Standardized Multicenter Critical Care Database Integrating Minute-Level Vital Signs, Laboratory Tests, Interventions, and Outcomes: Profile of the OneICU Database
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Introduction
Standrdized intensive care unit (ICU) databases with high frequency data collection from multicenter electronic medical records remain scarce. We aimed to describe the profile of newly developed OneICU database, which includes minute-level recordings of vital signs, laboratory values, interventions, and diagnosis codes, and to evaluate the importance of vital sign measurement frequency and laboratory data completeness for developing machine learning–based clinical decision support.
Methods
This retrospective, multicenter observational study collected critically ill patient data from 12 tertiary care hospitals from 2013 to 2025. Patient demographics, measurement frequency of vital signs and laboratory tests, as well as the availability of the Sequential Organ Failure Assessment (SOFA) score components were compared across three large ICU databases: OneICU, Medical Information Mart for Intensive Care (MIMIC)-IV and eICU. We then evaluated the prediction accuracy of five machine learning models to forecast hypotensive events 60–120 minutes in advance, defined as a median invasive mean arterial pressure (MAP) < 65 mmHg or vasopressor initiation, using different MAP sampling frequencies: OneICU 1 minute, OneICU 5 minute, OneICU hourly, MIMIC IV hourly, and eICU 5 minute.
Results
OneICU currently includes 152,269 ICU stays from 127,757 unique patients. Compared with MIMIC-IV and eICU, OneICU captured more frequent vital signs (minute-level vs. hourly in MIMIC-IV and every five minutes in eICU) and provided broader availability of SOFA score components. In particular, the respiratory component was available for 73.6 % of stays in OneICU versus 37.8 % in MIMIC IV and 30.9 % in eICU, and the liver component for 93.3 % versus 45.2 % and 41.3 %, respectively. The test-set area under the receiver operating characteristic curve was highest for the OneICU 1-minute model (0.942), followed by OneICU 5-minute model (0.939), OneICU hourly model (0.901), eICU 5-minute model (0.899), and MIMIC IV hourly model (0.799).
Conclusions
A high resolution, multicenter ICU database integrating minute level vital sign recordings with comprehensive SOFA score coverage is feasible and was associated with superior hypotension prediction performance. OneICU enables detailed analyses of ICU trajectories and addresses the current scarcity of large scale ICU data from Asian populations.
