The rapid accumulation of electronic health records (EHRs) and the advancements in data analysis technology have laid the foundation for research and clinical decision-making in the healthcare community. Graph neural networks (GNNs), a deep learning model family for graph embedding representations, have been widely used in the field of smart healthcare. However, traditional GNNs rely on the basic assumption that the graph structure extracted from the complex interactions among the EHRs must be a real topology. Noisy connections or false topology in the graph structure leads to inefficient disease prediction. We devise a new model named PM-GSL to improve diabetes clinical assistant diagnosis based on patient multi-relational graph structure learning. Specifically, we first build a patient multi-relational graph based on patient demographics, diagnostic information, laboratory tests, and complex interactions between medicines in EHRs. Second, to fully consider the heterogeneity of the patient multi-relational graph, we consider the node characteristics and the higher-order semantics of nodes. Thus, three candidate graphs are generated in the PM-GSL model: original subgraph, overall feature graph, and higher-order semantic graph. Finally, we fuse the three candidate graphs into a new heterogeneous graph and jointly optimize the graph structure with GNNs in the disease prediction task. The experimental results indicate that PM-GSL outperforms other state-of-the-art models in diabetes clinical assistant diagnosis tasks.
Keywords: diabetes assistant diagnosis; graph neural networks; graph structure learning; patient multi-relational graph.