Trustworthy Agents for Electronic Health Records through Confidence Estimation

TL;DR

This paper introduces a confidence-aware clinical agent that improves trustworthiness in EHR question answering by balancing accuracy and reliability, addressing hallucination risks of large language models in healthcare.

Contribution

It proposes HCAcc@k%, a new metric for evaluating accuracy-reliability trade-offs, and develops TrustEHRAgent, a confidence estimation method that outperforms baselines under strict reliability constraints.

Findings

TrustEHRAgent outperforms baselines at HCAcc@70% on MIMIC-III and eICU datasets.

Traditional metrics are insufficient for evaluating healthcare AI reliability.

The approach enhances trustworthiness by transparently expressing uncertainty.

Abstract

Large language models (LLMs) show promise for extracting information from Electronic Health Records (EHR) and supporting clinical decisions. However, deployment in clinical settings faces challenges due to hallucination risks. We propose Hallucination Controlled Accuracy at k% (HCAcc@k%), a novel metric quantifying the accuracy-reliability trade-off at varying confidence thresholds. We introduce TrustEHRAgent, a confidence-aware agent incorporating stepwise confidence estimation for clinical question answering. Experiments on MIMIC-III and eICU datasets show TrustEHRAgent outperforms baselines under strict reliability constraints, achieving improvements of 44.23%p and 25.34%p at HCAcc@70% while baseline methods fail at these thresholds. These results highlight limitations of traditional accuracy metrics in evaluating healthcare AI agents. Our work contributes to developing trustworthy clinical agents that deliver accurate information or transparently express uncertainty when confidence is low.