Related Experiment Video
Updated: Mar 6, 2026
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Development and Internal Validation of an Early Warning Predictive Model for Critically Ill Patients in the Emergency
1Department of Emergency, Hebei General Hospital, Shijiazhuang, People's Republic of China.
Objective:
This study aimed to develop and internally validate an early warning predictive model to identify the risk of critical illness among patients presenting to the emergency department (ED).
Methods:
A retrospective analysis was conducted using clinical data from 3859 patients admitted between November 1, 2021 and December 31, 2021. Patients were randomly assigned to a training cohort (n = 2,703) and a validation cohort (n = 1,156) in a 7:3 ratio. Fourteen readily accessible physiological indicators obtained during the early stage of emergency department presentation were adopted as predictive parameters. Independent predictors of early critical risk were identified in the training cohort using generalized additive models, stepwise multivariate logistic regression and clinical practical considerations. The resulting model was used to stratify risk levels.
Results:
No statistically significant differences were observed in in baseline characteristics between the training and validation cohorts (p>0.05). Sex, age, heart rate, respiratory rate, systolic blood pressure, pulse oximetry saturation, level of consciousness, pupil status, mental status, and pain score were identified as independent predictors of critical risk (all p<0.05). Risk stratification using conditional inference trees categorized patients into low-risk (p≤0.129), medium-risk (0.129<p≤0.867), and high-risk (p>0.867) groups. The model demonstrated strong discriminatory ability, with area under the curve values of 0.926 (95% CI: 0.913-0.940) in the training cohort and 0.914 (95% CI: 0.889-0.938) in the validation cohort. Calibration was satisfactory, as indicated by Hosmer-Lemeshow test p-values of 0.318 and 0.654, respectively.
Conclusion:
The developed predictive model demonstrated good discrimination, calibration, and clinical utility for the early identification of patients at critical risk in the ED setting. All predictors can be obtained during the initial clinical assessment, which facilitates real-time application in triage. This practical accessibility supports the model's potential integration into routine emergency workflows and primary healthcare settings.
