The key to synthesizing the features of electronic medical records (EMR) big data and using them for specific medical purposes, such as mortality and phenotype prediction, is to integrate the individual medical event and the overall multivariate time series feature extraction automatically, as well as to alleviate data imbalance problems. This paper provides a general feature extraction method to reduce manual intervention and automatically process large-scale data. The processing uses two variational auto-encoders (VAEs) to automatically extract individual and global features. It avoids the well-known posterior collapse problem of Transformer VAE through a uniquely designed “proportional and stabilizing” mechanism and forms a unique means to alleviate the data imbalance problem. We conducted experiments using ICU-STAY patients’ data from the MIMIC-III database and compared them with the mainstream EMR time series processing methods. The results show that the method extracts visible and comprehensive features, alleviates data imbalance problems and improves the accuracy in specific predicting tasks.