as a binary outcome variable. Machine learning models spanning 15 different algorithms, including XGBoost, LightGBM, and CatBoost, were trained on the training set using 8-fold cross-validation with hyperparameter tuning. Model performance was evaluated on the test set using the macro F1 score, accuracy, and AUROC. Results: Among the evaluated models, LightGBM achieved the best overall performance, with a macro F1 score of 0.69, an accuracy of 74.91%, and an AUROC of 0.78. CatBoost and XGBoost demonstrated comparable performance, yielding similarly strong predictive accuracy. Analysis with permutation importance and SHAP from the top-performing model identified prior healthcare utilization, selected laboratory values, and demographic characteristics as key predictors of ACU. Conclusion: These findings demonstrate the effectiveness of ML models for predicting ACU among people with diabetes. While integrating multimodal EHR data is valuable, our results suggest that comparable predictive performance can be achieved using a targeted subset of high-impact features, enabling efficient and scalable risk stratification in clinical settings. Disclosure H. Sharma: None. G. Yu: None. J. Lee: None. Z. Obradovic: None. R. McCoy: None. D. Rubin: None.