Automated Triage Models

Updated 21 April 2026

Automated triage models are computational systems that integrate rule‐based logic, Bayesian methods, and machine learning to prioritize tasks efficiently across diverse domains.
They leverage graph-based inference, probabilistic reasoning, and neural network frameworks to deliver scalable, auditable, and rapid decision support.
Applications span medical emergency triage, software bug prioritization, and incident management, highlighting their adaptability and practical impact.

Automated triage models are computational systems designed to support or replace human experts in prioritizing tasks—most notably, patients in medical and emergency contexts, but also bugs in software engineering, incident tickets, and backlog management. These models leverage a range of algorithmic paradigms, from hard-coded decision rules and fuzzy logic to inductive machine learning, graph methods, and multi-agent collaboration. The breadth of application and methodological diversity reflect both the universal necessity of efficient, risk-aware prioritization and the distinct constraints of each domain.

1. Foundations and Computational Models

Core to automated triage is the formalization of prioritization rules and clinical or operational expertise as an inference engine over a structured decision process. A canonical example is the "babylon check" system, which models the triage process as a directed graph $G$ with six node types: questions ( $Q$ ), answers ( $A$ ), logical triggers ( $T$ ), worry-point score nodes ( $S$ ), exempt nodes ( $E$ ), and outcome recommendations ( $O$ ), each with assigned priorities (Middleton et al., 2016). Red-flag detection and weighted scoring are used in parallel to select the optimal outcome:

Red-flag triggers $t_j$ combine selected answers, patient demographics, and worry score thresholds to generate urgent recommendations.
Weighted scoring sums real-valued weights $w_i$ over selected answers; discrete outcome tiers are indexed by ranges of the total score $R$ .

Formally: $Q$ 0

$Q$ 1

Such graph-based logic enables deterministic, fully auditable triage recommendations, supporting clinical safety and rapid execution.

In graph neural network (GNN) triage, patients become nodes in a feature-rich graph, linked by similarity (e.g., cosine similarity or proximity in the feature space). Triage is then posed as a node-classification problem with multiclass cross-entropy loss (Defilippo et al., 2024, Guzzi et al., 2023). This inductive formulation captures higher-order relationships and supports robust generalization.

2. Algorithmic Paradigms: Rule-based, Statistical, and AI-driven Approaches

Automated triage approaches fall into several broad algorithmic categories:

1. Deterministic, Expert-Designed Logic:

Systems like babylon check encode clinician-derived triggers and weights into a static, interpretable graph (Middleton et al., 2016). Fuzzy inference systems similarly employ expert-tuned membership functions and rule bases to produce continuous or categorical triage scores, as well as adaptive modules for waiting-time estimation through fuzzy Q-learning (0810.3671).

2. Probabilistic and Bayesian Reasoning:

Rule-based Bayesian networks enable robust, uncertainty-tolerant triage, as demonstrated in DARPA Triage Challenge (DTC) field robotics (Rusiecki et al., 21 Dec 2025). Expert-defined Conditional Probability Tables implement physiology-respecting causality, permitting exact inference even with missing or noisy data. Noisy-OR parameterizations, marginalization of missing features, and soft evidence fusion accommodate the realities of sensor dropouts and imperfect computer vision (Hughes et al., 9 Dec 2025).

3. Machine Learning and Inductive Models:

Gradient-boosted decision trees (LightGBM, CatBoost) achieve state-of-the-art accuracy in sparse, dynamic triage interviews, particularly when interview completeness varies. TabTransformer architectures further improve robustness to missingness via self-attention contextualization, although with prohibitive computational cost for large datasets (Krylova et al., 16 Apr 2025). Neural-network models such as modular CNNs for medical note/text, feed-forward classifiers over structured features, and joint architectures for multimodal screening have demonstrated superior positive predictive value in targeted screening applications (e.g., pneumonia and UTI at pediatric ED triage (Lu, 2023)).

Graph-based learning—leveraging patient similarity graphs and models such as GraphSAGE, GCN, and GATv2—outperforms both tabular baselines and non-graph deep models in multi-class triage prediction (Defilippo et al., 2024, Guzzi et al., 2023).

4. LLMs and Multi-Agent Systems:

Recent advances use instruction-tuned LLMs for end-to-end triage in multiple settings: bug assignment in software repositories, clinical event triage in EHR, and vulnerability prioritization in security workflows (Kiashemshaki et al., 28 Aug 2025, Wang et al., 18 Mar 2026, Haddad et al., 21 Oct 2025). LLMs have demonstrated strong shortlist (Top-10) accuracy—up to 0.99 in restricted settings—and moderate Cohen’s kappa agreement with expert decisions.

Multi-agent systems orchestrate specialist agents (e.g., data normalization, information-seeking, department assignment) to manage complexity and adapt to healthcare institutional heterogeneity (Cheng et al., 30 Jul 2025, Delikoyun et al., 17 Oct 2025). LLM-based multi-agent frameworks can both validate known biomarkers and discover novel risk features, supporting explainable and literature-grounded acute care triage (Delikoyun et al., 17 Oct 2025).

3. Data, Feature Engineering, and Workflow Integration

Automated triage requires comprehensive, high-integrity feature engineering:

Structured Data: Vitals, demographics, comorbidities, and medication history; scaled and imputed as necessary. Algorithms frequently use SMOTE+ENN to rebalance classes, and min–max normalization to scale features for neural or GNN processing (Defilippo et al., 2024, Guzzi et al., 2023).
Unstructured Data: Free-text chief complaints, nurse/patient dialogues, and electronic messages. Representations range from Doc2Vec and BioSentVec embeddings to FlauBERT or BERT-derived contextual encodings (Lansiaux et al., 1 Jul 2025).
Multi-modal Inputs: Clinical imaging (CXR), signal waveforms (radar for vitals), transcript/audio, and multi-view images for injury and mental-status classification. Onboard, edge-deployable VLMs and DINO-variant vision encoders are increasingly viable (Hughes et al., 9 Dec 2025, Li et al., 26 Aug 2025).

Workflow integration typically leverages EHR integration, mobile/web frontends, and robust configuration stores allowing non-developer editability. Question engines dynamically adapt the triage flow (Q → A → Q), and collaboration among agents or modules (both digital and robotic) delivers results through web APIs, UI dashboards, or real-time mesh networks for field response (Middleton et al., 2016, Hughes et al., 9 Dec 2025, Wang et al., 18 Mar 2026).

4. Evaluation, Metrics, and Clinical Performance

Performance metrics are tailored to domain and deployment requirements:

Accuracy, Precision, Recall, F1: Macro-averaged across triage levels or binary tasks (e.g., pneumonia/UTI screening) (Lu, 2023, Defilippo et al., 2024).
Positive Predictive Value (PPV), Negative Predictive Value (NPV): Especially important for conditions with asymmetric costs of missed detection or false positives (Lu, 2023).
Weighted/Unweighted Cohen’s Kappa: Agreement with expert adjudication, capturing ordinal structure in multi-level triage (Gatto et al., 19 Jan 2026, Lansiaux et al., 1 Jul 2025, Haddad et al., 21 Oct 2025).
Pairwise Ranking/Bradley-Terry Models: Casting urgency as a pairwise head-to-head sorting problem, as in inbox triage (Gatto et al., 19 Jan 2026).
Selective Prediction Metrics: For imaging, risk–coverage curves and area under these curves (AURC) quantify errors at different acceptance thresholds (Li et al., 26 Aug 2025).
Consultation Time, Auditability, Clinical Safety: Deployment studies report time savings, traceability, and no unsafe under-triaging (Middleton et al., 2016, Wang et al., 18 Mar 2026).

Table: Representative Results

Model/Domain	Metric	Value	Reference
Babylon Check	Triage accuracy	88.2–90.2%	(Middleton et al., 2016)
GraphSAGE (GNN)	Test accuracy	93.2% (cosine graph)	(Guzzi et al., 2023)
TriNet (CNN+NN)	PPV (UTI)	0.93	(Lu, 2023)
AT-CXR (CXR)	Full-coverage acc.	95.3% (LLM router)	(Li et al., 26 Aug 2025)
SCM Navigator	Sensitivity (SCM)	0.94	(Wang et al., 18 Mar 2026)
Pairwise LLM	Inbox Sorting Acc.	76% (UrgentReward-8B)	(Gatto et al., 19 Jan 2026)

5. Limitations, Bias, and Future Directions

Limitations include:

Heuristic and Stationary Assumptions: Pre-defined triggers and static worry weights may require retuning as guideline and population shifts occur. GNNs often assume similarity structure stationarity, which may break during epidemic surges (Middleton et al., 2016, Guzzi et al., 2023).
Sparse or Biased Data: ML models are sensitive to class imbalance, demographic skew, and incomplete recordation. Interpretability remains limited for deep and LLMs, despite SHAP and feature attribution efforts (Guzzi et al., 2023, Lansiaux et al., 1 Jul 2025).
Domain Transfer and Multimodality: Pediatric and specialty flows, language/dialect diversity, and multimodal integration (e.g., real-time image/vitals fusion) require further development.
Workflow Integration: Robust concurrency, explanation, and human-in-the-loop retention are necessary for safety and adoption (Wang et al., 18 Mar 2026).

Ongoing developments focus on:

Probabilistic and Bayesian updates for uncertainty management (Middleton et al., 2016).
Cohesive multi-agent and collaborative frameworks for complex, large-scale institutional heterogeneity (Cheng et al., 30 Jul 2025).
Inductive and continual learning pipelines to adapt to population and workflow change (Defilippo et al., 2024).
Model-agnostic interpretability and active learning with real-time human feedback.

6. Cross-Domain Applications and Generalization

Automated triage methodologies have expanded well beyond emergency medicine:

Bug Triaging: MDP-based policies (ADPTriage) for expert-task assignment optimize assignment accuracy and minimize fixing time by balancing immediate costs and future uncertainties (Jahanshahi et al., 2022). Instruction-tuned LLMs provide high-quality ranked developer shortlists for real-world issue assignment (Kiashemshaki et al., 28 Aug 2025).
Vulnerability Prioritization: LLMs, when paired with structured prompt design and reasoning chains, show moderate agreement with expert prioritization frameworks (e.g., SSVC) (Haddad et al., 21 Oct 2025).
Simulated Dialogue and Education: TriageSim constructs persona-conditioned simulated dialogue and audio from structured data, facilitating scaled research in conversational triage and assessment of speech-based LLM reasoning (Srirag et al., 2 Mar 2026).

The modeling frameworks and optimization strategies demonstrated in automated triage are readily adaptable to any scenario requiring fast, scalable, and error-averse prioritization over dynamic, heterogeneous task or incident streams.

Key references:

(Middleton et al., 2016, Defilippo et al., 2024, Guzzi et al., 2023, Lu, 2023, Krylova et al., 16 Apr 2025, Cheng et al., 30 Jul 2025, Lansiaux et al., 1 Jul 2025, Gatto et al., 19 Jan 2026, Hughes et al., 9 Dec 2025, Rusiecki et al., 21 Dec 2025, Jahanshahi et al., 2022, Li et al., 26 Aug 2025, Kiashemshaki et al., 28 Aug 2025, Delikoyun et al., 17 Oct 2025, Haddad et al., 21 Oct 2025, Srirag et al., 2 Mar 2026, 0810.3671, Marchiori et al., 2020, Wang et al., 18 Mar 2026, 1904.02726)