Graph Representation Learning For Medical Complication Prediction

With the advent of the era of medical big data,artificial intelligence-based Clinical Decision Support Systems(CDSS)are constructed to assist doctors in making clinical decisions.However,the treatment plans designed by CDSS are often unpractical for the difficulty of describing complication relations among diseases.In fact,revealing the potential complication relations is not only the important prerequisite for ensuring the effectiveness of treatment plans,but also the theoretical basis for prevention and control of diseases.Traditional methods are usually based on medical experiments and causal inference to find the latent complications despite the iterative trial-and-error processes.Recently,with the advances of graph representation learning techniques,it is promising to predict medical complications in properly constructed bioinformatics graphs.However,diseases often have heterogeneous attributes to reflect different biomedical characteristics,which contribute distinctly to the discovery of unknown complication relations.Moreover,due to the sparsity and terminology of disease attributes,it is challenging to measure the similarity between attributes of different disease nodes,which seriously interferes the medical complication prediction task.Firstly,for the challenges of heterogeneity and sparsity,this paper proposes a novel complication Complex Attributed Network Embedding framework(CANE)to jointly model the heterogeneous attributes and neighborhood of disease nodes.The core assumption of CANE is that the diseases with complication relations should have similar attributes or graph structures,so their representations should be more similar.To this end,we first obtain the similarity semantical embeddings for biomedical entities via matrix factorization and extract domain semantics in biomedical texts via dynamic word representation.Then,the convolutional neural fusion strategy and multiple attention mechanisms are proposed to fuse the heterogeneous attributes.Along this line,we introduce aggregation functions to explicitly enhance the representations of diseases with sampled neighborhood features.By anchoring the input of the aggregation functions as the fusion representation of attributes,CANE can provide structural guide for the fusion of heterogeneous attributes and realize the co-training of different modules.In this way,the critical attributed and neighborhood information is preserved into the representations of diseases,which are applied for complication prediction task next.Secondly,for the terminology of disease attributes,this paper proposes a novel Knowledge Powered Cooperative Semantic Fusion framework(KCSF)to capture deeper knowledge semantics for disease attribute representation.In reality,biomedical texts contained in disease attributes are full of scientific entities and domain knowledge,which can provide additional distinguishable semantics for enhancing the representations of diseases.However,existing works seldom consider such unique property of biomedical texts,which reduces their performance on medical complication prediction task.To this end,we first exploit knowledge graph technology to capture related entity semantics for biomedical texts,and then contextually model pure text semantics and entity semantics via sequence encoders.Along this line,mutual attention mechanism between entities and texts is designed to emphasize the crucial semantics of entities with the guide of texts,and vice versa.In this way,we achieve the mutual enhancement and fusion of the two types of semantics,which leads to more informative disease attribute representation.Furthermore,we reconstruct CANE with KCSF to encode more crucial biomedical semantics into the final representations of diseases.Lastly,comprehensive experiments on real world complication data not only clearly validate the effectiveness of CANE and KCSF but also demonstrate the potential of graph representation learning for medical complication prediction.