What question did this study set out to answer?

This research aims to create predictive models for emergency department disposition based on triage-stage information. It evaluates the impact of adding clinical data to enhance prediction accuracy.

March 13, 2026BMC Emergency Medicine0 citationsOpen Access

Machine learning-based model for triage-stage prediction of emergency department disposition

KYK Yoshimura TMTakeshi Moriguchi

Key Points

This research aims to create predictive models for emergency department disposition based on triage-stage information. It evaluates the impact of adding clinical data to enhance prediction accuracy.
Conducted a retrospective cohort study at a secondary-level emergency facility in Japan.
Analyzed 2,080 adult emergency department encounters to predict disposition outcomes.
Developed and compared three predictive models: Pre-test, Clinical, and Comprehensive, using logistic regression.
Assessed model performance using area under the curve and stratified 5-fold cross-validation.
Of the 2,080 encounters, 1,252 (60.2%) were discharged home, while 828 (39.8%) required admission or worse.
The Comprehensive model achieved an area under the curve of 0.921, outperforming the Clinical model (0.887) and the Pre-test model (0.796).
Sensitivity and specificity were highest for the Comprehensive model at 0.850 and 0.855, respectively, indicating better predictive performance.

Abstract

Early disposition decisions in the emergency department (ED) are sometimes made before laboratory results, imaging, or diagnostic labels become available. We developed disposition prediction models using only triage-stage information and quantified the incremental gain from adding downstream clinical data within the same cohort. We conducted a single-center retrospective cohort study at a secondary-level emergency medical facility in Japan, including adult ED encounters. The outcome was ED disposition, dichotomized as discharged home versus admission or worse (hospital admission, intensive care unit admission, death in the ED, or inter-hospital transfer). We trained L2-regularized logistic regression models using prespecified predictor sets. The Pre-test (triage-stage) model included demographics, trauma status, comorbidities, symptoms, and vital signs. The Clinical model was defined as the Pre-test model plus four binary clinical-process indicators (yes/no): blood gas analysis performed, computed tomography performed, magnetic resonance imaging performed, and consultation requested. The Comprehensive model was defined as the Clinical model plus laboratory tests and diagnosis-related variables. Discrimination was assessed using stratified 5-fold cross-validation with fold-wise preprocessing to prevent leakage. Among 2,080 encounters, 1,252 (60.2%) were discharged home and 828 (39.8%) had admission or worse. The area under the receiver operating characteristic curve was highest for the Comprehensive model 0.921 (95% confidence interval (CI), 0.909–0.933), followed by the Clinical model 0.887 (95% CI, 0.873–0.902) and the Pre-test model 0.796 (95% CI, 0.776–0.816). At the Youden index-derived threshold, sensitivity/specificity were 0.850/0.855 (Comprehensive), 0.855/0.804 (Clinical), and 0.676/0.784 (Pre-test). Permutation and ablation analyses suggested that discrimination in the Pre-test model was supported by multiple triage-stage predictors, including age and acute neurologic or mental status indicators. ED disposition can be predicted to some extent using triage-stage information alone, and performance improves when clinical-process indicators, laboratory tests, and diagnosis-related variables are added. These results clarify both the potential and the limitations of triage-only decision support for early ED workflow planning.

AI से पूछें

Bookmark

View Full Paper