CARE: A Conformal Safety Layer for Medical Summarization

TL;DR

Large language models (LLMs) used for medical summarization often miss critical information and make unsupported claims. Conformal Assessment for Risk Evaluation (CARE) is introduced as a post-hoc, model-agnostic safety layer that quantifies risk in LLM outputs.

Key Takeaways

• Integrate CARE for improved medical summarization safety.
• Use calibrated flags to enhance omission detection.
• Leverage LLMs with reduced clinician review burden.
• Achieve reliable outputs with limited labeled data.