PaperlayerView exact revision on arXiv

Source-linked AI summary

From Triage to Discharge: A Survey of NLP Tasks, Methods, and Open Challenges in the Emergency Department

Dipankar Srirag, Aditya Joshi, Salil Kanhere, Padmanesan Narasimhan

arXiv:2608.23627v1cs.CL

TL;DR

Emergency-department NLP research lacks a systematic ED-specific synthesis of task formulation, modelling choices, and evaluation practices. This survey analyses 46 papers across triage, diagnosis, and disposition, finding a shift toward pretrained and interactive systems alongside persistent limits in generalisability, clinical evaluation, and deployment.

  • Problem

    Existing surveys do not systematically examine how ED constraints shape NLP task formulation, modelling choices, and evaluation practices across the emergency-care workflow.

  • Method

    The survey synthesises 46 ED-NLP papers across triage, diagnosis, and disposition, covering tasks, modelling paradigms, evaluation practices, benchmarks, and shared tasks.

  • Results

    Across tasks, research shifts from task-specific neural architectures toward pretrained transformers and LLMs, with more open-ended, multimodal, process-oriented evaluation and interactive clinical tools.

  • Takeaways & Limitations

    ED-NLP research increasingly addresses summarisation, reasoning, and decision support in forms intended to reflect complex clinical workflows.

  • Takeaways & Limitations

    The corpus is skewed toward English- and Chinese-language research and chest-X-ray report generation, so generalisability to other languages, care settings, and imaging modalities requires caution.

Abstract

from arXiv · show

Emergency departments (EDs) operate under time pressure, generating multimodal data such as clinical conversations, triage notes, and discharge documents. Recent advances in natural language processing (NLP), particularly pretrained transformers and large language models, have created new opportunities to support language and time-intensive stages of emergency care. Yet existing surveys map clinical NLP across the broader hospital workflow or focus on specific tasks. This survey analyses 46 papers spanning the three phases of ED: triage, diagnosis, and disposition, covering tasks such as triage classification, clinical summarisation, automatic diagnosis, report generation, and discharge documentation. We examine modelling paradigms, evaluation practices, and emerging benchmarks and shared tasks. Across tasks, we identify common trends, including a shift from task-specific neural architectures to pretrained language models, growing interest in interactive clinical systems, and increasing attention to clinically grounded evaluation. Finally, we detail open challenges such as limited generalisability, noisy clinical inputs, and workflow constraints that inform future ED-NLP research.

1 Introduction

Emergency departments generate diverse language data across triage, diagnosis, and disposition, but existing surveys do not systematically connect ED-specific task formulation, modelling, and evaluation. This survey addresses that gap by analysing 46 papers across ED workflows and synthesising tasks, methods, benchmarks, and open challenges.

  • 1 Introduction: The survey frames ED-oriented NLP as a response to time pressure and documentation demands while examining whether evaluation reflects operational realities.The stated potential includes reducing documentation burden, surfacing clinical signals, and supporting consistent patient communication.
  • 1 Introduction: ED care spans triage, diagnosis and assessment, and disposition, producing structured and unstructured records from initial reports through final decisions.These records include triage notes, patient speech, consultation notes, radiology and laboratory reports, and discharge summaries.
  • 1 Introduction: Prior surveys rarely examine ED-specific NLP task formulation, modelling assumptions, and evaluation protocols systematically.Existing clinical surveys emphasise use cases and outcomes, while NLP surveys often focus on individual tasks or architectures.
  • 1 Introduction: The survey analyses 46 papers across ED phases, modelling paradigms, dataset language, input modality, evaluation, benchmarks, and shared tasks.Covered tasks include triage classification, clinical interaction summarisation, automatic diagnosis, radiology report generation, disposition prediction, and discharge documentation.

2 Research Trends

The reviewed ED-NLP literature concentrates on diagnosis and assessment, remains predominantly text-based and offline, and uses varied training strategies. Clinician evaluation and ED-specific deployment are uncommon, leaving real-world validation and broader generalisability as major gaps.

  • Clinical settings: 35/46 studies target diagnosis and assessment, compared with 5/46 on triage and 6 on disposition.The corpus contains 10 ED-native, 17 ED-anchored, and 19 ED-adjacent papers.
  • Clinical settings: 16/19 automatic-diagnosis papers are ED-adjacent, showing that much transferable work is not evaluated in emergency departments.This inclusion boundary conditions the survey’s estimates of research trends.
  • Clinical settings: 32/46 studies use English and 13/46 use Chinese, while other languages appear in only 3/46 studies.The imbalance largely reflects public clinical corpora, especially the MIMIC family, and indicates scarce multilingual datasets and evaluation resources.
  • Clinical settings: 43/46 studies use text, 22/46 include structured EHR variables, and only 4/46 incorporate medical imaging directly.Despite heterogeneous multimodal information, most current ED-NLP work remains text-centred.
  • Training strategies: Supervised learning appears in 13/46 studies, agent-based systems in 10/46, and transfer learning in 8/46, with no single dominant strategy.Agent-based and retrieval-augmented methods are concentrated in 2024 and 2025.
  • Evaluation paradigms: 24/46 studies rely on retrospective records and 21/46 on simulation or synthetic data, while only one study is prospective.Most findings therefore remain supported by offline rather than real-time clinical evaluation.
  • Evaluation paradigms: 11/46 papers supplement automated metrics with clinician or human judgement, and no paper reports deployment in an ED-specific workflow.The survey identifies real-world ED validation as the most concrete open challenge.

3 NLP Tasks in the ED

ED NLP tasks span triage, diagnosis, disposition, and radiology reporting, using increasingly interactive and pretrained approaches but facing evaluation and workflow constraints.

  • Triage Classification: Triage classification assigns ordered acuity levels, but most studies flatten the ordinal structure into standard multiclass prediction.Models use pretrained transformers, agent-based deliberation, or retrieved medical evidence from complaints, vital signs, and encounter data.
  • Triage Classification: Post-hoc EHR inputs can inflate triage accuracy because they include information unavailable at the moment of triage.This evaluation setup conflicts with real-world triage, which relies on limited initial information.
  • Clinical Interaction Summarisation: Clinical interaction summarisation converts multispeaker consultations into structured notes, commonly using SOAP sections and autoregressive generation.Approaches range from task-specific pipelines to pretrained, domain-adaptive, and prompt-based models.
  • Automatic Diagnosis: Automatic diagnosis interleaves symptom inquiries with a terminal diagnosis, framing the task as sequential decision-making over dialogue history.Studies use reinforcement learning, supervised, knowledge-grounded, and agent-based approaches.
  • Automatic Diagnosis: Diagnosis systems often optimise turn-efficient symptom acquisition, but turn-count rewards can mismatch clinician-time constraints in real ED workflows.Most evaluations use simulated patients or retrospective settings, with limited clinician-agreement and workflow assessment.
  • Radiology Report Generation: Radiology report generation for chest X-rays produces structured Findings and Impression sections, while evaluation combines text-generation, clinical-concept, and diagnostic metrics.The task is relevant to ED workflows because many chest X-ray examinations occur during ED stays.
  • Tasks in Patient Disposition: Disposition prediction combines structured variables, free-text complaints, visual signals, or linearised heterogeneous data to predict discharge, ward, or ICU outcomes.Discharge generation instead produces patient-facing instructions, with evaluation requiring calibration and subgroup sensitivity because error costs and patient needs vary.

4 Benchmarks and Shared Tasks

Benchmarks and shared tasks are expanding clinical NLP evaluation beyond exam-style recall toward workflow-oriented generation, reasoning, prediction, and documentation safety.

  • Benchmarks: Early clinical language-model benchmarks mainly measured factual recall through exam-style question answering rather than clinical decision-making or communication.Newer benchmarks include open-ended summarisation, diagnosis, and disposition tasks.
  • Benchmarks: ClinicBench aggregates datasets for clinical language understanding, reasoning, and generation, reflecting broader workflow-oriented evaluation.Other resources combine multimodal ED data for time-sensitive prediction or assess documentation errors and dialogue-to-note generation.
  • Shared Tasks: MEDIQA-Chat 2023 compared section-wise and full-note generation from doctor-patient conversations under common datasets and protocols.Submitted systems ranged from fine-tuned dialogue-aware transformers to retrieval-based models and in-context LLM approaches.
  • Shared Tasks: The Discharge Me! challenge evaluated brief hospital-course and discharge-instruction generation from EHR data using prompt-based, fine-tuned, and hybrid systems.Hybrid pipelines combined structured extraction with controlled generation, while submissions also explored dynamic input filtering.

5 Conclusion and Future Directions

The survey finds a shift toward pretrained transformers and LLMs, open-ended and process-oriented evaluation, and interactive clinical tools across ED-NLP tasks. It identifies generalisability, robustness, evaluation validity, and deployment as key open challenges.

  • ED-NLP research has shifted from task-specific neural architectures toward pretrained transformer models and LLMs.
  • Evaluation is increasingly open-ended, multimodal, and process-oriented, moving beyond exam-style question answering toward realistic clinical workflows.
  • Many systems are framed as interactive tools for summarisation, reasoning, and decision support rather than isolated prediction models.
  • Single-institution datasets, online health communities, and institution-specific documentation and workflows limit ecological validity and generalisability.
  • Real-world ED data remains noisy and incomplete because of transcription errors, time pressure, informal descriptions, and language barriers.
  • Automated metrics often correlate poorly with clinical usefulness and safety, while clinician evaluation appears in only 11 of 46 papers.

Limitations

The survey’s scope and evidence base constrain how broadly its findings should be interpreted. Coverage is concentrated in selected ED tasks, languages, regions, imaging modalities, and heterogeneous study designs.

  • The survey covers ED tasks aligned with triage, diagnosis and assessment, and disposition, excluding adjacent areas such as outpatient triage and post-discharge follow-up.
  • Its methodological findings apply most directly to English- and Chinese-language ED-NLP research, so generalisability to other languages and care settings requires caution.
  • Radiology report-generation coverage is restricted to chest X-rays because MIMIC-CXR concentrates the available ED-adjacent data there.
  • Because included studies differ substantially in task formulation, dataset design, and evaluation methodology, the survey avoids head-to-head benchmark rankings.

Ethical Considerations

The survey treats ED-NLP systems as supplementary tools requiring clinician oversight because incorrect or overtrusted outputs could contribute to unsafe decisions. Privacy, bias, hallucination, and limited generalisability remain central risks before real-world use.

  • NLP systems should supplement rather than replace clinical expertise because incorrect, incomplete, or poorly calibrated outputs could contribute to unsafe decisions.
  • Fluent LLM-generated text may encourage over-reliance even when the underlying reasoning is unsound.
  • API-based deployments raise privacy concerns, while disease prediction and triage models face demographic bias and limited cross-setting generalisability.
  • The survey calls for clinician oversight, prospective validation, subgroup-aware evaluation, and institutional scrutiny before real-world deployment.

A Emergency Department Phases

ED care is organised into triage and arrival, diagnosis and assessment, and disposition. Each phase collects and acts on different clinical information, motivating the survey’s task groupings.

  • A Emergency Department Phases: The ED workflow begins with arrival and triage, where nurses rapidly assess patients and assign structured triage categories.
  • A Emergency Department Phases: Triage tools must align with nurse-led structured protocols rather than diagnostic reasoning.
  • A Emergency Department Phases: During diagnosis and assessment, physicians combine hand-off notes, elicited symptoms, examination findings, and laboratory or imaging investigations.
  • A Emergency Department Phases: Documentation in diagnosis and assessment often follows SOAP: Subjective, Objective, Assessment, and Plan.
  • A Emergency Department Phases: Disposition formalises whether patients are discharged with follow-up instructions or admitted, while discharge summaries consolidate findings, treatments, and ongoing recommendations.

B The Promise of NLP in the ED

Clinical NLP has progressed from sparse and task-specific models toward pretrained transformers and specialised clinical language models, creating opportunities for ED support. However, ED deployment remains constrained by short, data-sparse encounters, privacy and safety requirements, and limited real-world validation.

  • NLP advances: Pretrained transformer architectures and large-scale pretraining have reshaped natural language understanding and generation.Earlier approaches included bag-of-words, linear, and probabilistic models for retrieval, classification, and sequence labelling.
  • Clinical NLP: Clinical pretraining and continued pretraining help models capture medical terminology, abbreviations, and documentation patterns.Specialised models have shown gains on clinical concept extraction, classification, inference, and medical question answering benchmarks.
  • ED applications: ED NLP has mainly predicted triage scores, admission, or critical illness from unstructured triage notes.Combining free text with structured data often improves results, but most studies are retrospective and have high risk of bias.
  • Deployment constraints: Only one reviewed triage study reported clinical deployment, leaving a gap between retrospective predictive performance and real-world impact.The review identifies limited validation and workflow integration despite strong predictive accuracy for some outcomes.
  • Deployment constraints: Short, data-sparse ED encounters and rapid documentation constrain methods that depend on longitudinal histories or rich contextual data.Privacy, safety, and workflow integration requirements also make generic language models harder to reuse without adaptation.

C Scope and Search Criteria

The survey uses a PRISMA-style, multi-stage search and screening process targeting NLP research anchored to ED workflows. Inclusion requires alignment with an ED phase, an ED-relevant language resource, a primary NLP contribution, and ED-anchored evaluation.

  • Search strategy: The search covered ACL Anthology, AAAI, NeurIPS, ACM, ISCA, Nature, and PLOS ONE venues.The initial search was completed in August 2025, with later relevant papers retrospectively added through submission.
  • Search strategy: Queries combined ED-related, clinical, NLP, artificial intelligence, and machine learning terms with Boolean operators and domain synonyms.At least one emergency-department, emergency-room, triage, clinical, or hospital term was required.
  • Inclusion criteria: Included papers had to map their task to at least one of the three defined ED phases.The phases are triage and arrival, diagnosis and assessment, and patient disposition.
  • Inclusion criteria: Included studies had to use ED-relevant language resources and make an NLP method a primary research contribution.Examples include triage notes, chief complaints, symptom descriptions, diagnostic impressions, and patient-clinician interactions.
  • Inclusion criteria: Evaluation had to be anchored to ED workflows through ED-specific data, tasks, or explicit acute-care framing.General clinical benchmarks or broad inpatient corpora alone were insufficient.
  • Screening process: Three sequential screening stages were conducted by the lead author, with independent review of 20% of Stage 2 abstracts yielding Cohen’s κ of 0.80.The stages were title screening, abstract screening, and full-text review; disagreements were resolved after Stage 2 review.

D Papers Included in the Survey

The survey catalogues all 46 included papers and organises their analysis around clinical settings, training paradigms, and evaluation paradigms.

  • Corpus overview: Tables 2 and 3 provide a paper-level catalogue of all 46 included studies.The catalogue supports analysis across clinical settings, training paradigms, and evaluation paradigms.

E Reported Performances

Reported performance is fragmented across ED NLP tasks, with constrained prediction generally showing stronger headline scores than open-ended generation. Training remains centred on supervised and transfer learning, while newer paradigms are increasingly used but difficult to compare because datasets, task formulations, and evaluation protocols vary widely.

  • Cross-task patterns: Performance is strongly task dependent: triage and disposition prediction generally report stronger headline scores than clinical interaction summarisation and radiology report generation.The contrast is between constrained prediction and open-ended generation tasks.
  • Cross-task patterns: Automatic diagnosis has the greatest evaluation heterogeneity because prior studies use disease prediction, symptom acquisition, dialogue generation, and multi-step reasoning formulations.Its broad range of task formulations complicates direct comparison.
  • Training paradigms: Supervised and transfer learning remain the most established training approaches across the reviewed tasks.They are particularly common in classification and structured generation settings with relatively stable labels and evaluation criteria.
  • Training paradigms: Agent-based, prompt-based, and retrieval-augmented methods are more common for multi-step reasoning, external knowledge, or flexible generation.These methods appear in tasks such as automatic diagnosis, but their gains remain difficult to interpret consistently across varying evaluation settings.
  • Evaluation paradigms: Classification studies commonly use accuracy, AUC, AUROC, and AUPRC, whereas summarisation and generation studies use ROUGE, BLEU, factuality, and concept-level measures.The reported values are a reference across heterogeneous studies rather than evidence of a single best-performing modelling strategy.
  • Evaluation paradigms: The performance tables are not directly comparable because datasets, label spaces, task formulations, and evaluation protocols differ substantially.Values remain on the original scale used by each paper.
Loading 2608.23627v1…