In order to solve the problems of the insufficient mining of relationship among inter-disease joint effects and readmission and the weak generalization ability of related models， a readmission prediction model based on graph contrastive learning was proposed， called HealthGraph. Firstly， the disease co-occurrence information in the dataset was used to construct a disease code map， so that the correlation information among diseases was fully explored. Then， a patient data augmentation method was proposed with the guidance of the idea of graph contrastive learning， and the topology related to the task was captured by the graph sampler adaptively， and a new view was constructed to improve the data richness， thereby improving generalization performance of the model. Finally， readmission prediction was carried out by combining the initial disease code map embedding and the new view embedding. The respiratory and circulatory system diseases datasets were constructed on real dataset MIMIC-Ⅲ and extensive experiments were conducted. The results show that compared with REverse Time AttentIoN model （RETAIN） and the Stage-aware neural Network model （StageNet）， the proposed model has the accuracy and F1 indicators improved by about 1 percentage point. In addition， results of two groups of ablation experiments verify the effectiveness of the proposed model in improving the accuracy and generalization of readmission prediction.