Triage: Prioritization under Scarcity

• Triage is a resource-aware decision process that assigns priority levels to cases when resources are insufficient for full evaluation, balancing urgency with available capacity.
• It is applied across domains—from emergency medicine and mass casualty incidents to algorithmic deferral and remanufacturing—using methods like graph neural networks, boosting, and reinforcement learning.
• Research emphasizes operational metrics such as waiting time, error asymmetry, and cost tradeoffs, framing triage as a decision-theoretic problem of selecting when and how to act under constraints.

Triage is a resource-constrained decision process for prioritizing cases, tasks, or units when full evaluation or immediate service for all arrivals is infeasible. In emergency medicine, it assigns urgency so that life-threatening cases receive immediate attention and scarce staff, equipment, beds, and time are allocated appropriately; in mass casualty incidents, it operationalizes a rapid rule-based sorting procedure; in algorithmic settings, it can mean deferring selected instances to humans, reordering queues, or allocating a limited compute budget across competing tasks; and in remanufacturing it denotes an agile, prioritized, intentionally less exact sorting strategy used under extreme arrival conditions (Defilippo et al., 2024, Hagiwara et al., 18 Nov 2025, Okati et al., 2021, Gavidel et al., 2018). Across these uses, triage is not exhaustive diagnosis or perfect ranking. It is a fast, resource-aware control layer that decides who or what should be handled first, who can wait, what additional information should be collected, and when a system should stop gathering evidence and commit to an action.

1. Core concept and decision logic

At its most general, triage appears whenever arrivals, uncertainty, and capacity limitations coincide. Emergency-department papers formulate it as rapid assignment of a severity or priority code; radiology work formulates it as queue prioritization of studies with time-sensitive findings; remanufacturing treats it as a complementary emergency-mode sorting policy when normal inspection cannot be sustained; and recent work in autonomous AI systems reuses the term for prospective selection, sequencing, and allocation under a finite budget (Thompson et al., 2023, Gavidel et al., 2018, Nazi et al., 13 May 2026). The shared structure is a constrained allocation problem: an item arrives, the decision-maker has incomplete information, resources are finite, and the objective is to reduce harmful delay or wasted effort.

The action space varies by domain, but it is consistently small and operationally consequential. In clinical triage, actions include assigning urgency categories, routing to telecare or face-to-face care, asking another question, or deferring to a human expert. In circular-economy triage, the choice is whether to continue disassembly or commit immediately to a circular-economy route such as reuse, repurpose, recycle, or disposal (Fox et al., 17 Dec 2025). In benchmarked LLM triage, the plan must jointly encode selection, sequencing, and per-problem allocation under a shared token budget (Nazi et al., 13 May 2026). This suggests a common abstraction: triage is a control policy over scarce resources, not merely a classifier over labels.

A second common feature is asymmetry of error. Medical under-triage can delay lifesaving care, while over-triage wastes scarce resources (Ahmed et al., 2022). In algorithmic deferral, assigning a case to the wrong handler can increase expected loss (Okati et al., 2021). In remanufacturing, failure to recognize extreme arrivals can produce delayed market response, overstocking, overproduction, or lost profit (Gavidel et al., 2018). Triage therefore depends not only on prediction quality, but also on the operational cost of delay, congestion, or misallocation.

2. Clinical forms, taxonomies, and outputs

Emergency-department triage is the most direct and widely represented form. One graph-based ED study frames the task as multiclass classification over four categories—Red, Orange, Yellow, and Green—using historical patient records to assign severity labels for queue management and resource allocation (Defilippo et al., 2024). Another e-triage study uses the Emergency Severity Index, a five-level scale in which level 1 is the most urgent and level 5 the least urgent, then collapses the labels into three clinically motivated classes: “most urgent” for levels 1–2, “urgent” for level 3, and “less urgent” for levels 4–5 (Ahmed et al., 2022). A French proof-of-concept study uses the six-level FRENCH scale as the primary gold standard, comparing AI predictions against consensus labels curated by senior emergency physicians and triage nurses (Lansiaux et al., 1 Jul 2025).

Mass casualty triage is more explicitly algorithmic. The START method—Simple Triage and Rapid Treatment—assigns four categories, Green, Yellow, Red, and Black, through a fixed sequence: ability to walk, breathing, respiratory rate, perfusion, and mental status (Hagiwara et al., 18 Nov 2025). The procedure is designed for disaster scenes and uses objective physiological criteria rather than subjective judgment. Its role is narrower than diagnosis: it standardizes immediate prioritization under severe time, staffing, and resource constraints.

Telemedical triage broadens the output beyond urgency labeling to routing and timing. The Artificial Intelligence Triage Engine evaluates an initial symptom description, asks a small number of targeted follow-up questions, and recommends the most appropriate point of care and time frame for treatment (Marchiori et al., 2020). The output may be immediate telecare contact, scheduled teleconsultation, or referral to a face-to-face provider. CT-based COVID-19 triage similarly separates two tasks: identification of likely COVID-19 studies so that potentially infected patients can be isolated early, and severity quantification so that severe cases can be highlighted for urgent management (Goncharov et al., 2020). In both cases, triage is an ordering and routing layer, not a substitute for definitive diagnosis.

Radiological computer-aided triage and notification makes the same logic explicit in queueing terms. A CADt device yields a binary output for each non-emergent image—AI-positive or AI-negative—while emergent cases retain highest priority (Thompson et al., 2023). The triage action is therefore not a diagnosis but a queue transformation: without CADt, non-emergent images are read FIFO; with CADt, AI-positive cases are moved ahead of AI-negative cases.

3. Computational methodologies for automated triage

Recent medical triage systems span graph learning, boosting, reinforcement learning, NLP, LLM hybrids, and knowledge-graph-based dialogue systems. A central line of work converts ED triage from independent-row tabular classification into graph-based node classification. In "Leveraging graph neural networks for supporting Automatic Triage of Patients" (Defilippo et al., 2024), each patient is a node, edges encode similarity computed from clinical features using cosine similarity, Euclidean distance, Manhattan distance, or Minkowski distance, and severity prediction is performed by GCN, GATv2Conv, or GraphSAGE. Historical ED records define the graph; a new patient is added as a new node; and the model is inductive, so inference for the new node does not require retraining the entire network. The study reports that GraphSAGE consistently performs best among the tested GNNs and surpasses SVM and KNN on the same preprocessed features. The related network-science paper uses a similar patient-similarity graph and likewise reports GraphSAGE as the best-performing architecture across graph types (Guzzi et al., 2023).

A different ED line emphasizes tabular boosting and hyperparameter optimization. The ASA-based e-triage framework uses retrospective EHR data from four large ED locations, feature selection, SMOTE, and tuning of XGBoost and CatBoost with simulated annealing, adaptive simulated annealing, or grid search (Ahmed et al., 2022). The reported best model is ASA-CaB with RF_RFE, reaching accuracy 83.3%, precision 83.2%, recall 83.3%, and F1 83.2%. SHAP analysis is used to examine the contribution of variables such as ED location ID, temperature, oxygen saturation, BMI, and age. The same paper is explicit that the system is intended as decision support for nurses rather than a stand-alone replacement.

Telemedicine triage introduces a broader pipeline. The Artificial Intelligence Triage Engine is trained on more than 900k German teleconsultation records collected over more than 7 years, extracts symptoms, diagnoses, and attributes such as negation, location, intensity, duration, and activity context, builds a triage-specific ontology from over one million medical expressions, and reasons over a knowledge graph with more than 1M nodes and about 23M edges (Marchiori et al., 2020). The system combines an NLP pipeline, a dynamic question-generation module, and a recommendation engine. It keeps waiting time between interactions below one second, aims to reach final recommendations in only a few exchanges, and emphasizes transparency by retrieving similar historical patients to justify its recommendations.

Sequential questioning has also been modeled directly as reinforcement learning. A Deep Q-Learning triage agent is trained on 1374 clinician-authored clinical vignettes, each associated with an average of 3.36 expert triage decisions in the Methods section and 3.8 in the abstract (Buchard et al., 2020). The state is a ternary evidence vector over 597 observable medical evidences, actions are ask or one of four terminal triage labels—Red, Yellow, Green, Blue—and the key novelty is a dynamic target for the ask action so that the model learns when to stop gathering information. The best DyQN configuration achieves appropriateness 0.85 and safety 0.93 on the test set, approximately matching the human proxy on those metrics, while also learning a stopping policy.

More recent ED comparisons bring LLMs directly into triage prediction. The seven-month retrospective proof-of-concept at Roger Salengro Hospital compares TRIAGEMASTER (NLP with Doc2Vec plus a feedforward network), URGENTIAPARSE (FlauBERT embeddings combined with XGBoost), and EMERGINET (JEPA with dual encoders and VICReg regularization) (Lansiaux et al., 1 Jul 2025). On the composite Z-score, URGENTIAPARSE scores 2.514, EMERGINET 0.438, TRIAGEMASTER -3.511, and nurse triage -4.343. URGENTIAPARSE also attains F1 micro 0.899, F1 macro 0.894, exact agreement 0.900, and near agreement 0.928 on FRENCH prediction. At the same time, the paper notes that its learning curves suggest overfitting despite strong final metrics, so the result is explicitly proof-of-concept rather than deployment-ready.

4. Formal optimization, deferral, and stopping

Triage is not only a classification problem; it is also a decision-theoretic and operational optimization problem. The clearest formalization appears in algorithmic triage, where a model is allowed to defer some instances to human experts. The objective is

$$L(\pi,m)=\mathbb{E}_{x,y,h}\Big[(1-\pi(x))\,\ell(m(x),y)+\pi(x)\,\ell(h,y)\Big],$$

under a triage budget $\mathbb{E}_x[\pi(x)] \le b$, where $\pi(x)=1$ means defer to the human and $\pi(x)=0$ means predict with the model (Okati et al., 2021). A central result is that, for any fixed model, the optimal triage policy is a deterministic threshold rule based on the per-instance difference between the model’s expected loss and the human’s expected loss. The same paper proves that the model trained for full automation can be suboptimal under triage, because the triage-trained model need only be accurate on the subset it retains. This sharply distinguishes triage from a standard reject option.

The reinforcement-learning triage formulation addresses a different control question: when to continue information acquisition and when to stop. In the Deep Q-Learning framework, every episode starts with only one piece of evidence revealed; the ask action reveals one additional missing evidence; and the terminal actions commit to urgency labels (Buchard et al., 2020). The unusual reward design is counterfactual, with triage-action reward tied to the distribution of clinician labels for the vignette. The system therefore learns a stopping criterion from expert behavior, not from a hard-coded tree.

Queueing-theoretic work makes triage operational in terms of waiting time rather than only label accuracy. For CADt systems, the central performance metric is

$$\delta W_{D} \equiv W^{\text{CADt}}_{D} - W^{\text{no-CADt}}_{D},$$

the change in mean waiting time for diseased patients before image review (Thompson et al., 2023). A negative $\delta W_D$ means the triage device saves time for diseased patients. The paper studies four workflow models with one or two radiologists, emergent and non-emergent cases, and equal or unequal reading times, and concludes quantitatively that CADt is more effective in busy, short-staffed settings.

Circular-economy triage provides a further optimization perspective. There, the state is augmented with disassembly history, $\tilde v=(v,\tau)$, so that the decision process satisfies the Markov property (Fox et al., 17 Dec 2025). At each state, the system compares the value of continuing disassembly against committing to a circular-economy route, with recursive value

$$U(\tilde v)=\max\Big[\max_{k\in\mathcal K_v}(\cdots),\ \max_{e\in \mathrm{out}(\tilde v)}[-c_e+U(\tilde v')]\Big].$$

Although far removed from emergency medicine, the structure is recognizably triage-like: act now or gather more access and information at additional cost.

5. Evaluation, fairness, and synthetic benchmark construction

A persistent difficulty in triage research is that the relevant data are rare, private, incomplete, or expensive to curate. Mass casualty incidents occur infrequently, and scene documentation can be incomplete or inconsistent; nurse–patient triage dialogue is constrained by privacy and regulation; and retrospective ED labels may reflect historical noise or clinician heterogeneity (Hagiwara et al., 18 Nov 2025, Srirag et al., 2 Mar 2026, Defilippo et al., 2024). This has led to an increasingly methodological literature on evaluation data themselves.

Syn-STARTS addresses the scarcity of START-triage benchmarks by generating synthetic adult MCI cases with an LLM and a deterministic validation pipeline (Hagiwara et al., 18 Nov 2025). Each case contains a ground-truth triage tag, structured vitals, and a narrative patient description, and only cases passing START consistency, medical plausibility, and narrative consistency checks are retained. The paper constructs 2,000 validated cases, 500 per tag, and shows two forms of benchmark fidelity: in a blinded A/B test with 3 medical doctors, experts could not reliably distinguish Syn-STARTS cases from the manually curated TRIAGE-adult benchmark; and model accuracies on the synthetic matched-distribution sets correlate strongly with TRIAGE-adult, with Pearson’s $r = 0.92$. It also shows that increasing dataset scale from $n=12$ to $\mathbb{E}_x[\pi(x)] \le b$0 lowers the standard deviation of accuracy estimates.

TriageSim tackles a different evaluation bottleneck: spoken triage interaction (Srirag et al., 2 Mar 2026). Starting from structured EHRs and pedagogical resources, it generates persona-conditioned, multi-turn nurse–patient dialogues under ATS or ESI policy prompts, then synthesizes speech with accent variation and ED-like ambient noise. The corpus contains 814 conversations, 280K tokens, and 26.14 hours of audio. Automated and manual evaluations are split by fidelity type: patient disfluency rises monotonically with the intended control level, the average WER is 10.8, and a clinician with 5+ years of emergency experience rates a random subset of 50 conversations, yielding chief-complaint similarity 0.83 and red-flag precision 0.94, recall 0.96, and F1 0.94. Downstream conversational triage classification remains modest, with weighted Cohen’s $\mathbb{E}_x[\pi(x)] \le b$1 between 0.23 and 0.33 depending on modality and policy, which the paper interprets as evidence that the main difficulty is clinical reasoning over conversation rather than transcription noise.

Fairness research has similarly shifted attention from the final triage label to the full process path. Using MIMICEL, a MIMIC-IV-ED-derived event log with 7,422,277 rows and 413,893 cases, a process-mining study evaluates four business process outcomes—time, re-do, deviation, and decision—across age, gender, race, language, and insurance (Andreswari et al., 16 Jan 2026). Kruskal–Wallis and chi-square tests, along with $\mathbb{E}_x[\pi(x)] \le b$2 and Cramér’s $\mathbb{E}_x[\pi(x)] \le b$3, show that time effects are mostly negligible except small age effects at acuity levels 4 and 5, whereas decision disparities are much stronger, especially for insurance and language. The most striking reported value is insurance at ESI level 5 with $\mathbb{E}_x[\pi(x)] \le b$4, a very large effect. The paper maps these outcomes to distributive, procedural, and interactional justice, thereby treating triage fairness as a process-quality problem rather than only an output-label problem.

6. Extensions beyond medicine and the widened meaning of triage

The term has been extended well beyond emergency care, but the extensions preserve the same resource-aware logic. In remanufacturing, triage is defined as an agile, prioritized, intentionally less exact sorting strategy for extreme value scenarios in which arrival rate exceeds remanufacturing capacity (Gavidel et al., 2018). The supporting statistical model uses Extreme Value Analysis, the Generalized Extreme Value distribution, and, in the valve case study over 476 operational days, yields MLE estimates $\mathbb{E}_x[\pi(x)] \le b$5, $\mathbb{E}_x[\pi(x)] \le b$6, and $\mathbb{E}_x[\pi(x)] \le b$7, indicating Fréchet-type heavy-tailed behavior. Central-tendency models assign essentially zero probability to the largest observed arrivals, yet the data contain 9 cases out of 158 above 40 arrivals. Here triage is explicitly a complementary emergency-mode policy rather than routine inspection.

In fault-tolerant quantum computation, "Triage" names a dual-mode scheduler for real-time classical decoding under scarce decoder resources (Chen et al., 6 May 2026). The system uses a steady mode based on a weighted urgency and cost-efficiency heuristic,

$\mathbb{E}_x[\pi(x)] \le b$8

and an emergency mode that resolves the causal cone of synchronization-critical operations. Across benchmarks, it reports an average logical error-rate reduction of 52.6% compared to standard temporal parallelism. The analogy to medical triage is structural rather than semantic: a finite pool of decoders must be directed to the most urgent dependencies before stalls propagate.

A closely related abstraction appears in the LLM benchmark "TRIAGE: Evaluating Prospective Metacognitive Control in LLMs under Resource Constraints" (Nazi et al., 13 May 2026). There, a model receives a pool of problems and a shared token budget

$\mathbb{E}_x[\pi(x)] \le b$9

then commits to a single ordered plan. Performance is measured by the triage efficiency ratio

$\pi(x)=1$0

The paper finds that current models show substantial gaps in prospective metacognitive control, with $\pi(x)=1$1 negative for the large majority of configurations when budgets are binding at $\pi(x)=1$2. Here triage means portfolio-level self-allocation of compute before execution begins.

Other technical uses preserve the theme of adaptive escalation. In respiratory audio classification, TRIAGE—Tiered Retrieval and Inference for Audio with Gated Escalation—is a fully zero-shot framework that routes samples through progressively richer stages and finalizes easy cases early (Wang et al., 14 Apr 2026). It reports mean AUROC 0.744 across nine tasks and finalizes 46% of samples at the cheapest tier. In regression, TRIAGE—TRaining Identification Auditing of Good Examples—uses conformal predictive distributions to characterize training samples as under-estimated, over-estimated, or well-estimated through the score

$\pi(x)=1$3

thereby repurposing triage as data auditing and selection (Seedat et al., 2023). These usages are terminological extensions, but they retain the same core intuition: selective commitment under finite resources is itself a first-class problem.

Taken together, the contemporary literature treats triage as a general theory of constrained prioritization. In medicine, it remains anchored to urgency, safety, and care routing. In technical systems, it has become a wider design pattern for selection, escalation, deferral, and scheduling under scarcity. The convergent implication is not that all forms of triage are equivalent, but that they share a common operational grammar: limited capacity, heterogeneous urgency, incomplete information, and the need to act before full certainty is available.