摘要
目的构建机器学习辅助的基于真实世界数据的药品有效性与安全性评价方法。方法构建包含数据收集、提取、治理及评价的方法学框架,创新性提出基于多维亚组分析的评价路径。针对病程记录等非结构文本,采用“标签-模型-复核”模式实现结构化转换;针对病例结局缺失数据,运用机器学习算法预测和填补;通过多因素分析方法筛选影响结局指标的关键因素,据此对患者进行多维亚组分析,在多个亚组内开展不同品种/剂型药品的评价。应用所建立的方法对坦度螺酮胶囊剂和片剂进行有效性和安全性评价。结果共纳入12265例住院患者和144483例门诊患者诊疗数据,焦虑程度预测模型的曲线下面积(AUC)均>0.9。治疗(30±10)和(60±20)d后,坦度螺酮胶囊有效率分别为93.40%和93.44%,片剂有效率分别为91.64%和92.87%。在905个亚组中,70.94%(642/905)的亚组胶囊剂有效率高于片剂,7.29%(66/905)的亚组片剂更优。安全性方面,胶囊剂的药品不良反应(ADR)发生率1.53%,片剂1.63%,差异无统计学意义。在439个亚组中,82.69%(363/439)亚组胶囊剂ADR发生率更低,仅3.87%(17/439)亚组片剂ADR发生率低于胶囊剂。结论利用机器学习技术在真实世界数据的清洗和结构化中具有显著优势。所建立的机器学习辅助基于真实世界数据的药品有效性和安全性评价方法,可精准识别不同特征人群中药物相对有效性和安全性差异,为真实世界证据支持的药品评价实践提供技术参考。
Objective To develop methods for evaluating drug effectiveness and safety based on real-world data.Methods A methodological framework encompassing data collection,extraction,governance,and evaluation for real world data(RWD)-based comprehensive clinical drug evaluation was constructed,with an innovative approach of multi-dimensional subgroup analysis proposed.For unstructured texts such as medical records,a“label-model-review”mode was adopted to convert them into structured data.Machine learning algorithms were applied to predict and impute missing case outcomes.Key factors influencing outcome indicators were identified through multi-factor analysis,and patients were stratified into multiple subgroups based on these factors to evaluate different drug varieties/formulations within each subgroup.The above methods were empirically applied to evaluate the efficacy and safety of tandospirone capsules and tablets.Results A total of 12265 inpatients and 144483 outpatients were included.The AUC values of the anxiety level prediction models all exceeded 0.9.The effective rates of tandospirone capsules and tablets at(30±10)days and(60±20)days of treatment were 93.40%vs.91.64%,93.44%vs.92.87%,respectively.Among 905 efficacy-evaluated subgroups,capsules demonstrated superior effectiveness to tablets in 642 subgroups(70.94%),with only 66 subgroups(7.29%)showing the inverse pattern.Safety analyses revealed the ADR incidence with capsules was 1.53%,which was lower than the 1.63%for tablets,with no statistically significant difference.This advantage persisted in 363/439 safety subgroups(82.69%),whereas tablets showed lower risk in merely 17 subgroups(3.87%).Conclusions The study demonstrates the significant advantages of applying machine learning and artificial intelligence technologies in the cleaning and structuring of real-world data.The developed machine learning-assisted evaluation framework enables precise identification of differences in the relative effectiveness and safety of drugs across patient subpopulations,providing a methodological reference for real-world evidence-based drug evaluation practices.
作者
邹林珂
刘馨宇
邓博
沈浩
侯正尧
高光洁
李梦婷
梁诗悦
温亚林
常欢
杨勇
龙恩武
李晋奇
吴行伟
ZOU Linke;LIU Xinyu;DENG Bo;SHEN Hao;HOU Zhengyao;GAO Guangjie;LI Mengting;LIANG Shiyue;WEN Yalin;CHANG Huan;YANG Yong;LONG Enwu;LI Jinqi;WU Xingwei(Department of Pharmacy,Sichuan Academy of Medical Sciences&Sichuan Provincial People's Hospital,Chengdu 610072,China;Personalized Drug Therapy Key Laboratory of Sichuan Provincial,School of Medicine,University of Electronic Science and Technology of China,Chengdu 610054,China)
出处
《医药导报》
2026年第3期534-542,共9页
Herald of Medicine
基金
国家自然科学基金面上项目(72174038)
四川省自然科学基金面上项目(2024NSFSC0598)
四川省卫生健康信息中心科研项目(2023HX044)
四川省卫健委临床研究普通项目(24LCYJPT03)
成都市科技局技术创新研发项目(2024-YF05-01192-SN)。
关键词
药品临床综合评价
真实世界数据
机器学习
亚组分析
坦度螺酮
Clinical comprehensive evaluation
Real-world data
Machine learning
Subgroup analysis
Tandospirone
