ED-Triage: Clinical & Computational Insights

• ED-triage is a process that uses both traditional acuity scales and modern machine-learning systems to assign patient urgency and predict resource needs from initial contact data.
• It integrates structured data like vital signs with unstructured inputs such as triage notes and behavioral cues to improve decision-making and patient routing.
• Recent advances in multimodal fusion, Bayesian models, and graph learning have enhanced accuracy, operational efficiency, and fairness in dynamic, real-time emergency care.

ED-triage denotes the clinical and computational processes by which emergency departments assign urgency, anticipate resource needs, determine immediate disposition, and, increasingly, trigger downstream testing or surveillance from information available at first contact. In contemporary research, it spans traditional nurse-led acuity scales such as the Emergency Severity Index (ESI), CTAS, ATS, MTS, KTAS, and FRENCH, as well as machine-learning systems that operate on structured vitals, demographic variables, chief complaints, triage notes, and sequential bedside cues (Yang et al., 10 Apr 2026, Defilippo et al., 2024, Han et al., 2024, Lansiaux et al., 1 Jul 2025).

1. Clinical scope and formal triage frameworks

ED triage is classically the first sorting step of emergency care: patients are assigned a priority level that governs waiting time, monitoring intensity, and access to staff and resources. In the ESI formulation emphasized in recent work, the five levels are: ESI 1 for immediate life-saving intervention, ESI 2 for high risk, altered mental status, or severe pain/distress, ESI 3 for stable patients needing many resources, ESI 4 for stable patients needing one resource, and ESI 5 for stable patients needing no resources (Yang et al., 10 Apr 2026). Other papers situate ED-triage within a broader international ecology of scales, including CTAS, ATS, MTS, KTAS, and the six-level FRENCH scale used in French emergency care (Defilippo et al., 2024, Han et al., 2024, Lansiaux et al., 1 Jul 2025).

The research literature also treats triage as more than acuity assignment. Some systems predict the number of ED resources a patient will consume, especially for the overloaded middle of ESI, where many patients are placed in ESI 3 and then processed first-come, first-served (Gligorijevic et al., 2018). Others treat triage as a routing and scheduling problem, combining urgency with expected physician time and queue optimization (0810.3671). A plausible synthesis is that ED-triage has evolved from a static ordinal label into a multi-objective inference layer spanning urgency, throughput, destination, and safety.

2. Data modalities and the informational structure of triage

The dominant data sources at triage are structured and unstructured. Structured inputs recurrently include age, sex, pain score, arrival mode, and vital signs such as temperature, heart rate, respiratory rate, oxygen saturation, and blood pressure; unstructured inputs include chief complaint fields, free-text triage notes, medication lists, nursing assessments, and brief histories (Yang et al., 10 Apr 2026, Ivanov et al., 2022, Lu, 2023). Several studies deliberately restrict themselves to information available before labs and imaging, preserving the operational meaning of triage as a front-door decision.

A consistent finding is that triage text is short, clinically dense, and institution-specific. ED triage notes are described as brief free-text records written under time pressure, with non-standard grammar, misspellings, abbreviations, and local shorthand; these properties are central to both their value and their difficulty (Chen et al., 1 Jun 2026). In pediatric settings, text can be especially important because developmental variation makes adult-trained vital-sign heuristics unreliable, while chief complaints encode cues such as feeding difficulty, parental concern, and behavior change that are not captured in tabular vitals (Yang et al., 10 Apr 2026). Beyond acuity prediction, triage notes also support surveillance tasks such as self-harm detection and vaccine-safety signal detection, precisely because they record the earliest narrative linking symptoms, timing, and context (Khademi et al., 24 Jul 2025, Chen et al., 1 Jun 2026).

Recent work also broadens the notion of triage input to include sequential behavioral signals. A Bayesian framework uses time to first urination as a bedside cue that updates admission risk over the ED stay, explicitly modeling the distinction between patients who have voided and those who have not (Senda et al., 17 Oct 2025). This suggests that ED-triage need not remain a single-time-point function of arrival data.

3. Modeling paradigms

Current ED-triage research spans classical boosting, multimodal deep learning, graph learning, and LLM-based orchestration. A prominent structured-plus-text approach is KATE, which combines machine learning with clinical NLP over 147,052 analytic encounters from two hospitals and predicts ESI directly from triage-time information; on a clinician-adjudicated gold set, KATE achieved 75.9% accuracy versus 59.8% for nurses, and at the ESI 2 versus ESI 3 boundary it reached 80.0% versus 41.4% for nurses (Ivanov et al., 2020). Another line of work uses a deep attention model over chief complaint, past medical history, medication list, and nursing assessment, achieving an AUC of approximately 88% for identifying resource-intensive patients and an accuracy of approximately 44% for exact resource-category prediction (Gligorijevic et al., 2018).

Multimodal fusion is especially active. A late-fusion model for 5-level ESI predicts from vitals through XGBoost and from chief complaint through Bio_ClinicalBERT, then combines the two softmax vectors with logistic regression; when trained only on adults and evaluated zero-shot on children, symmetric modality dropout at 30–40% improved pediatric Quadratic Weighted Kappa from 0.331 to 0.351 while slightly improving adult performance (Yang et al., 10 Apr 2026). In a different tabular direction, an adaptive simulated annealing–optimized CatBoost model for merged ESI classes reached 83.3% accuracy, 83.2% precision, 83.3% recall, and 83.2% F1 (Ahmed et al., 2022).

Graph representations recast triage as node classification on a patient-similarity network. In one such formulation, each patient visit is a node, edges encode cosine or distance-based similarity, and GraphSAGE outperforms GCN, GATv2, SVM, and KNN when graph density is neither too sparse nor too dense (Defilippo et al., 2024). LLM systems further extend the design space: a KTAS-oriented multi-agent CDSS built on Llama-3-70B, CrewAI, LangChain, and RxNorm separates triage nurse, emergency physician, pharmacist, and coordinator roles, and in a 43-case evaluation the multi-agent system achieved perfect accuracy for KTAS 1 and stronger triage decisiveness than a single-agent baseline (Han et al., 2024). A French proof-of-concept comparing NLP, LLM, and JEPA architectures found the LLM model URGENTIAPARSE to be strongest for FRENCH triage, with weighted kappa 0.800, Spearman 0.802, and exact agreement 0.900 against expert gold-standard labels (Lansiaux et al., 1 Jul 2025).

4. Prediction targets and evaluation regimes

Although acuity assignment remains central, ED-triage research now targets a wider set of outcomes. These include exact ESI level, merged urgency strata, resource categories, hospital admission, disease-specific screening, self-harm surveillance, vaccine-safety surveillance, and department routing (Gligorijevic et al., 2018, Ivanov et al., 2022, Lu, 2023, Chen et al., 1 Jun 2026, Cheng et al., 30 Jul 2025). This diversification explains the broad metric landscape.

For ordered acuity labels, Quadratic Weighted Kappa is prominent because it penalizes larger ordinal errors more heavily. In adult-to-pediatric ESI transfer, QWK is the main measure, supplemented by accuracy, balanced accuracy, and macro F1 (Yang et al., 10 Apr 2026). For disease screening at triage, discrimination metrics dominate: the sepsis model KATE Sepsis, trained on 512,949 adult encounters from 16 hospitals, achieved AUC 0.9423 with sensitivity 71.09% and specificity 94.81%, compared with 0.6826, 40.8%, and 95.72% for standard SIRS-plus-infection screening (Ivanov et al., 2022). For high-specificity medical directives, TriNet reported PPV 0.86 and TNR 0.97 for pneumonia and PPV 0.93 and TNR 0.97 for UTI, deliberately accepting lower TPR to reduce over-testing (Lu, 2023).

Surveillance tasks use still different criteria. An evidence-augmented self-harm classifier from ED triage notes achieved AUPRC 0.887 ± 0.016 in internal validation and 0.884 ± 0.012 in external validation, with prospective AUPRC 0.881 ± 0.008 at the development site and 0.879 ± 0.012 and 0.816 ± 0.015 at two external sites without site-specific retraining; the same pipeline identified the primary self-harm method with 95% accuracy (Chen et al., 1 Jun 2026). For vaccine-safety signal detection from Victorian ED triage notes, an active-learning transformer classifier improved from F1 0.295 in Round 1 to 0.972 in Round 4 on a deployment-environment evaluation set, with precision 0.955 and recall 0.991, clearly surpassing pattern matching (Khademi et al., 24 Jul 2025). These results indicate that “ED-triage” is now evaluated both as a clinical classification problem and as a high-throughput surveillance substrate.

5. Generalization, lexical heterogeneity, and demographic bias

Generalization is a recurring technical fault line. One form is demographic transfer: adult-trained triage models can misread pediatric physiology because normal heart rate, blood pressure, and respiratory rate vary with age and development (Yang et al., 10 Apr 2026). Another is institutional transfer through language. In cross-hospital self-harm modeling, shared unigram usage between hospitals had cosine similarity 0.822, but bigrams fell to 0.628 and trigrams to 0.473, and a previously reported TF–IDF plus gradient boosting model dropped from AUPRC 0.85 within-site to 0.78 when transferred across hospitals (Chen et al., 1 Jun 2026). The transferable evidence-augmented self-harm pipeline is notable precisely because it reduces that degradation without site-specific retraining (Chen et al., 1 Jun 2026).

The literature also documents fairness-relevant disparities in triage itself. In a large adult ED analysis using propensity-score matching, men were more likely than women to be triaged as urgent and admitted, and had shorter stays; Black, Hispanic/Latino, Asian, and Other race patients were less likely than White patients to be triaged as urgent, and Black patients were more likely than White patients to experience long stays before admission (Coggan et al., 28 Mar 2025). These findings matter for AI because models trained on historical triage labels may inherit the same structural inequities even when race is excluded from the feature set. This suggests that subgroup calibration, transport evaluation, and error auditing are not ancillary but constitutive requirements for ED-triage systems.

A related issue is modality imbalance. In multimodal EHR models, structured tabular data often dominate training, leading to “modality collapse” in which free text is effectively ignored; symmetric modality dropout was proposed precisely to counter that tendency and improve cross-age transfer (Yang et al., 10 Apr 2026). The broader implication is that generalization failure often reflects representational imbalance rather than mere model capacity.

6. Operational integration, dynamic triage, and future directions

ED-triage systems are increasingly designed for direct workflow participation. One early integrated system combined a fuzzy inference triage score, fuzzy Q-learning for physician-time estimation, and a genetic-algorithm queue optimizer; in simulation on a 17-patient queue, average waiting time fell from 169 minutes to 121 minutes, a reduction of 48 minutes, while urgent patients were prioritized (0810.3671). More recent work operationalizes triage as test triggering or service routing: TriNet silently orders confirmatory tests at triage for high-confidence pneumonia and UTI cases (Lu, 2023), and a collaborative multi-agent system for department recommendation reaches 89.2% accuracy in primary-department classification and 73.9% in secondary-department classification after four rounds of patient interaction across 9 primary and 62 secondary departments (Cheng et al., 30 Jul 2025).

Dynamic updating is another frontier. A Bayesian sequential model for time to first urination fits the population cumulative admission curve with integrated squared error 0.002, RMSE 0.003, KS 0.008, and coverage 0.98; at the patient level, age/sex adjustment improves $\mathrm{AUC}(t)$ from 0.50 to 0.70, after which Platt recalibration refines probability scaling without changing discrimination (Senda et al., 17 Oct 2025). This is a concrete move away from static triage snapshots toward time-consistent risk trajectories.

Across the literature, however, retrospective design, single-center or regionally bounded datasets, label noise, and limited prospective validation remain pervasive limitations. Stated future directions include early-fusion and joint-fusion architectures, targeted debiasing and variable analysis, external and prospective validation, additional modalities such as imaging or physiological waveforms, federated or privacy-preserving learning, integration with curated knowledge bases rather than general web search, and explicit fairness auditing across demographic subgroups (Yang et al., 10 Apr 2026, Chen et al., 1 Jun 2026, Han et al., 2024, Senda et al., 17 Oct 2025). A plausible implication is that the next phase of ED-triage research will be defined less by isolated accuracy gains than by stable transport across hospitals, calibrated dynamic updating, and safe embedding into real emergency workflows.