This research demonstrates how synthesized annotation guidelines can significantly improve large language models' performance in clinical information extraction without requiring deep domain knowledge.

• LLMs with synthesized guidelines outperform baseline few-shot approaches by 14.8% on clinical NER tasks
• The approach is knowledge-lite, eliminating the need for domain experts to create detailed guidelines
• Performance improvements are consistent across different model sizes and clinical datasets
• Guidelines synthesized by GPT-4 proved as effective as human-created guidelines

For healthcare organizations, this approach offers a cost-effective way to deploy advanced NLP solutions with minimal domain expertise requirements, potentially accelerating clinical data extraction and analysis workflows.

Synthesized Annotation Guidelines are Knowledge-Lite Boosters for Clinical Information Extraction