With the rapid growth of patient medical data, the challenges of enhancing the quality of medical data and ensuring patient privacy are becoming increasingly prominent. Conventional data anonymization methods often fail to effectively protect medical data while guaranteeing the exploitation of data value, prompting us to seek advanced techniques to balance between data analysis and real data protection. This study proposes to utilize the knowledge graph attention network to model the implicit structure of patient medical records, and generate virtual medical data. This approach also enhances the interpretability of the pattern learning and data generation process. The whole framework comprises classifying data fields, preprocessing the data, training knowledge network, generation of core data fields and generation of the rest data fields with a regression model. The proposed model is evaluated through Jensen-Shannon divergence and Wasserstein distance metrics. Preliminary results indicate that the generated patient medical data follows the pattern of original data. This study provides a new direction for the field of medical data generation.