Innovative Machine Learning Paradigms for Predicting Emergency Department Visits in T2DM Patients: A Comprehensive Analysis Across South Korea's Quintet Cohorts (Preprint)

• DOI: 10.2196/preprints.76639
• Publication Date: 2025-05-06T13:50:08Z
• Authors: Sunyoung Kim; Jaeyu Park; Hyunji Sang; Jiyoung Hwang; Tae Sun Park; You-Cheol Hwang; Hyunjung Lim; Ye Rim Lee; Hyejin Kang; Dong Keon Yon; Sang Youl Rhee

UNSTRUCTURED Background: Visiting emergency departments (ED) among patients with Type 2 Diabetes Mellitus (T2DM) is associated with adverse outcomes, including increased risks of hospitalization and mortality. Utilizing real-world clinical data, including prescription information, to predict the likelihood of ED visits could benefit primary care. This study aimed to apply machine learning (ML) methods to predict the likelihood of ED visits based on the electronic medical records (EMR) of patients with T2DM. Methods: We analyzed data from five independent EHR-based cohorts: Data from five institutions were combined and randomly divided into a training set (n=176,576) for model training and a test set (n=44,144) for evaluation. The primary outcome was the occurrence of first ED visits. Various machine learning (ML) models were evaluated through hyperparameter tuning within the training set, and the area under the receiver operating characteristic (AUROC) curve was calculated for the test set. Results: Among the 64,436 patients screened, xxx met the inclusion criteria, with 11,549 (17.92%) having at least on ED visit in the training set, while among the 220,720 patients screened, 49,770 (22.55%) met the inclusion criteria, with 49,770 (22.55%) having at least one ED visit in the test set The CatBoost model exhibited superior performance, achieving a mean AUROC of 78 % (95% CI, 94.4-94.9) in the training set and an AUROC of 87 % (95% CI, 94.4-94.9) in the test set. Among the top 20 strong predictive variables, diastolic blood pressure (DBP) was the most significant variable identified. Conclusions: In this study, we developed a predictive model for assessing the risk of ED visits in patients with T2DM. The model was validated using data from other hospitals, demonstrating its applicability in clinical settings for identifying patients at a heightened risk of requiring emergency care.