基于多层前馈神经网络的中药药性量化研究

Quantitative study on medicinal properties of traditional Chinese medicine based on BP neural network

目的为解决中药药性描述的抽象、模糊导致难以准确把握其本质特性的问题,提出一种基于多层前馈神经网络(BP神经网络)的药向量训练(quantitative model of traditional Chinese medicine’s properties based on BP neural network,QM-BP)模型,实现中药药性的量化表示。方法首先对中药及其对应的功效进行整理,获得"中药-功效"样本对;其次,构建"中药-药向量-功效"3层结构的QM-BP模型,并利用中药的药性数据对模型进行初始化;最后,基于QM-BP模型使用"中药-功效"样本进行训练,得到BP药向量。结果将《中药学》教材所涉及的474味中药及其528个功效基于QM-BP模型训练并结合临床分析,发现训练后得到的BP药向量比药性的初始量化值更能反映中药的属性特征。此外,由于BP药向量与词向量具有相似的性质,发现功效相似的药物对应的BP药向量在欧几里得距离中距离较近,而功效差异较大的中药药向量在欧几里得距离中距离较远。结论利用BP神经网络构建药向量训练模型,在中药药性与功效具有关联性的基础上,对药性量化值进行修正,以期使药性量化值更精确。今后可优化QM-BP模型并开展药对、复方分析,以期探明中药药性及组方配伍中蕴藏的内在规律。

Objective It is difficult to accurately grasp the essential characteristics of medicinal properties of traditional Chinese medicine due to the abstraction and vagueness. This paper proposes a Quantitative Model of Traditional Chinese Medicine’s Properties based on BP Neural Network(QM-BP Model) to train and realize quantitative representations of Chinese herbal medicine(CHM). Methods Data for analysis were obtained and organized by conceptual analysis. Sample pairs of the associations were obtained based on the relationships of CHM and their efficacy. Then a QM-BP model with three-tier structure in form of CHM-drug vector-efficacy was constructed, initialized and trained according to prior organized CHM data. Finally, rules of correlation of CHM and their efficacy was obtained by training dataset with drug vectors representing quantitative attributes of CHM. Results Based on the training of QM-BP model, 474 TCM and 528 effects included in the textbook of TCM were trained and combined based on the training of QM-BP model. It was found that the BP drug vectors representing drug properties after training reflected the attribute characteristics of CHM better than the initial quantitative values. In addition, as BP drug vector and word vector have similar properties, the BP drug vectors for CHM with similar efficacy was relatively close in Euclidean distance while the CHM with different efficacies were relatively far in Euclidean distance. Conclusion In this paper, a BP neural network was adopted to construct a medicine vector training model. Based on the correlation between the medicinal properties and efficacy of TCM, the quantified values of the medicinal properties were modified to represent medicinal properties more accurately. In future work, the QM-BP model can be applied to the analysis of herb pairs and prescriptions to analyze the rules of combination related to medicinal properties and the compatibility within prescriptions.