Few-shot LLM framework extracts data from Spanish mammography reports

GPT few-shot with RAG achieved F1 0.89-0.94 for named entity extraction vs fine-tuned 0.97 in Spanish mammography reports, but relation extraction lower (0.78 vs 0.99) without task training.

• Inference cost was fractions of a cent per report, with latency in seconds.
• Local open-weight models preserved privacy but had substantially lower accuracy.
• The framework still requires an initial annotated subset to populate the retrieval store.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.