Domain-Adapted Small Language Models for Reliable Clinical Triage
Researchers fine-tuned Qwen2.5-7B to accurately predict Emergency Severity Index (ESI) levels using pediatric triage data from Children’s National Hospital. By training it on both expert scenarios and real patient records, the model showed improved accuracy over raw text input. The best method involved generating concise clinical vignettes from unstructured notes for prediction, achieving a strong balance of performance and efficiency.