Predicting Hospital Readmission Using Graph Representation Learning Based on Patient and Disease Bipartite Graph

Accurate hospital readmission prediction is conducive to reducing medical waste, improving the quality and efficiency of public health services, and providing better medical services for more people. The readmission of each patient is closely related to their disease history. Therefore, it is of great help to accurately predict the readmission by using the patient’s diagnosis history information. However, the diagnosis history of some patients may be very short, and it is difficult to use the features of individual patients to predict their readmission. In this paper, a hospital readmission prediction model based on patient and disease bipartite graph, PDGraph, is proposed. In this method, heterogeneous graph is used to establish the correlations between patients and diseases, which can express the historical disease information of patients and the latent relationships between patients with the same disease. By constructing the bipartite graph of patients and diseases, one patient establishes an indirect relationship with patients with the same diseases through disease nodes. Thus, the features of other related patients can be used to assist the hospital readmission prediction and improve the prediction effect. Then, PDGraph embedding generation algorithm is designed to aggregate the information of disease and related patients to each patient to improve the predictive performance. Our proposed model was tested on a real dataset, and the results show that the proposed method is more accurate in the prediction task than baselines.