Figure 4.

The change of feature space with the training process. In this feature space, each point represents an electronic medical record and different colors indicate different diseases. At the beginning of training (Epoch = 0), since the model parameters are randomly initialized, all the electronic medical records in the feature space are randomly distributed and indivisible. After 5 epoch, electronic medical records of different diseases began to have a trend of separation. After 10 epoch, the electronic medical records of all kinds of diseases have been separated, except for some areas and the edge of each category. When the training reaches 100 epoch, we can clearly see that the samples of each disease have been completely separated, and the electronic medical records of the same disease are also gathered together.