LLMs Outperform Traditional Classifiers in Dynamic Outpatient Referral
A recent study published on arXiv (2503.08292) examines outpatient referral (OR) as a dynamic process influenced by the acquisition of information and the reduction of uncertainty, disputing the typical view of OR as a static classification issue. The researchers evaluated both fixed patient information scenarios and dynamic multi-turn dialogue situations to determine if large language models (LLMs) enhance referral results through improved predictions or more effective inquiries. Results indicate that while LLMs show minimal benefits over traditional classifiers in static referral precision, they excel in dynamic contexts by posing insightful follow-up questions that diminish uncertainty regarding potential departments. The key strength of LLMs is their capability for interactive questioning rather than their predictive precision.
Key facts
- Study published on arXiv with ID 2503.08292
- Outpatient referral is a core clinical workflow assigning patients to hospital departments
- Research compares static and dynamic scenarios for LLM performance
- LLMs show limited advantage in static referral accuracy
- LLMs outperform traditional classifiers in dynamic settings
- LLMs ask discriminative follow-up questions to reduce uncertainty
- Primary value of LLMs is interactive questioning ability
- Study challenges static classification approach to outpatient referral
Entities
Institutions
- arXiv