PREDICTING EMERGENCY ADMISSION USING PRE-VISIT INFORMATION: A MACHINE LEARNING APPROACH INTEGRATING ELECTRONIC MEDICAL RECORDS AND ADMINISTRATIVE CLAIMS DATA
Author(s)
TOMOKI ISHIKAWA, MBA,MPH,Ph.D.1, Keita Shibahashi, Ph.D.,MPH,MD2, Katsuhiko Ogasawra, Ph.D.,MBA3, Nobuya Yamashita, MBA4, Reina Taguchi, Ph.D.,MPH5, Hiroyuki Shiotsuki, Ph.D.1, Yasuaki Saijo, Ph.D.,MD1.
1Department of Social Medicine, Asahikawa Medical University, Hokkaido, Japan, 2Tertiary Emergency Medical Center, Tokyo Metropolitan Bokutoh Hospital, Tokyo, Tokyo, Japan, 3Faculty of Health Sciences, Hokkaido University, Hokkaido, Japan, 4ONO PHARMACEUTICAL CO., LTD., Osaka, Japan, 5National Center for Geriatrics and Gerontrogy, Obu, Japan.
1Department of Social Medicine, Asahikawa Medical University, Hokkaido, Japan, 2Tertiary Emergency Medical Center, Tokyo Metropolitan Bokutoh Hospital, Tokyo, Tokyo, Japan, 3Faculty of Health Sciences, Hokkaido University, Hokkaido, Japan, 4ONO PHARMACEUTICAL CO., LTD., Osaka, Japan, 5National Center for Geriatrics and Gerontrogy, Obu, Japan.
OBJECTIVES: In Japan, non-severe cases account for approximately 47% of emergency transports, contributing to the strain on emergency medical resources. Existing pre-hospital triage tools rely on symptom-based algorithms without accounting for individual patient characteristics. We aimed to develop a machine learning model predicting hospital admission following emergency department visits using pre-visit information, to support appropriate utilization of emergency care.
METHODS:
We used the Millennium Medical Record Database, a nationwide Japanese database (>2.8 million patients) integrating electronic medical records and administrative claims. Emergency visits between November 2016 and June 2024 were identified through procedure codes and free-text annotations (e.g., "transport," "walk-in"). The primary outcome was hospital admission. Features included age, sex, prior healthcare utilization, prescription history, comorbidities, and pre-visit symptom categories. To assess the incremental value of healthcare utilization history, we compared a minimum feature model (demographics and symptoms only) with a full feature model additionally incorporating healthcare utilization, prescriptions, and comorbidities. An XGBoost model was trained and evaluated using 5-fold cross-validation for each feature set.
RESULTS: Of 541 cases, 327 (60.4%) were ambulance-transported and 214 (39.6%) were walk-in. Median age was 72 years (IQR: 46-82); 51.1% were male. The overall admission rate was 44.1%. Admission rates were 57.1% for ambulance and 24.3% for walk-in patients, suggesting a mismatch between transport mode and clinical severity. The minimum feature model achieved an AUC of 0.661 (95% CI: 0.616-0.707), whereas the full feature model achieved a substantially higher AUC of 0.810 (95% CI: 0.773-0.846).
CONCLUSIONS: An machine learning model based on pre-visit information accurately predicted emergency hospital admission. Incorporating prior healthcare utilization markedly improved predictive performance over symptom-based features, supporting potential integration into pre-hospital triage as a complement to symptom-based assessment.
METHODS:
We used the Millennium Medical Record Database, a nationwide Japanese database (>2.8 million patients) integrating electronic medical records and administrative claims. Emergency visits between November 2016 and June 2024 were identified through procedure codes and free-text annotations (e.g., "transport," "walk-in"). The primary outcome was hospital admission. Features included age, sex, prior healthcare utilization, prescription history, comorbidities, and pre-visit symptom categories. To assess the incremental value of healthcare utilization history, we compared a minimum feature model (demographics and symptoms only) with a full feature model additionally incorporating healthcare utilization, prescriptions, and comorbidities. An XGBoost model was trained and evaluated using 5-fold cross-validation for each feature set.
RESULTS: Of 541 cases, 327 (60.4%) were ambulance-transported and 214 (39.6%) were walk-in. Median age was 72 years (IQR: 46-82); 51.1% were male. The overall admission rate was 44.1%. Admission rates were 57.1% for ambulance and 24.3% for walk-in patients, suggesting a mismatch between transport mode and clinical severity. The minimum feature model achieved an AUC of 0.661 (95% CI: 0.616-0.707), whereas the full feature model achieved a substantially higher AUC of 0.810 (95% CI: 0.773-0.846).
CONCLUSIONS: An machine learning model based on pre-visit information accurately predicted emergency hospital admission. Incorporating prior healthcare utilization markedly improved predictive performance over symptom-based features, supporting potential integration into pre-hospital triage as a complement to symptom-based assessment.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD155
Topic
Health Policy & Regulatory, Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
No Additional Disease & Conditions/Specialized Treatment Areas