Submitted:
21 September 2026
Posted:
22 September 2026
You are already at the latest version
Abstract
Background/Objectives: Frozen triage narratives may contain prognostic information not captured by urgency categories or vital signs. We evaluated their incremental value for hospital admission and subsequent critical outcomes among adults assigned Emergency Severity Index (ESI) levels 3-5. Methods: This single-center retrospective study used routine electronic records from 1 April 2025 to 30 June 2026. Models were developed using 51,735 visits (4852 admissions) in April-December 2025 and evaluated without refitting in 40,433 visits in January-June 2026. Hospital admission included accepted interhospital inpatient transfers. Separate models predicted index-hospital ICU transfer or death among completed, observable episodes, including 285 development events. ESI-only, structured, text-only and combined logistic models were compared using discrimination, calibration and patient-cluster bootstrap intervals. Grouped cross-validation and fixed model variants assessed overfitting and source-quality sensitivity. Results: Temporal evaluation included 3517 admissions (8.70%). Adding text increased admission AUROC from 0.763 to 0.835 (95% CI 0.828-0.841) and average precision from 0.246 to 0.340. The combined model had a calibration slope of 1.083 and observed-to-expected ratio of 0.982. Among 39,697 secondary-analysis visits, 177 had ICU transfer or death (0.446%); combined AUROC was 0.924 (95% CI 0.908-0.940) and average precision 0.076. Text gains persisted without the seven measured fields or their missingness indicators. Conclusions: Frozen triage text added predictive information in a later calendar period. Admission prediction estimates recorded disposition rather than medical necessity. Low critical-outcome precision, incomplete follow-up exclusions and single-center design preclude clinical deployment without further validation. No specific research funding or registration identifier was reported.
Keywords:
emergency department
; triage
; natural language processing
; hospital admission
; temporal evaluation
; calibration
; intensive care
; mortality
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.