With the wide application of electronic health records (EHR) in healthcare\nfacilities, health event prediction with deep learning has gained more and more\nattention. A common feature of EHR data used for deep-learning-based\npredictions is historical diagnoses. Existing work mainly regards a diagnosis\nas an independent disease and does not consider clinical relations among\ndiseases in a visit. Many machine learning approaches assume disease\nrepresentations are static in different visits of a patient. However, in real\npractice, multiple diseases that are frequently diagnosed at the same time\nreflect hidden patterns that are conducive to prognosis. Moreover, the\ndevelopment of a disease is not static since some diseases can emerge or\ndisappear and show various symptoms in different visits of a patient. To\neffectively utilize this combinational disease information and explore the\ndynamics of diseases, we propose a novel context-aware learning framework using\ntransition functions on dynamic disease graphs. Specifically, we construct a\nglobal disease co-occurrence graph with multiple node properties for disease\ncombinations. We design dynamic subgraphs for each patient's visit to leverage\nglobal and local contexts. We further define three diagnosis roles in each\nvisit based on the variation of node properties to model disease transition\nprocesses. Experimental results on two real-world EHR datasets show that the\nproposed model outperforms state of the art in predicting health events.\n