- 期刊JMIR medical informatics（JMIR Med Inform）
- 发表时间2026-10-29
- 作者Su X, Zhao X, Yao G, Wang Y
- DOI10.2196/101606
- PMID42969579
摘要机译
[背景]关于高血压伴焦虑、抑郁或失眠的中医文献包含对证候、治疗原则、方剂和草药的描述。聚焦的知识图谱可以组织这些关联及其来源，但其对生成答案的贡献需要单独评估。
[目的]构建高血压情绪共病中医知识资源，评估被策展关系的回顾子集，并评估相关问答实施，包括固定生成模型图检索的增量效果。
[方法]Qwen2.5-7B-Instruct从419个中文记录的标题、关键词和摘要中生成候选者，然后由专家策划，辅以教科书和理论。专家验证仅限于从早期100个关系评论样本中发布的图表中存在的80个关系。
三个构建之外的临床医生撰写了73个问题，另外两个临床医生使用预先指定的四类标准盲目地对匿名回答进行评分。评估包括仅检索和仅模型对照、生成模型比较和六种检索配置。另外一个匹配的DeepSeek-V4-Pro实验在有和没有图形检索的情况下，每个条件下每个问题产生一个响应。
使用精确的双面McNemar检验比较严格正确性；部分正确性和记录的安全问题分别报告。[结果]该资源包含956个索引术语、101个图节点和312个与记录源归属的关系。专家确认发布的图中存在79/80个审查关系（98.8%，95%CI 93.3%-99.8%）；提取输出中没有274/312个最终关系。
在匹配实验中，有图检索的严格准确率为50/73（68%，95%CI 57%-78%），无图检索的严格准确率为57/73（78%，95%CI 67%-86%）。三个不和谐对倾向于图辅助，10个倾向于纯模型生成（P=.09）。部分正确的回答分别为22/73（30%）和15/73（21%）。两种条件都有72/73correct-or-partially-correct回答（99%）；
同样的问题在两种情况下都有记录的潜在安全问题。在带有图检索的单独生成模型评级回合中，Qwen的严格准确率为9/73（12%），DeepSeek的严格准确率为26/73（36%）。[结论]精选图提供了结构化的、可检查的领域知识，并具有来源归属；专家确认适用于已发布关系的审查子集。
匹配实验没有证明从图检索中提高了严格的答案准确性，也没有建立等价性。资源和评估为有针对性的开发提供了基础；文献导航和临床使用的好处需要直接验证。
展开英文摘要原文
【BACKGROUND】Traditional Chinese medicine (TCM) literature on hypertension with anxiety, depression, or insomnia contains descriptions of syndrome patterns, treatment principles, formulas, and herbs. A focused knowledge graph can organize these associations and their sources, but its contribution to generated answers requires separate evaluation.
【OBJECTIVE】To construct a TCM knowledge resource for hypertension with emotional comorbidity, assess a reviewed subset of curated relations, and evaluate the associated question-answering implementation, including the incremental effect of graph retrieval with a fixed generation model.
【METHODS】Qwen2.5-7B-Instruct generated candidates from the titles, keywords, and abstracts of 419 Chinese-language records, followed by expert curation with textbook and theoretical supplementation. Expert validation was restricted to the 80 relations present in the released graph from an earlier 100-relation review sample. Three clinicians external to construction authored 73 questions, and two additional clinicians blindly rated anonymized responses using a prespecified four-category rubric. Evaluations included retrieval-only and model-only controls, generation-model comparisons, and six retrieval configurations. An additional matched DeepSeek-V4-Pro experiment generated one response per question per condition with and without graph retrieval. Strict correctness was compared using the exact two-sided McNemar test; partial correctness and recorded safety concerns were reported separately.
【RESULTS】The resource contained 956 indexed terms, 101 graph nodes, and 312 relations with recorded source attribution. Experts confirmed 79/80 reviewed relations present in the released graph (98.8%, 95% CI 93.3%-99.8%); 274/312 final relations were absent from the extraction output. In the matched experiment, strict accuracy was 50/73 (68%, 95% CI 57%-78%) with graph retrieval and 57/73 (78%, 95% CI 67%-86%) without it. Three discordant pairs favored graph assistance and 10 favored model-only generation (P=.09). Partially correct responses numbered 22/73 (30%) and 15/73 (21%), respectively. Both conditions had 72/73 correct-or-partially-correct responses (99%); the same question had a recorded potential safety concern in both conditions. In the separate generation-model rating round with graph retrieval, strict accuracy was 9/73 (12%) for Qwen and 26/73 (36%) for DeepSeek.
【CONCLUSIONS】The curated graph provides structured, inspectable domain knowledge with source attribution; expert confirmation applies to the reviewed subset of released relations. The matched experiment did not demonstrate improved strict answer accuracy from graph retrieval and did not establish equivalence. The resource and evaluations provide a basis for targeted development; benefits for literature navigation and clinical use require direct validation.
本页由系统自动抓取 PubMed 元数据生成；标题与摘要版权归原作者与期刊所有；机译内容仅供参考，请以英文原文为准；收录不代表本站观点，亦不作为用药或医疗建议。