Small Language Models Show Promise for Reliable Clinical Triage
An investigation released on arXiv assesses the potential of open-source small language models (SLMs) as effective, privacy-conscious tools for assigning Emergency Severity Index (ESI) in emergency departments. The researchers conducted a thorough comparison of various SLMs using different prompting methods, discovering that clinical vignettes—brief summaries of triage cases—produced the most precise predictions. The SLM Qwen2.5-7B exhibited the best combination of accuracy, stability, and computational efficiency. By employing extensive domain adaptation with expert-reviewed and silver-standard pediatric triage data, fine-tuned Qwen2.5-7B models significantly minimized discordance and critical errors, surpassing all baseline SLMs and advanced proprietary large language models. This study tackles ongoing issues related to inconsistent free-text triage documentation that leads to mistriage and inefficiencies in workflow.
Key facts
- Study evaluates open-source small language models for ESI assignment
- Clinical vignettes yielded most accurate predictions
- Qwen2.5-7B showed best balance of accuracy, stability, and efficiency
- Domain adaptation used expert-curated and silver-standard pediatric triage data
- Fine-tuned models reduced discordance and clinically significant errors
- Outperformed all baseline SLMs and advanced proprietary LLMs
- Addresses mistriage and workflow inefficiencies from variable documentation
- Published on arXiv with ID 2604.26766
Entities
Institutions
- arXiv