Page 100 - 《中国药房》2026年2期
P. 100
基于AI幻觉抑制的药学智能问答平台的构建与效能验证 Δ
2
1
1
1
温正旺 ,王嘉莹 ,杨文月 ,杨昊煜 ,马 霄 ,刘 云 (1.邯郸市第一医院药学部,河北 邯郸 056003;2.东软
1 #
1*
集团有限公司,河北 邯郸 056003)
中图分类号 R95 文献标志码 A 文章编号 1001-0408(2026)02-0226-06
DOI 10.6039/j.issn.1001-0408.2026.02.16
摘 要 目的 构建低“人工智能(AI)幻觉”的药学智能问答平台,提升用药咨询的准确性、一致性与可追溯性。方法 利用Python
代码对药品说明书进行批量结构化整理并构建本地药学知识库,基于大型语言模型实现检索与问答流程设计,并在Dify平台完成
系统集成与本地化部署。通过设计典型临床用药问题,从达峰时间、半衰期检索及肾功能减退患者剂量调整方案推理等维度,将
药学智能问答平台的输出结果与在线版DeepSeek进行对比验证,评估其检索和推理结果的准确性与可靠性。结果 基于本地药
品说明书构建的药学智能问答平台在达峰时间、半衰期及剂量调整方案的检索和推理准确率均为100%。相比之下,在线版Deep‐
Seek在3个维度方面的准确率分别为30%(6/20)、50%(10/20)和38%(23/60)。结论 构建的药学智能问答平台能够根据临床提问
精准检索并提炼本地知识库信息,能避免AI幻觉的出现,为医务人员提供可靠的用药决策支持。
关键词 药学智能问答平台;AI幻觉;大型语言模型;DeepSeek;人工智能
Construction and efficacy verification of an intelligent pharmaceutical Q&A platform based on AI
hallucination-suppression
1
2
1
1
1
WEN Zhengwang ,WANG Jiaying ,YANG Wenyue ,YANG Haoyu ,MA Xiao ,LIU Yun (1. Dept. of
1
Pharmacy, Handan First Hospital, Hebei Handan 056003, China;2. Neusoft Group Co., Ltd., Hebei Handan
056003, China)
ABSTRACT OBJECTIVE To construct an intelligent pharmaceutical Q&A platform for precision medication with low “artificial
intelligence (AI) hallucination”, aiming to enhance the accuracy, consistency, and traceability of medication consultations.
METHODS Medication package inserts were batch-processed and converted into structured data through Python programming to
build a local pharmaceutical knowledge base. The retrieval and question-answering processes were designed based on large language
models, and system integration and localized deployment were completed on Dify platform. By designing typical clinical medication
questions and comparing the output of the intelligent pharmaceutical Q&A platform with the online version of DeepSeek across
dimensions such as peak time retrieval, half-life, and dosage adjustment reasoning for patients with renal impairment, the accuracy
and reliability of its retrieval and reasoning results were evaluated. RESULTS The intelligent pharmaceutical Q&A platform,
constructed based on local drug package inserts, achieved 100% accuracy in retrieval and reasoning for peak time, half-life, and
dosage adjustment schemes. In comparison, the online version of DeepSeek demonstrated accuracies of 30%(6/20), 50%(10/20),
and 38%(23/60) across these three dimensions, respectively. CONCLUSIONS The constructed intelligent pharmaceutical Q&A
platform is capable of accurately retrieving and extracting information from the local knowledge base based on clinical inquiries,
thereby avoiding the occurrence of AI hallucinations and providing reliable medication decision support for healthcare professionals.
KEYWORDS intelligent pharmaceutical Q&A platform; AI hallucination; large language models; DeepSeek; artificial intelligence
近年来,随着“互联网+”与医疗大数据的快速发展, 处理和医学语义理解能力,在用药咨询、健康问答等任
人工智能(artificial intelligence,AI)技术在医药服务领 务中显示出一定潜力,可在一定程度上缓解基层医疗药
[1]
域的应用不断拓展 。ChatGPT、DeepSeek 等大型语言 学人力不足的问题。然而,通用型 LLM 在专业药学场
模型(large language models,LLM)具备较强的自然语言 景中的可靠性仍面临较大挑战。多项医学AI研究指出,
Δ 基金项目 中国青年创业就业基金会中国青年医学创新研究科 LLM 在处理药物相互作用、特殊人群用药、个体化剂量
[2]
研课题(No.P250320108853) 调整等高风险任务时,存在风险管控相对不足等情况 ,
*第一作者 副主任药师。研究方向:人工智能药学。E-mail:
易产生“AI幻觉”,即生成与事实不符或缺乏证据支持的
zhengwangwen@krae.edu.kg
内容,此类错误可能导致剂量偏差、禁忌遗漏或不当建
# 通信作者 副主任医师,硕士。研究方向:消化道恶性肿瘤、人工
智能药学。E-mail:bsg_ly@hotmail.com 议,对患者安全构成严重威胁。
· 226 · China Pharmacy 2026 Vol. 37 No. 2 中国药房 2026年第37卷第2期

