计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (8): 2130-2139.DOI: 10.3778/j.issn.1673-9418.2309100

• 人工智能·模式识别 • 上一篇    下一篇

基于对话结构与图注意力网络的药物推荐算法

陈江美,张文德,谭睿璞   

  1. 1. 福州大学 经济与管理学院,福州 350108
    2. 福州大学 信息管理研究所,福州 350108
    3. 福建江夏学院 电子信息科学学院,福州 350108
  • 出版日期:2024-08-01 发布日期:2024-07-29

Medication Recommendation Algorithm Based on Dialogue Structure and Graph Attention Network

CHEN Jiangmei, ZHANG Wende, TAN Ruipu   

  1. 1. School of Economics & Management, Fuzhou University, Fuzhou 350108, China
    2. Institute of Information Management, Fuzhou University, Fuzhou 350108, China
    3. College of Electronics and Information Science, Fujian Jiangxia University, Fuzhou 350108, China
  • Online:2024-08-01 Published:2024-07-29

摘要: 现有的药物推荐算法大多基于历史电子健康记录,但该数据难以反映患者的当前健康状况,也无法捕捉患者实时性的健康需求,导致推荐效果不佳。为此,融合了在线对话和疾病信息,提出一种基于对话结构与图注意力网络的药物推荐算法。集成灰关联分析与图注意力网络,运用灰关联分析学习节点间的关联,提出了一种新的关联感知图结构,以弥补传统图网络难以捕捉节点关联的不足。构建了对话分层编码器,基于新的图注意力网络编码话语与对话表示,并设计两种关联图结构学习节点的邻接关系,以生成蕴含上下文语义的对话结构表示。基于知识图谱和新的图网络学习疾病表示,将其与对话表示融合,实现药物的预测与推荐。实验结果表明,提出的算法在各评估方法下均优于基线方法,与性能最好的基线DNN相比,提出算法的F1和Jaccard分别提高了1.8%和3.5%,表明了提出算法能有效提高推荐性能。

关键词: 药物推荐, 对话结构, 图注意力网络, 灰关联分析

Abstract: Existing studies on medication recommendation are mostly based on the electronic health records. However, these data are difficult to reflect the current health status of patients and learn the current health needs of patients. Thus, this paper fuses the dialogue and disease information, and proposes a dialogue-based structure and graph attention network recommendation algorithm. Firstly, the grey relational analysis and graph attention network are integrated and the relations of nodes are measured using grey relational analysis to provide a novel relation-aware graph structure. It can improve the traditional graph networks to learn the node relations. Secondly, a dialogue hierarchical encoder is established to encode the representations of utterance and dialogue via the new graph structure. Meanwhile, two graph structures are designed to learn the node correlation for generating the contextual dialogue representation. Finally, the representations of disease by knowledge graph and graph network are incorporated with dialogue to predict and recommend. Experimental results show that the proposed algorithm is superior to baselines in terms of all metrics. Compared with the best baseline DNN, the performance of the proposed algorithm improves 1.8% and 3.5% in terms of F1 and Jaccard, which shows that the proposed algorithm can improve the recommendation performance.

Key words: medication recommendation, dialogue structure, graph attention network, grey relational analysis