时间标签：数据日期 2026-11-03 · 生成时间 2026-11-03 05:30（Asia/Shanghai）｜ 检索窗口 PubMed edat 2026/10/31 – 2026/11/03 ｜ 情报发生日期 2026-11-03
研究问题
To evaluate how different embedding strategies impact retrieval effectiveness in clinical IE, we selected a range of embedding models inspired by prior research [ 14 , 15 ].；For each framework, we tested top-k retrieval with k values of 1, 5, 10, 15, and 20 to determine how the number of retrieved documents impacts clinical information extraction performance.；The null hypothesis in each test assumed that there was no significant difference in performance between groups, while the alternative hypothesis suggested a significant difference.
方法
In this experiment, we used FAISS [ 21 ] for its efficient index structure for fast retrieval, scalability, and flexibility in handling high-dimensional vector data [ 22 , 23 ].；The evaluation was performed on the Gadi supercomputer provided by the Australian National Computational Infrastructure (NCI) [ 24 ], which features 48 CPUs, four NVIDIA Tesla V100 GPUs (each with 32 GB of memory), and 382 GB of system memory.；Statistical Analysis We first assessed the distribution of each performance metric using the Shapiro–Wilk test for normality, following prior recommendations [ 25 ].
核心结论
In contrast, the aged care clinical notes in our dataset include substantial informal documentation, such as colloquial expressions, non-standard abbreviations, shorthand notations, and narrative-style observations.；These results indicated that while no configuration was significantly superior, 3-shot prompting offered marginal benefits depending on the task.；While our goal was to identify optimal configurations for deployment scenarios where fine-tuning is impractical due to limited labelled data, computational resource constraint, and regulatory constraints, understanding the performance ceiling achievable with fine-tuning would contextualise the practical significance of our findings.
局限
However, their free-text format poses significant challenges for computational processing, as they are often written in inconsistent styles and include diverse terminology, abbreviations, acronyms, and narrative structures [ 3 ].；In this context, retrieval-augmented generation (RAG) addresses not merely a factual knowledge gap, but also a local and institutional knowledge gap that parametric knowledge alone cannot reliably bridge.
对本项目的价值
参考价值：属「衰弱 / 肌少症功能结局」方向；脑老化 / 神经退行 / 认知栏目可用作本项目文献库与综述素材。 ｜ 价值分 1.9/5（★★☆☆☆）
全文来源
PubMed PMC JATS（正文 43,979 字符，题名指纹 0.91，已过 DOI 逐字 + 题名指纹 + 正文长度三重归属校验）
原文题录
- 中文题名：RAG 配置在电子病历临床信息抽取中的评测：养老照护案例研究
- 英文标题：Evaluating RAG Configurations for Clinical Information Extraction from EHR Notes: Aged Care Case Study.
- 期刊与标识：J Healthc Inform Res · PMID 42964527 · DOI 10.1007/s41666-026-00254-8 · PMC PMC13719912 · EDAT 2026-10-31