基于属性知识嵌入的LDA模型在中药推荐中的应用

Application of Latent Dirichlet Allocation model based on attribute knowledge embedding in traditional Chinese medicine recommendation

目的:数据驱动的药材推荐方法帮助中医医师在真实的临床实践中更精确、更智能地制定科学的治疗处方,也可以为中医诊断和治疗的发展提供科学依据。方法:通过文本挖掘方法分析了24127条中医处方记录,在中医理论的基础上模拟生成处方的过程,并将症状和药材的丰富信息及其相互关系等领域知识纳入考虑。提出了一种基于属性知识网络嵌入的LDA(Latent Dirichlet Allocation)主题模型的中药推荐方法,中药的属性知识网络包含药材丰富的属性信息以及蕴含的药理作用,对主题模型进行了增强。结果:研究结果表明,在最佳嵌入系数下,模型的预测困惑度、准确性以及平均AUC相较于基线主题模型均有更好的表现。结论:所提出的方法有利于提升模型稳定性、药材推荐准确度,更好地承担诊疗模式挖掘等任务。

Objective:A data-driven medicinal materials recommendation method helps traditional Chinese medicine(TCM)physicians to make scientific treatment prescriptions more accurately and intelligently in real clinical practice,and can also provide a scientific basis for the development of TCM diagnosis and treatment.Methods:24127 TCM prescription records were analyzed by text mining method,and the process of generating prescriptions was simulated based on TCM theory,and the knowledge of the rich information of symptoms and medicinal materials and their interrelationships was taken into account.A TCM recommendation method based on Latent Dirichlet Allocation(LDA)topic model embedded with attribute knowledge network was proposed,and the attribute knowledge network of TCM contains rich attribute information of medicinal materials and pharmacological effects,which enhanced the topic model.Results:The experimental results showed that the prediction perplexity,accuracy and average Area Under Curve(AUC)of the model performed better compared with those of the baseline topic model under the optimal embedding coefficients.Conclusion:The proposed method is beneficial in improving model stability and accuracy of medicinal materials recommendation,and better undertaks tasks such as diagnosis and treatment pattern mining.