AI Predicts Emergency Room Visits for Diabetes Patients in Advance... Achieves 87% Accuracy

National Institute of Health Analyzes Clinical Data from 220,000 Patients in Korea
Blood Pressure, Kidney Function, and Blood Glucose Management Are Key Predictors

A team of domestic researchers has developed a technology that predicts the risk of emergency room visits for diabetes patients in advance by utilizing artificial intelligence (AI).

The Korea Disease Control and Prevention Agency (KDCA) National Institute of Health announced on August 11 that it has developed a machine learning model to predict emergency room visit risk by analyzing electronic medical records (EMR) of 220,720 type 2 diabetes patients from five medical institutions in Korea.

The research team analyzed patient data accumulated from 2008 to 2022, based on large-scale multicenter diabetes big data established by the National Institute of Health. Among the subjects analyzed, 22.6% (49,770 patients) visited the emergency room at least once within one year before or after being diagnosed with diabetes. Although diabetes is generally managed through outpatient care, the risk of emergency room visits, hospitalization, or death can increase if acute complications occur, such as severe hypoglycemia, hyperglycemia, or diabetic ketoacidosis, which makes early risk prediction critical.

Patients who visited the emergency room were found to be older on average and had relatively poorer blood glucose control and kidney function compared to those who did not. Additionally, the proportions of insulin and diuretic use were about twice as high, and the prevalence of comorbid hypertension and cerebrovascular diseases was also higher among emergency room visitors.

The researchers compared multiple AI models using 55 clinical variables routinely checked during medical care—such as blood pressure, blood glucose test results, kidney function, and medication prescription history. As a result, the 'CatBoost' model demonstrated the most outstanding performance, achieving a prediction accuracy of 87% and an AUROC (Area Under the Receiver Operating Characteristic Curve: a measure of a model's ability to distinguish between patients who visited and did not visit the emergency room) of 0.87 on validation data, which significantly outperformed the conventional logistic regression model (AUROC 0.73).

The analysis found that the most critical factors for predicting emergency room visit risk were diastolic blood pressure, serum creatinine (an indicator of kidney function), and systolic blood pressure, in that order. The researchers noted that, since most of these factors can be managed through medication and lifestyle improvements, early identification and proactive intervention for high-risk patients during medical treatment could reduce emergency room visits and acute complications. They further highlighted that the ability to predict risk using only existing clinical information, without any additional tests, is also a major advantage.

Juhyun Lim, Head of the Division of Endocrinology and Nephrology Research at the National Institute of Health, said, "This study is an example showing how real-world clinical data, based on multicenter diabetes big data, can be used to predict patients' health risks and devise preventive strategies. In the future, we aim to distribute predictive models that can be used in primary care facilities as well, to help identify and manage the risk of diabetes complications early."

Wonho Kim, Head of the Division of Integrated Research on Chronic Diseases at the National Institute of Health, commented, "This study offers the possibility of transitioning from a focus on post-treatment care to a more proactive preventive and management system that predicts and manages risk. For diabetes patients, it is essential to prevent emergencies and to pay close attention to even minor health changes through lifestyle management and regular health examinations."

The results of this study were published in the July edition of the international journal PLOS ONE.

This content was produced with the assistance of AI translation services.