Scholarlink

<p>Ambient and artificial intelligence scribes are rapidly adopted to address documentation burden, productivity and quality challenges, yet their evaluation is limited by reliance on the clinical note as the primary reference standard. Clinical documentation, a downstream artifact of conversational care, may omit, distort or introduce information not present in the source interaction. This paper proposes an episode-based, conversation-linked research dataset architecture using hospital phone encounters as a pragmatic implementation pathway. The proposed infrastructure links time-aligned audio, transcripts, and derived acoustic representations to electronic health record artifacts and downstream outcomes, with explicit representation of linkage uncertainty and versioned provenance. Such a dataset enables systematic assessment of intrinsic documentation quality and extrinsic fidelity to source conversations, including safety-netting and uncertainty preservation. The paper also reports governance considerations related to consent variability, secondary research use, medicolegal implications, and equitable performance across demographic and linguistic subgroups. By defining a governance-aware dataset class, this approach supports reproducible evaluation and responsible development of AI-assisted clinical documentation systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hospital Phone Encounters as Upstream Evidence for AI Scribe Evaluation

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Emre Sezgin

摘要

Ambient and artificial intelligence scribes are rapidly adopted to address documentation burden, productivity and quality challenges, yet their evaluation is limited by reliance on the clinical note as the primary reference standard. Clinical documentation, a downstream artifact of conversational care, may omit, distort or introduce information not present in the source interaction. This paper proposes an episode-based, conversation-linked research dataset architecture using hospital phone encounters as a pragmatic implementation pathway. The proposed infrastructure links time-aligned audio, transcripts, and derived acoustic representations to electronic health record artifacts and downstream outcomes, with explicit representation of linkage uncertainty and versioned provenance. Such a dataset enables systematic assessment of intrinsic documentation quality and extrinsic fidelity to source conversations, including safety-netting and uncertainty preservation. The paper also reports governance considerations related to consent variability, secondary research use, medicolegal implications, and equitable performance across demographic and linguistic subgroups. By defining a governance-aware dataset class, this approach supports reproducible evaluation and responsible development of AI-assisted clinical documentation systems.