Performance evaluation and benchmarking across 16 large language models on a comprehensive real-world emergency department triage data set

Technical Summary: Performance Evaluation and Benchmarking of Large Language Models on Real-World Emergency Department Triage Data

Problem Statement
Emergency Department (ED) triage is a high-stakes process determining patient prioritization and resource allocation under time pressure. While Large Language Models (LLMs) have been proposed as decision-support tools for triage, existing evaluations predominantly rely on simulated scenarios or curated datasets (e.g., MIMIC-IV, NEJM Healer). There is a critical lack of evidence regarding the performance, calibration, and reproducibility of contemporary LLMs when applied to comprehensive, uncurated, real-world clinical data. Furthermore, the impact of prompting strategies versus model architecture on triage accuracy remains unclear.

Methodology
This retrospective cross-sectional benchmarking study was conducted at a tertiary academic emergency center in Germany featuring an integrated Central Point of Assessment (CPA).

• Dataset: The study utilized 16,107 consecutive adult walk-in encounters (October 2023–February 2024). Data were extracted from a structured clinical decision support system (CDSS) capturing presenting complaints (CEDIS-PCL), vital signs (ABCDE framework), and final triage decisions made by specialized nursing staff. The data were fully anonymized.
• Models Evaluated: The study assessed 16 configurations, including 14 individual LLMs (proprietary and open-source, varying in size and reasoning capabilities), one ensemble strategy ("Council of Models"), and a specific stepwise prompting approach.
• Evaluation Protocol:
  • Inputs: Models received structured clinical variables for each encounter.
  • Outputs: Models were tasked with predicting the Emergency Severity Index (ESI) level (1–5), sectoral allocation (ED vs. Urgent Care Practice/UCP), and a self-reported verbalized confidence level (0–100%).
  • Strategies:
    • Standard Prompting: Direct classification based on available data.
    • Council of Models: Aggregation of predictions from nine models, with a "chairman" model determining the final ESI.
    • ESI Stepwise: A structured, decision-tree approach guiding the model sequentially through the ESI algorithm (life-saving intervention need, high-risk features, danger-zone vitals, resource estimation).
  • Reproducibility: A subset of 100 patients was queried 100 times each (10,000 total classifications) using identical inputs to measure stochastic variability.
• Metrics: Primary outcomes included agreement measured by quadratic-weighted Cohen's $\kappa$ ($\kappa_{qw}$), misclassification patterns (over/under-triage), Expected Calibration Error (ECE), and output reproducibility.

Key Results

• Triage Accuracy (ESI Classification):
  • Performance was highly heterogeneous across models. Quadratic-weighted $\kappa$ values ranged from 0.18 (GPT 5 Nano) to 0.75.
  • Only the ESI Stepwise approach achieved "substantial" agreement ($\kappa_{qw} = 0.747$), approaching the lower bound of reported human inter-rater reliability.
  • Most other models demonstrated only "moderate" to "fair" agreement ($\kappa_{qw}$ 0.18–0.52).
  • The most common errors occurred between adjacent levels, particularly ESI 3 and ESI 4.
• Sectoral Allocation (ED vs. UCP):
  • Agreement on sectoral assignment was uniformly low across all models ($\kappa \leq 0.30$).
  • Misallocations were frequent; for instance, some models misallocated up to 75% of patients intended for the ED to the UCP, or vice versa.
• Bias and Calibration:
  • Overconfidence: All 14 individually evaluated models were systematically overconfident. Overconfidence indices ranged from +5.2 to +33.7 percentage points (mean verbalized confidence minus actual accuracy).
  • Calibration: Expected Calibration Error (ECE) was high (0.099–0.355). Verbalized confidence showed no meaningful association with correct answers (Confidence Accuracy $\approx$ 0), indicating models cannot reliably signal their own uncertainty.
  • Directional Bias: 13 of 16 models exhibited net over-triage (assigning higher acuity than nurses), while the ESI Stepwise approach and GPT 5 Nano variants showed significant under-triage.
• Reproducibility:
  • Output was non-deterministic for a substantial minority of cases. In the reproducibility experiment, 23% of patients received their most common ESI classification in fewer than 70 out of 100 runs.
  • Majority voting over 100 runs did not significantly improve accuracy, suggesting the variability stems from genuine decision-boundary uncertainty rather than simple sampling noise.

Key Contributions

1. Real-World Benchmarking: This is the first study to systematically evaluate over a dozen contemporary LLMs on a comprehensive, uncurated real-world dataset from an active emergency department, moving beyond fictional vignettes or curated archives.
2. Prompting vs. Architecture: The study demonstrates that the prompting strategy is more influential than model size or architecture. A structured, algorithm-guided stepwise approach significantly outperformed high-reasoning models using unstructured prompts, suggesting that encoding clinical logic into the prompt is critical for performance.
3. Safety and Reliability Analysis: The paper provides empirical quantification of LLM output variability in a clinical setting, revealing that a significant portion of cases yield inconsistent classifications even with identical inputs. It also highlights the critical safety vulnerability of uncalibrated verbalized confidence scores.

Significance and Claims
The authors conclude that current large language models demonstrate heterogeneous and generally limited performance in real-world emergency triage tasks. While LLMs hold potential as supportive tools, their current characteristics—specifically the lack of calibration, non-deterministic outputs for a subset of patients, and poor sectoral allocation accuracy—do not support autonomous or safety-critical deployment in clinical triage environments.

The study emphasizes that improvements in calibration, reliability, and workflow integration are prerequisites for clinical implementation. It suggests that future development should prioritize structured algorithm-guided prompting and regulatory-compliant validation in prospective settings over simply selecting larger or more complex models. The findings underscore the need for external calibration mechanisms and caution against presenting self-reported confidence scores to clinicians as probability estimates.