CARE: A Conformal Safety Layer for Medical Summarization

2026-06-07

AI Summary

This work addresses the critical challenges of key information omission and factual hallucination in medical summarization by large language models, for which existing error detection methods lack formal risk control. The authors propose CARE—a model-agnostic, training-free post-hoc safety layer that, for the first time, enables joint conformal calibration of both omission and hallucination risks. By integrating conformal prediction, joint threshold-space optimization, and a unified importance–coverage modeling framework, CARE achieves rigorous risk bounds (α = 0.15) with 95% confidence across five medical summarization tasks using only around 100 annotated samples. Preliminary clinical evaluation demonstrates that its calibrated uncertainty markers improve omission detection rates by an average of 28.6 percentage points, substantially reducing the burden of manual review.

Abstract

Large language models (LLMs) are increasingly used for medical summarization, but their outputs can omit medically important information and introduce unsupported claims. Existing error-detection methods produce heuristic or uncalibrated scores, providing no formal control over missed errors and no principled way to trade off safety against clinician review burden. We introduce Conformal Assessment for Risk Evaluation (CARE), a post-hoc, model-agnostic safety layer that uses conformal risk control to overlay calibrated omission and hallucination flags onto summaries from any LLM without retraining. CARE provides finite-sample, distribution-free guarantees through two controllers: a hallucination controller that bounds the probability of a document containing any unflagged hallucinated sentence, and an omission controller that bounds the expected fraction of important omissions not surfaced for review. Unlike hallucination detection, omissions depend jointly on whether a source sentence is important and whether it is covered by the summary. We show that calibrating only one dimension can violate the target risk bound, while marginal decompositions remain valid but overly conservative. By jointly calibrating over the full $(τ,γ)$ threshold space, CARE preserves formal guarantees while surfacing up to 5$\times$ fewer sentences than alternative calibrated baselines. Across five medical summarization tasks, CARE satisfies the target risk bound at $α= 0.15$ with 95% confidence across 100 calibration/test resplits, using only ~100 labeled documents per domain. In a preliminary clinician study (75 document reviews), calibrated flags improved omission detection by 28.6 percentage points on average. These results show that sentence-level safety guarantees are feasible for LLM-assisted medical summarization and offer a tunable mechanism for balancing residual risk and review effort.

Problem

• medical summarization
• hallucination
• omission
• conformal risk control
• safety guarantees

Innovation

• conformal risk control
• medical summarization
• hallucination detection
• omission detection
• model-agnostic safety layer