This study aims to generate medical text data that integrates spatiotemporal attributes to solve the problems of small size and low availability of open source text datasets due to the privacy of medical data. This paper optimizes the LeakGAN model by using a multi-layer LSTM structure and increasing the number of hidden and embedding layers to help generate networks that capture long-distance dependencies in data. In addition, we designed a spatiotemporal data generation method based on random algorithms and Markov chains to generate the spatiotemporal attributes of electronic medical records. In order to comprehensively evaluate the generation effect of the model, this paper evaluates from similarity-based indicators and likelihood-based indicators. The results show that the method designed in this article is 2.038 better than other models on NLL-oracle, 0.016 on NLL-test, and 0.0069 on EmbSim. The method proposed in this article has good ability to fit real data and can avoid the high cost and privacy leakage problems faced when manually constructing electronic medical record data sets by generating realistic medical electronic medical record synthetic data. This method also provides new solutions to the problem of insufficient data sets faced by more downstream applications (such as data auditing, anomaly detection, etc.).