Design and Feasibility Assessment of a Compact Emergency Unit in Rural and Remote Areas: A Multicenter Analysis of KTAS-Based Triage Data

Background/Objectives: Emergency department (ED) overcrowding burdens rural and remote areas where geographic isolation limits timely care. The Compact Emergency Unit (CEU)—a 24 h facility with remote physician oversight—has been proposed but lacks an empirical foundation. We aimed to (1) quantify CEU-eligible (final KTAS 4–5) patients in a multicenter ED cohort; (2) compare their operational metrics with non-eligible patients; (3) characterize hourly demand for facility planning; and (4) develop machine-learning models for non-discharge prediction within this low-acuity stratum. Methods: Retrospective analysis of 12 months (January–December 2025) of NEDIS data from two Korean university-affiliated EDs. Effect sizes (Cliff’s δ, Cramér’s V) were reported alongside p-values. Three classifiers (logistic regression, random forest, and XGBoost) were developed with patient-level cross-validation, comparing a 16-feature baseline and a 22-feature set augmented with arrival vital signs. Calibration and decision curve analysis were performed. Results: Of 34,544 valid triage visits (27,743 unique patients), 9871 (28.6%) were CEU-eligible. They had shorter LOS (92 vs. 171 min; Cliff’s δ = −0.51), 98.8% symptomatic home discharge, and a median of 0 specialty consultations. Nighttime visits comprised 43.7% of CEU-eligible encounters, peaking at 20:00 (1.76 visits/h/day). The non-discharge rate was 1.20% (118/9871). The vital-augmented random forest reached AUROC 0.794 (95% CI 0.758–0.829); XGBoost calibration was near-perfect (ECE 0.020). A combined ML-or-vital-sign screening rule raised non-discharge sensitivity to 94.1%. Conclusions: Approximately 29% of ED visits could be CEU-suitable. Single-modality machine learning is insufficient for safety-critical triage, but a layered ML-plus-vitals screening approach achieves operationally relevant sensitivity. Prospective implementation studies are required before clinical deployment.