Abstract: Standard clinical pathways play an important role in standardizing the treatment process and improving the therapeutic effect.However,current clinical pathways are designed for all patients of a disease,which cannot reflect the personalized information of patients or hospitals.To realize the personalized clinical pathways that meet the characteristics of patients and hospitals,the clinical pathway of drug treatment from prescription data recorded in medical information systems was mined.The prescription data and DrugBank database were used to generate the daily drug efficacy documents of patients.The token-bigram Latent Dirichlet Allocation (LDA) model was used to cluster these documents.The therapeutic topics were obtained at the same time.Finally,the probabilistic suffix tree model was trained as the clinical path of drug treatment by taking the patients' topic sequence of drug efficacy as input.The model could assist experts developing personalized clinical pathways and could be used for recommending follow-up medication as well.By taking the prescription data of pneumonia patients in MIMIC-Ⅲ database as an example,the feasibility and validity of the proposed method was verified.

Key words: process mining, token-bigram latent Dirichlet allocation model, probabilistic suffix tree, clinical path

摘要： 标准临床路径对于规范治疗流程、提高治疗效果具有重要作用,但当前的临床路径是面向同一病种的所有患者制定的,无法体现患者或者医疗部门的个性化信息。为了实现符合患者和医疗部门特点的个性化临床路径,从医疗信息化系统中记录的患者处方数据出发,进行药物治疗临床路径的挖掘。首先由处方数据结合DrugBank数据库生成患者的每日用药疗效文档;然后使用词对隐狄利克雷分布模型对这些药物疗效文档进行主题聚类,得到患者每日所用药物对应的疗效主题;最后以各个患者的药物疗效主题序列为输入,训练概率后缀树模型作为药物治疗的临床路径模型,该模型既可以辅助专家进行个性化临床路径的制定,也可以用于患者后续服用药物的推荐。以MIMIC-Ⅲ数据库中肺炎患者的处方数据为实例,对所提方法的可行性和有效性进行了验证。

关键词: 过程挖掘, 词对隐狄利克雷分布模型, 概率后缀树, 临床路径