Accuracy of the large language model ChatGPT in adult emergency department triage: a systematic review and meta-analysis.

Deep Analysis: ChatGPT in Adult Emergency Department Triage

Clinical Hook

The perpetual challenge of efficient and accurate triage in overcrowded emergency departments drives an urgent search for innovative solutions, making the potential of large language models like ChatGPT a compelling, albeit complex, area of investigation.

PICO Breakdown

• P (Population): Adult patients presenting to the emergency department for triage.
• I (Intervention/Index Test): The large language model ChatGPT, used to assign triage levels or assess patient acuity.
• C (Comparison/Reference Standard): Human-assigned triage (typically by experienced ED nurses or physicians) using established triage systems (e.g., Emergency Severity Index [ESI], Canadian Triage and Acuity Scale [CTAS], Manchester Triage System [MTS]), or expert consensus regarding appropriate triage levels.
• O (Outcome): Accuracy of ChatGPT in triage, typically measured by metrics such as sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall agreement or kappa statistics, often with a focus on its ability to correctly identify high-acuity patients and avoid undertriage.

Critical Appraisal

This systematic review and meta-analysis provides a high-level synthesis of current evidence regarding ChatGPT's performance in ED triage, offering valuable insights but also highlighting significant limitations inherent in both the underlying technology and the nascent research landscape.

Strengths of the Study:

• Robust Methodology: As a systematic review and meta-analysis, this study represents the highest level of evidence for synthesizing existing research. It presumably followed PRISMA guidelines, ensuring a comprehensive search strategy and systematic evaluation of included studies, minimizing publication bias.
• Timely and Relevant Topic: The rapid advancement and public uptake of LLMs necessitate such evaluations, particularly in critical healthcare settings like the ED.
• Quantitative Synthesis: By pooling data, the meta-analysis attempts to provide a more precise estimate of ChatGPT's accuracy than individual studies, offering a summary performance metric.

Limitations and Considerations:

• Heterogeneity of Included Studies: This is a critical concern for any meta-analysis involving rapidly evolving technology.
  • ChatGPT Version Variability: The studies likely included different versions of ChatGPT (e.g., GPT-3.5, GPT-4, various iterations within these), each with distinct training data, capabilities, and performance characteristics. Lumping these together can obscure individual model performance and make the aggregated result less representative of any single current model.
  • Prompt Engineering Variability: The phrasing and structure of prompts given to ChatGPT significantly impact its output. Studies likely used diverse prompting strategies, leading to considerable input heterogeneity.
  • Reference Standard Variability: The specific human triage systems (ESI, CTAS, MTS) and the expertise of the human adjudicators (nurses, physicians, consensus panels) can vary, making direct comparisons challenging.
  • Outcome Definition Inconsistency: The precise definitions of "accuracy," "undertriage," and "overtriage" (e.g., specific ESI levels, binary high vs. low acuity) can differ across studies.
  • Study Design Diversity: Included studies could range from retrospective analyses of anonymized charts to simulated scenarios or even small prospective pilots, each with different risks of bias and generalizability.
• Quality of Included Primary Studies: While the meta-analysis synthesizes existing data, it is only as strong as the primary studies it includes. Many initial studies on LLMs in healthcare are pilot-level, retrospective, or utilize simulated data, which may not accurately reflect real-world performance.
• Focus on "Accuracy" vs. "Safety": While overall accuracy is reported, the most critical metric in ED triage is safety – specifically, the sensitivity for identifying high-acuity patients and avoiding undertriage. A model might have high overall accuracy but unacceptable undertriage rates for critical conditions. The abstract's conclusion of "not sufficiently accurate or reliable" likely implicitly addresses this safety concern.
• Black Box Nature: The lack of transparency in LLM decision-making makes it difficult to understand why certain triage recommendations are made, hindering error analysis and trust-building.
• Dynamic Nature of LLMs: The rapid pace of LLM development means that a review of past performance may quickly become outdated. What was true for GPT-3.5 might not be true for a newer, more advanced model, impacting the long-term relevance of the findings.
• Generalizability to Real-World Settings: Many studies might be based on curated datasets or specific clinical scenarios rather than the full spectrum of patient presentations, language nuances, and cognitive biases encountered in a busy ED.

In summary, the review appropriately concludes that current ChatGPT versions are not ready for independent ED triage. This is largely due to observed performance gaps (especially concerning safety) and the methodological challenges of robustly evaluating an evolving technology across diverse clinical contexts. The heterogeneity across studies, particularly regarding LLM versions and prompt engineering, likely contributed significantly to the "not sufficiently accurate or reliable" finding, as inconsistent performance across different iterations and applications would prevent a confident endorsement.

Practice Application

The findings of this systematic review and meta-analysis provide crucial guidance for clinicians, hospital administrators, and researchers regarding the present capabilities of ChatGPT in adult emergency department triage.

Immediate Implications for Clinical Practice:

• No Independent Use: The most important takeaway is that ChatGPT, in its current state, cannot and should not be used independently for adult emergency department triage. Its accuracy and reliability are insufficient, meaning it poses an unacceptable risk of undertriage for high-acuity patients, which could lead to significant patient harm or adverse outcomes.
• Maintain Human Oversight: Triage remains a complex process requiring nuanced clinical judgment, empathy, and the ability to interpret non-verbal cues – all beyond the current capabilities of LLMs. Human clinicians must continue to be the primary decision-makers in triage.

Potential Future Roles and Research Directions:

While independent use is premature, this review does not negate the potential of AI in ED triage. Future applications will likely involve augmentation rather than replacement:

• Decision Support Tool: ChatGPT or similar LLMs could evolve into sophisticated decision support tools, flagging potential high-acuity cases that might be missed, suggesting differential diagnoses, or providing quick access to clinical guidelines during the triage process. This would require rigorous validation for specific use cases.
• Training and Education: LLMs could be valuable tools for training new triage nurses, simulating various patient scenarios, and providing feedback on triage decisions based on vast knowledge bases.
• Automated Documentation and Summarization: LLMs could assist with rapidly summarizing chief complaints, extracting key information from patient notes, or drafting initial triage documentation, thereby potentially improving efficiency.
• Predictive Analytics: Beyond direct triage, LLMs could contribute to predictive models that forecast ED flow, identify patients at risk of deterioration, or optimize resource allocation based on incoming patient profiles.

Prerequisites for Future Clinical Integration:

Before any LLM can be safely integrated into ED triage workflows, several critical advancements and considerations are necessary:

• Demonstrably Higher Accuracy and Safety: Specifically, extremely high sensitivity for identifying high-acuity patients, minimizing undertriage to near-zero levels.
• Robust Prospective Validation: Large-scale, real-world prospective studies are needed, comparing LLM performance against human experts in diverse clinical settings, using standardized protocols.
• Transparency and Explainability: The "black box" nature must be addressed. Clinicians need to understand how the LLM arrived at its recommendation to build trust and identify potential biases or errors.
• Integration with EHR Systems: Seamless integration with electronic health records is essential for real-time data access and automated documentation.
• Ethical Frameworks and Regulatory Oversight: Clear guidelines for accountability, data privacy, bias mitigation, and liability must be established.
• Continuous Learning and Updates: LLMs in clinical use would require continuous monitoring, retraining, and updating to reflect new medical knowledge and adapt to evolving patient populations.

In conclusion, the paper serves as a vital reality check. While the promise of AI in healthcare is immense, particularly in high-stakes environments like the ED, current LLM technology is not yet ready for independent decision-making in triage. The focus must now shift towards developing, meticulously validating, and safely integrating LLMs as sophisticated assistants under unwavering human oversight, rather than replacements.