机构特定LLM提示可恢复去标识化系统遗漏的PHI
原标题：Institution-Specific LLM Prompting Recovers PHI That De-identification Systems and Their Gold Standards Both Miss
AI 摘要
贝勒医学院等机构的研究者在 arXiv 发表论文，评估了大型语言模型（LLM）在电子健康记录去标识化中识别机构特定受保护健康信息（PHI）的能力。研究对 100 份儿科肿瘤学笔记进行基准测试，发现 LLM（最佳 F1=0.918）优于专用系统（TiDE F1=0.779），且通过命名缺失类别和防止过度编辑的提示，可恢复 79% 的遗漏 PHI，并提升精确率。专家审查确认 LLM 标记的额外标识符均为真实 PHI，重新标注后召回率达 0.981。结论认为 LLM 是专用去标识化系统的适应性替代方案，机构特定提示开发应作为主要适配手段。
正文节选
Institution-Specific LLM Prompting Recovers PHI That De-identification Systems and Their Gold Standards Both Miss Article type: Research and Applications Authors. Daniel Palacios, BS1,2,3,4,, Matthew Brady Neeley, BS1,2,3,4,, Angel Adetomike Otto, MS5, Shalini Dhamodharan, MS5, John P. Woodhouse, BA5, Chi-fan Lin, MS5, Mark Zobeck, MD, MPH5,, Zhandong Liu, PhD1,2,3,4,, Hyun-Hwan Jeong, PhD2,3,4,. These authors contributed equally. Co-corresponding authors. Affiliations. - 1. - Quantitative and