💬 NLP

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

This survey reviews 46 papers on natural language processing in emergency departments, analyzing methods and trends across triage, diagnosis, and discharge phases while highlighting open challenges such as generalizability, noisy data, and workflow constraints.

Original authors: Dipankar Srirag, Aditya Joshi, Salil Kanhere, Padmanesan Narasimhan

Published 2026-08-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Dipankar Srirag, Aditya Joshi, Salil Kanhere, Padmanesan Narasimhan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Emergency rooms are places where time moves differently. In these chaotic, high-stakes environments, doctors and nurses must make life-or-death decisions within a narrow window, often just four hours from a patient's arrival to their final destination. Every second counts, and the pressure is immense. Amidst the shouting, the beeping monitors, and the rush of bodies, a massive amount of information is generated: whispered complaints, shouted instructions, scribbled notes, and complex medical reports. For decades, this flood of spoken and written words has been a burden, taking valuable time away from the actual work of healing. Now, a new kind of tool is emerging to help manage this deluge. It is not a robot nurse or a futuristic scanner, but a form of computer software known as natural language processing. This technology allows machines to read, understand, and summarize human language, offering a potential lifeline to overwhelmed medical staff by turning chaotic conversations into clear, organized records.

A team of researchers from the University of New South Wales in Australia recently took a close look at how this technology is currently being used in emergency rooms. They did not just look at one specific trick or one new gadget; instead, they gathered and analyzed forty-six different scientific studies to see the full picture of what is happening right now. Their goal was to understand how computers are being taught to handle the three main stages of an emergency room visit: the initial sorting of patients by urgency, the process of diagnosing what is wrong, and the final decision on whether to send a patient home or admit them to a hospital. By reviewing this wide range of work, they discovered that while computers are getting better at reading medical text, they are still far from being able to replace the human judgment required in a crisis.

The researchers found that most of the current work focuses on the middle stage of an emergency visit: the diagnosis. This is where doctors spend the most time talking to patients and reviewing test results, creating a rich supply of text for computers to study. In these studies, software is being trained to summarize long doctor-patient conversations into short, structured notes, or to suggest possible diseases based on a list of symptoms. However, the very first stage of the emergency room visit, known as triage, receives surprisingly little attention. This is the moment when a nurse quickly decides how sick a patient is and how fast they need to be seen. Despite being the most time-critical part of the process, only a handful of studies have tried to teach computers to help with this initial sorting. Similarly, the final stage, where a patient is sent home or to a ward, has seen limited research, even though clear instructions for patients leaving the hospital are crucial for their recovery.

When the researchers looked at how these computer systems are built, they saw a clear shift in strategy. In the past, scientists built custom-made computer programs for each specific task, like a specialized tool for every job. Today, the trend has moved toward using large, pre-trained language models. These are massive computer systems that have already read huge amounts of text and learned the general rules of language before being taught anything about medicine. It is like taking a person who already knows how to speak and read fluently and then giving them a crash course in medical terminology, rather than teaching them the alphabet from scratch. This approach has become the standard, with most new studies using these powerful, pre-existing models to understand medical notes and conversations.

Despite these advances, the researchers identified a significant gap between how these systems are tested and how they actually work in a real emergency room. Most of the studies they reviewed relied on old records or computer simulations. They fed the software clean, perfect transcripts of conversations that had already happened, often with all the answers already known. In a real emergency room, however, the situation is messy. Speech recognition software often makes mistakes, patients might speak in broken sentences or different languages, and doctors are under extreme time pressure. The study suggests that because the testing conditions are so much cleaner than reality, the computer systems might look much smarter on paper than they would in a busy, noisy hospital corridor. Furthermore, very few of these systems have ever been tested in a live emergency room setting. The researchers found that not a single paper in their review reported a system that was actually deployed and used by doctors during real patient care.

Another major concern is how these systems are judged. In the scientific papers, success is often measured by how closely a computer's output matches a reference text, using automated scores. But in medicine, a small error can be dangerous. A computer might get the grammar perfect but miss a critical detail about a patient's allergy, or it might suggest a diagnosis that is technically possible but highly unlikely. The researchers noted that only a small number of studies included feedback from actual doctors to see if the computer's work was truly useful or safe. Without this human check, it is difficult to know if these tools are ready for the real world. The study also highlighted that the data used to train these systems comes mostly from English and Chinese sources, meaning the technology might not work well for patients speaking other languages or from different cultural backgrounds.

Ultimately, this survey paints a picture of a field that is full of promise but still in its early, experimental days. The technology to read and summarize medical text exists and is improving rapidly, driven by powerful new computer models. Yet, the path from a computer program that works in a lab to a tool that doctors can trust in a life-or-death emergency is long and fraught with challenges. The researchers conclude that before these systems can become a standard part of emergency care, they must be tested in real, messy hospital environments, evaluated by human doctors, and designed to handle the noise and uncertainty of actual patient interactions. Until then, these tools remain powerful assistants in the making, waiting to prove they can handle the chaos of the emergency room as well as the humans who work there.

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

Problem Statement
Emergency Departments (EDs) operate under severe time pressure, generating multimodal data including clinical conversations, triage notes, and discharge documents. While Natural Language Processing (NLP) has advanced significantly with pre-trained transformers and Large Language Models (LLMs), existing surveys either map clinical NLP across the broader hospital workflow or focus on specific tasks without addressing the unique constraints of the ED. There is a fragmentation in the literature regarding how ED-oriented NLP tasks are formulated, how modeling paradigms are chosen, and whether evaluation practices reflect the operational realities of emergency care (e.g., time pressure, noisy inputs, and workflow constraints). This survey addresses the gap by systematically analyzing NLP research specifically applied to the three phases of the ED workflow: triage, diagnosis/assessment, and disposition.

Methodology
The authors conducted a systematic survey of 46 papers spanning the ED workflow. The selection process followed PRISMA-style reporting, searching venues in NLP, AI, and health informatics (e.g., ACL, AAAI, NeurIPS, Nature) using keywords related to emergency departments, triage, and NLP.

The analysis was structured across three primary dimensions:

Clinical Settings: Categorizing papers by ED phase (Triage, Diagnosis, Disposition), language (English, Chinese, etc.), and ED-specificity (ED-native, ED-anchored, ED-adjacent).
Training Paradigms: Analyzing input modalities (text, structured EHR, imaging) and training strategies (supervised, transfer learning, agent-based, reinforcement learning, retrieval-augmented generation, etc.).
Evaluation Paradigms: Examining study designs (retrospective, synthetic, prospective), the use of clinician evaluation versus automated metrics, and deployment status.

The survey formalizes ED NLP tasks under a common conditional prediction framework ( $\hat{y} = \arg \max_{y \in Y} p_\theta(y | x)$ ), where $x$ represents information available at a specific ED phase and $y$ is the required output.

Key Contributions and Findings

1. Task-Specific Analysis
The survey details six core NLP tasks across the ED workflow:

Triage Classification: Assigning urgency levels (e.g., ESI 1–5). Current methods rely heavily on transfer learning with pretrained transformers (combining chief complaints and vital signs) and emerging agent-based or retrieval-augmented LLM systems. A critical finding is that most evaluations use post-hoc data containing information not available at the moment of triage, inflating accuracy.
Clinical Interaction Summarisation: Converting doctor-patient dialogues into structured notes (e.g., SOAP format). The field has shifted from task-specific neural architectures (e.g., pointer-generator networks) to transfer learning and prompt-based LLM approaches. Evaluations often assume clean transcripts, ignoring real-world ASR errors and informal phrasing.
Automatic Diagnosis: A sequential task where models interleave symptom inquiries with diagnosis. Approaches include reinforcement learning (optimizing for inquiry efficiency), supervised learning, knowledge-grounded methods (using knowledge graphs), and multi-agent systems. The survey notes a mismatch between turn-count optimization in models and the time-bounded nature of real ED diagnosis.
Radiology Report Generation (RRG): Specifically for chest X-rays. Methods have evolved to include clinical context (vitals, medications) and retrieval-augmented generation to reduce hallucinations. Evaluation relies on automated metrics (ROUGE, BLEU) which correlate poorly with clinical correctness.
Disposition Prediction: Forecasting admission vs. discharge and destination (ward/ICU). Models utilize multimodal inputs (text, vitals, and increasingly video) but often optimize for aggregate AUROC, ignoring the asymmetric clinical cost of errors (e.g., under-triage of critical patients).
Discharge Instruction Generation: Translating care plans into patient-facing guidance. Techniques include neural machine translation for simplification and knowledge-grounded generation to improve factuality.

2. Research Trends

Shift in Modeling: There is a clear transition from task-specific neural architectures to pre-trained transformer models and LLMs.
Emerging Paradigms: Growing interest in interactive clinical systems, including agent-based workflows and retrieval-augmented generation (RAG), particularly in 2024–2025.
Evaluation Gaps: The vast majority of studies (24/46) rely on retrospective records or synthetic data. Only one study adopted a prospective design. Furthermore, only 11/46 papers included clinician evaluation, with most relying on automated metrics that do not align well with clinical utility.
Deployment: No paper in the corpus reports clinical deployment within an ED-specific workflow.

3. Open Challenges
The authors identify several critical barriers to real-world application:

Data Limitations: Reliance on single-institution datasets (e.g., MIMIC family) limits generalizability. There is a scarcity of multilingual datasets.
Robustness: Systems often assume clean, structured inputs, whereas ED data is noisy (ASR errors, incomplete histories, informal language).
Evaluation Misalignment: Standard metrics (accuracy, ROUGE) fail to capture clinical safety and the ordinal nature of triage. Clinician evaluation is rare and rarely uses standardized protocols.
Deployment & Privacy: Barriers include API-based privacy concerns, integration with EHR infrastructure, and the lack of prospective validation.

Significance and Claims
The paper claims to be the first survey to comprehensively analyze NLP research specifically across the three distinct phases of the ED workflow (triage, diagnosis, disposition). It highlights that while methodological advances (LLMs, agents) are rapid, the field suffers from a "validation gap" where models are not tested against real-time clinical constraints or evaluated by clinicians.

The authors argue that future ED-NLP research must move beyond offline benchmarks toward:

Realistic Benchmarks: Utilizing resources like MC-BEC and MedJourney that reflect multimodal, process-centric workflows.
Clinically Grounded Evaluation: Adopting checklist-based protocols and subgroup analyses to ensure safety and fairness.
Prospective Validation: Moving from retrospective/simulated studies to prospective trials in live ED settings.
Workflow Integration: Developing systems that handle noisy, time-pressured inputs and integrate transparently with clinician workflows.

The survey concludes that while NLP holds promise for reducing documentation burden and supporting decision-making, the transition from "offline benchmarks to real-time clinical impact" requires closer collaboration between NLP researchers, clinicians, and healthcare institutions, alongside more rigorous, clinically aligned evaluation standards.

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

More like this