Triage: Prioritization Under Scarce Resources

• Triage is the systematic prioritization of cases based on urgency and resource constraints, employing protocols like START to maximize overall outcomes.
• It applies to diverse domains such as mass casualty incidents, emergency departments, telemedicine, imaging, and even circular economy routing.
• Recent research leverages synthetic benchmarks, machine learning, and explainability methods to enhance decision support, calibration, and fairness.

Searching arXiv for the main Syn-STARTS paper and closely related triage benchmarks to ground the article in recent literature. arXiv_search(query="(Hagiwara et al., 18 Nov 2025) OR Syn-STARTS Synthesized START Triage Scenario Generation Framework for Scalable LLM Evaluation", max_results=5) arXiv_search(query="Syn-STARTS: Synthesized START Triage Scenario Generation Framework for Scalable LLM Evaluation", max_results=5) Triage is the structured prioritization of cases under scarcity, usually to determine urgency, routing, or resource allocation before definitive diagnosis. In mass casualty incidents it is defined by the need to maximize total survival under severe time and resource constraints; in emergency departments it assigns acuity and queue priority; in telemedicine it recommends point of care and time frame; in imaging it reorders worklists; and in recent AI literature it has also become a formal framework for multimodal referral planning, conversational simulation, fairness auditing, and resource-bounded decision allocation (Hagiwara et al., 18 Nov 2025, Marchiori et al., 2020, Mosquera et al., 2020, He et al., 18 Mar 2026, Nazi et al., 13 May 2026). Across these settings, a recurrent theme is that triage is primarily a problem of risk assessment and routing rather than full diagnosis, with high asymmetry between under-triage and over-triage and a correspondingly strong demand for calibrated, explainable, and operationally robust decision systems (Srirag et al., 2 Mar 2026, Jang et al., 8 Jun 2026).

1. Operational definitions and domains of use

In disaster medicine, triage is a rapid bedside algorithm for prioritizing victims when there are more patients than can be treated immediately with available resources. The START method is described as “the most widely adopted global standard method for triage” in mass casualty incidents and is explicitly designed to minimize subjective judgment by relying on respirations, circulation, and consciousness (Hagiwara et al., 18 Nov 2025). In emergency departments, triage is a front-door process that assigns an urgency code or severity index and thereby shapes queue management, treatment timing, and resource allocation; the Italian ED literature summarized in graph-based triage work describes four phases—immediate “on the door” evaluation, subjective and objective evaluation, triage decision, and re-evaluation—while related work models four urgency classes Red, Orange, Yellow, and Green as a supervised classification problem (Defilippo et al., 2024, Guzzi et al., 2023).

Recent work broadens this clinical definition without discarding its core logic. In telemedicine, the AI Triage Engine is designed to determine the appropriate point of care and time frame through adaptive questioning over symptom narratives, previous illnesses, and demographic information, thereby operationalizing triage as remote care routing rather than bedside acuity tagging (Marchiori et al., 2020). In dental outpatient workflows, triage is formulated as an entry-point routing decision from chief complaint plus panoramic radiograph to a complete referral plan across eight specialty domains and twenty-two fine-grained treatment labels, making it explicitly hierarchical and multi-label (He et al., 18 Mar 2026). In conversational emergency simulation, triage is framed as a brief spoken interaction in which a nurse elicits symptoms, checks red flags and vital signs, and assigns an acuity level under ATS or ESI protocols; that work stresses that triage is risk assessment, not diagnosis (Srirag et al., 2 Mar 2026).

Imaging triage translates the same idea into prioritization of studies rather than patients. Chest radiograph triage systems flag abnormal studies and reorder radiology worklists; CT-based COVID-19 triage separates identification of likely infected studies from severity quantification so that radiologists can prioritize isolation-sensitive and high-severity cases (Mosquera et al., 2020, Goncharov et al., 2020). A plausible implication is that “triage” now names a family of routing problems in which the object being prioritized may be a patient, a study, a conversation, or a downstream referral plan, but the formal concern remains the same: bounded resources and unequal cost of delay.

2. Rule-based structures and hierarchical label systems

The canonical example of formal triage logic is START. Its decision sequence is rigid: if the patient can walk, tag as Green; if not breathing, open the airway and tag Red if breathing starts or Black if it does not; if breathing, assign Red for respiratory rate over 30; if respiration is not critical, assign Red for capillary refill over 2 seconds or absent radial pulse; if perfusion is adequate, assign Red if the patient cannot follow simple commands and Yellow otherwise (Hagiwara et al., 18 Nov 2025). This makes START unusually amenable to algorithmic operationalization because the label is a deterministic function of a small set of observations.

Other systems are less minimal but still explicitly protocolized. The Emergency Severity Index first checks for immediate life-saving intervention, then high-risk features or severe pain, and then estimates resource needs and danger-zone vital signs across five levels (Andreswari et al., 16 Jan 2026). The FRENCH scale is described as a six-level urgency stratification system intended for consistency, reproducibility, and standardization across French emergency services (Lansiaux et al., 1 Jul 2025). These schemes differ in granularity and clinical scope, but they share the feature that triage categories are not free-form narrative judgments; they are bounded label systems tied to treatment urgency, queue order, or referral destination.

Hierarchical triage adds another layer. Dental-TriageBench organizes triage as eight coarse specialty domains—such as Periodontology, Endodontics, and OMFS—with twenty-two fine-grained labels nested beneath them, so that triage becomes a structured referral plan rather than a single urgency class (He et al., 18 Mar 2026). This matters because multi-problem patients require simultaneous coverage of several domains. The benchmark reports an average of 2.60 positive domains per case and only 21.1% single-domain cases, showing that realistic triage often involves concurrent routing decisions rather than a single winner-take-all classification (He et al., 18 Mar 2026).

3. Data scarcity, synthetic benchmarks, and multimodal corpora

A central difficulty in triage research is the scarcity or inaccessibility of directly usable data. Mass casualty incidents are infrequent, and the START study argues that sufficient real-world records are difficult to accumulate at the scene; as a result, benchmark construction has increasingly turned toward synthetic generation (Hagiwara et al., 18 Nov 2025). Syn-STARTS addresses this by generating structured JSON cases with a ground-truth tag, a vitals_info object, and a patient_description, then validating every candidate through START consistency, medical plausibility, and narrative consistency (Hagiwara et al., 18 Nov 2025). With $N = 500$ per tag, the framework constructs 2,000 validated cases. Clinicians in a blinded A/B test could not reliably distinguish these synthetic narratives from manually curated adult START cases, and model accuracies on matched synthetic versus manual datasets were strongly correlated with Pearson’s $r = 0.92, p < 0.01$ (Hagiwara et al., 18 Nov 2025). The paper also reports that larger synthetic datasets reduced variance in estimated LLM accuracy, for example GPT-4 from $0.78 \pm 0.058$ at $n=12$ to $0.82 \pm 0.018$ at $n=200$ (Hagiwara et al., 18 Nov 2025).

Synthetic dialogue generation extends this logic into spoken triage. TriageSim converts structured EHR and pedagogical scenarios into 814 persona-conditioned nurse–patient conversations with aligned audio, totaling 26.14 hours, and evaluates them for linguistic, behavioural, acoustic, and medical fidelity (Srirag et al., 2 Mar 2026). The generated corpus yields weighted Cohen’s $\kappa$ values of 0.31 and 0.23 on ATS and ESI from synthetic text, 0.33 and 0.24 from ASR transcripts, and 0.28 and 0.27 from direct audio, while the synthetic audio itself has an overall WER of 10.8 (Srirag et al., 2 Mar 2026). Chief complaint reconstruction from a random subset of 50 conversations reached mean cosine similarity 0.83, and nurse red-flag logging achieved precision 0.94, recall 0.96, and F1 0.94 against expert annotation (Srirag et al., 2 Mar 2026).

Multimodal benchmarking reveals a different dimension of scarcity: authentic aligned image-text triage data. Dental-TriageBench provides 246 de-identified cases with chief complaint, OPG, hierarchical labels, and expert-authored golden reasoning trajectories (He et al., 18 Mar 2026). It shows that both modalities are required and that models fail especially on multi-domain cases, where they tend to output overly narrow referral sets with omission-heavy errors (He et al., 18 Mar 2026). This suggests that benchmark design for triage cannot be reduced to flat classification accuracy on isolated labels; it must preserve concurrency, hierarchy, and modality interaction.

4. Computational approaches to triage decision support

Early automation focused on explicit control logic. A South African accident and emergency system used a fuzzy inference system to compute a triage score from vital signs, consciousness, and pain, a fuzzy Q-learning module to estimate patient–doctor time, and a genetic algorithm to optimize queue order (0810.3671). In simulation, average waiting time decreased from about 169 minutes to 121 minutes, a reduction of 48 minutes, with under-triage under 2% and over-triage within accepted limits (0810.3671). This line of work treated triage not only as categorization but as a scheduling problem over urgency-weighted waiting time.

A second line learns triage directly from expert behaviour. Deep Q-Learning on 1,374 curated clinical vignettes learned both when to stop asking questions and which triage action to commit to, yielding safe decisions in 94% of cases and matching expert decisions in 85% of cases (Buchard et al., 2020). The main DyQN OR-query agent achieved appropriateness 0.85, safety 0.93, and an average of 13.34 questions on the test set, and it asked substantially more questions on unseen vignettes than on training vignettes, indicating novelty-sensitive behaviour (Buchard et al., 2020).

Graph-based methods recast triage as node classification on patient similarity networks. On a 6,551-row ED dataset with 16 features, graphs were constructed using cosine, Euclidean, Manhattan, and Minkowski similarity criteria, and GraphSAGE on a cosine similarity graph with threshold 0.95 performed best among the tested GNNs, outperforming SVM and KNN baselines on tabular data (Defilippo et al., 2024). Related work argues that such models make patient prioritization less dependent on isolated rows and more dependent on similarity to clinically comparable cases (Guzzi et al., 2023). In the telemedicine setting, a larger-scale system built from more than 900,000 German-language teleconsultation records combined NLP, ontology construction, a knowledge graph with roughly 23 million edges, adaptive question generation, and recommendation models, reaching F1 scores of 87.5% for high-risk, 74.0% for medium-risk, and 90.4% for low-risk predictions and operating as a certified system at Medgate (Marchiori et al., 2020).

Recent comparisons increasingly favour transformer-based architectures. In a 657-patient French ED proof-of-concept, URGENTIAPARSE, a FlauBERT-based LLM system combined with XGBoost, obtained a composite score of 2.514 on FRENCH triage prediction, compared with 0.438 for the JEPA-based EMERGINET, $-3.511$ for the NLP-based TRIAGEMASTER, and $-4.343$ for nurse triage against the expert gold-standard (Lansiaux et al., 1 Jul 2025). Imaging triage shows a parallel pattern. The chest x-ray system TRx achieved AUROC 0.75 on an external dataset and 0.87 on a local dataset for abnormal study detection, with local sensitivity 86% and specificity 88%, using a late fusion of segmentation, detection, and classification networks over four radiographic signs (Mosquera et al., 2020). CT-based COVID-19 triage formalized identification and severity as joint tasks; the Multitask-Spatial-1 model achieved ROC AUC 0.93 for COVID versus all others and Spearman correlation 0.97 for severity quantification (Goncharov et al., 2020).

5. Evaluation, explainability, and fairness

Triage evaluation is not monolithic. Emergency triage prediction studies commonly report F1-score, weighted kappa, Spearman correlation, MAE, and RMSE, explicitly treating urgency levels as ordinal rather than nominal labels (Lansiaux et al., 1 Jul 2025). Early warning work over irregularly sampled medical time series shifts emphasis toward AUPRC, Brier score, and Expected Calibration Error, because triage under resource constraints depends on calibrated probabilities and cross-patient comparability rather than only class accuracy (Jang et al., 8 Jun 2026). The TRIAGE framework for ISMTS argues that one-sided chain-of-thought induces “risk polarization,” collapsing graded risk into near-deterministic answers, and shows that dialectical reasoning over competing outcomes improves average AUPRC by 3.3%, reduces calibration error by 81%, and improves reasoning quality by 20% relative to competitive baselines (Jang et al., 8 Jun 2026).

Explainability likewise varies by modality. TRx provides per-finding heatmaps and a unified heatmap over chest radiographs, explicitly aiming at clinical implementation and second-reader use (Mosquera et al., 2020). URGENTIAPARSE couples FlauBERT with XGBoost and uses SHAP to trace predictions back to specific textual and structured features (Lansiaux et al., 1 Jul 2025). TRIAGE for ISMTS produces outcome-specific rationales rather than post-hoc feature attributions, and its rationales score higher than post-hoc explanations generated from a strong time-series baseline (Jang et al., 8 Jun 2026). A plausible implication is that triage explanation is moving from saliency-only paradigms toward rationales that are explicitly auditable against protocols or expert reasoning trajectories.

Fairness analysis reveals that accuracy alone is insufficient. Using the MIMICEL event log derived from MIMIC-IV ED, process-mining work evaluates time, re-do, deviation, and decision as business process outcomes across age, gender, race, language, and insurance (Andreswari et al., 16 Jan 2026). Within the same ESI acuity levels, race and gender effects are generally negligible or small, whereas insurance and language show medium to large effects for deviation and discharge decision, with especially strong associations in ESI 1–3 and very large insurance effects on decision at ESI 5 (Andreswari et al., 16 Jan 2026). Dental-TriageBench adds a different safety lens: models exhibit omission-heavy errors and overly narrow referral sets, whereas human triage tends toward slightly higher over-referral but lower omission risk (He et al., 18 Mar 2026). Together, these findings challenge a common misconception that a triage model is adequate once its overall agreement is high; in practice, omission patterns, calibration, and group disparities are integral parts of safety assessment.

6. Limits, controversies, and emerging extensions

Current triage AI remains narrow relative to real deployment. Syn-STARTS evaluates only adherence to START for static, isolated cases and does not model incident-level reasoning, multi-patient coordination, time evolution, or multimodal evidence; the authors are explicit that START correctness is not patient-outcome correctness (Hagiwara et al., 18 Nov 2025). TriageSim is synthetic, uses a single-clinician medical fidelity evaluation, and shows only modest agreement for conversational acuity classification, implying that clinical reasoning from noisy dialogue remains difficult even when ASR is not the primary bottleneck (Srirag et al., 2 Mar 2026). The French ED comparison reports clear overfitting in the learning curves of both the LLM and JEPA systems and emphasizes the need for multicentre prospective validation before workflow integration (Lansiaux et al., 1 Jul 2025).

At the same time, the term has begun to migrate beyond bedside medicine. Circular economy triage defines product end-of-life routing as a stop-versus-continue decision on a state-augmented disassembly graph, where the choice is between further disassembly and committing to reuse, repurpose, recycle, or disposal under condition-aware utility and operational constraints (Fox et al., 17 Dec 2025). A separate framework called TRIAGE evaluates whether LLMs can prospectively plan selection, sequencing, and per-problem token allocation across a task pool under a finite budget, thereby importing the logic of triage into metacognitive control under resource constraints (Nazi et al., 13 May 2026). This suggests that triage is increasingly functioning as a general formal language for priority setting under scarcity, with medicine remaining the central but no longer exclusive domain.

Across these literatures, the stable core of triage is not any single label set or algorithm. It is the requirement to make consequential routing decisions early, with incomplete information, limited resources, and asymmetric costs of error. Contemporary research increasingly treats that requirement as a benchmarkable computational object—sometimes through rule-based protocols such as START, sometimes through synthetic or multimodal corpora, and sometimes through learned systems that must be accurate, calibrated, explainable, and fair at the same time (Hagiwara et al., 18 Nov 2025, He et al., 18 Mar 2026, Jang et al., 8 Jun 2026).