Multi-channel hypergraph topic neural network for clinical treatment pattern mining

Recently, increasing attention has been paid to mining clinical treatment patterns from electronic medical records (EMRs), which provide physicians with explicit knowledge to guide the patient’s treatment. However, the current work mainly focuses on the topic probability models. These shallow models fail to consider the complex high-order correlations. Hypergraphs provide a natural way to establish complex high-order correlations, but their potential to discover the semantics of latent topic has not been fully explored. In this paper, we integrate the topic model in hypergraph learning and propose a multi-channel hypergraph topic neural network (C3-HGTNN) to discover latent topic treatment patterns with learning high-order correlations. Specifically, the hypergraph network is constructed based on the interactions in the treatment traces, which describe the latent high-order correlations. Then, based on the heterogeneity of the nodes in the hypergraph, we generate multiple channels and perform a convolution operation on each channel to fully encode the hypergraph. Finally, we develop an extensive representation of the treatment trace and activity by combining the embeddings from various channels and generating relevant topic distributions to identify latent treatment patterns. Also, we demonstrate that our model is a form of deep inference for traditional relational topic models, filling the gap in learning high-order correlation in traditional topic models. Extensive experiments on our dataset and public dataset show the effectiveness and superiority of our model compared to the competitive baselines.