An Improved Heterogeneous Graph Convolutional Network for Inter-relational Medicine Representation Learning

Medicine representation learning, which aims at uncovering hidden medicine relationships has emerged as a significant technique to imitate a doctor’s cognitive reasoning process. The majority of present research focuses on the intuitive relationships between medication and diagnosis, however, ignores the inherent properties of medicines. This study uses a heterogeneous graph convolutional network (HGCN) and a spectral clustering (SC) algorithm to investigate the associated knowledge underlying clinical treatment. Based on the chronic obstructive pulmonary disease (COPD) clinical data, we construct a medicine-property heterogeneous network consisting of two types of nodes involving medicines and their properties, and three types of edges referring to the intermedicine, interproperty, and medicine–property relations. HGCN is used to aggregate the neighbor nodes’ information and then generalize the medicine and property embeddings. Then, SC is leveraged to divide these embeddings into the syndromes to which they belong. To verify the model performance, a series of experiments referring to the baseline comparison, ablation study, and parameter sensitivity test have been carried out. Compared to three baseline models and their variants on six evaluation metrics, the experimental results demonstrate that the HGCN-SC model outperforms the baseline approaches in medicine combination identification and has around 3.0% improvement in accuracy over the SC.