基于数据挖掘技术分析艾迪注射液的不良反应

Analysis of ADR of Aidi Injection Based on Data Mining Technology

目的:采用数据挖掘技术研究艾迪注射液不良反应(ADR)发生特点及发生规律,为临床合理、安全使用中药注射剂提供借鉴。方法:收集近5年艾迪注射液ADR文献案例和国家ADR监测系统中某两家医院上报数据,利用IBM SPSS Modeler软件,综合运用决策树(C5.0)算法、贝叶斯网络、神经网络和关联规则Apriori算法,开展数据挖掘研究。结果:决策树(C5.0)算法结果显示,用药剂量是最关键的影响因素(重要性57.44%),其次是过敏史、溶媒用量、性别。贝叶斯网络分析结果显示,ADR累及器官受原患疾病影响最大(重要性29.84%),其次为用药剂量(重要性20.32%)。神经网络分析结果与贝叶斯网络基本吻合,原患疾病是最重要影响因素(重要性25.31%),其次为年龄(重要性16.39%),再次是用药剂量(重要性15.33%)。关联规则Apriori算法置信度最高的规则显示,对无过敏史的男性患者,以250 mL的5%葡萄糖为溶媒滴注艾迪注射液,发生ADR的例数为24例,占比13.41%,其中54.17%的案例表现为皮肤及附件损害。结论:挖掘结果准确率偏低,其中贝叶斯网络正确率最高,为62.57%,决策树(C5.0)算法正确率为39.66%,神经网络“ADR累及器官分类总体正确率”为32.4%。关联规则Apriori算法得到的6条规则支持度均在10%以上,且提升度均大于1,关联分析具有意义,但置信度均低于60%,条件概率偏低。整体挖掘结果不够理想,可能与样本量有限、临床用药复杂性相关,尚需大样本量数据及最优的挖掘算法来研究和验证。

Objective:Using data mining technology to study the ADR characteristics and regularrityof Aidi injection,to provide reference for rational and safe use of TCM injections in clinic.Methods:The ADR literature cases of Aidi injection and the data reported by two hospitals in the national ADR monitoring system in recent 5 years were collected,with IBM SPSS Modeler software,adopting Decision tree(C5.0)algorithm,Bayes network,Neural network and association rule Apriori algorithm,combining these ways to carry out data mining research.Results:Decision tree(C5.0)algorithm indicates dosage was the key factor(importance 57.44%),followed by history of allergic,quantity of solvent and gender.Bayes network analysis revealed that damaged organ caused by ADR mostly involved with original disease(importance 29.84%),dosage was the second(importance 20.32%).Result of neural network analysis agree with Bayes network basically,original disease was the major factor(importance 25.31%),next element was age(importance 16.39%),then dosage(importance 15.33%).With association rule Apriori algorithm,rule with the highest confidence,showed that there were 24 ADR cases occurred in male patients with no allergic history after taking the medicine,using glucose(250 mL,5%)as a solvent,the ratio was 13.41%,inside proportions of lesions of skin and appendages was 54.17%.Conclusion:The accuracy rate of data mining was relatively low,the precision of Bayes network was 62.57%,the highest.Decision tree(C5.0)was 39.66%.Adopting neural network,the whole precision of damaged organ category caused by ADR was 32.4%.Six rules got through association rule Apriori algorithm,all their support were higher than 10%,promotion were higher than 1,association analysis was meaningful,but confidence rate lower than 60%,conditional probability was relatively low either.Outcome of data mining was not ideal in general,the result may be related with limited sample size and complexity of clinical medication.It needs more sample data and optimized algorithm to study and verify.