Figure 7 | Scientific Reports

Figure 7

From: Novel architecture for gated recurrent unit autoencoder trained on time series from electronic health records enables detection of ICU patient subgroups

Figure 7

Neural network model. The model consists of an encoder and a decoder part, but these parts are trained jointly. Neural network layers are shown using green boxes with round corners; blue boxes represent tensors, and vector operations like concatenation are shown as yellow boxes with dashed lines. The GRU block is part of the encoder and the decoder and consists of a bidirectional GRU layer, whose outputs are summed for each time step. In the encoder, the resulting features are average pooled over time and then processed using a dense layer, resulting in the encoder’s output: the bottleneck. The decoder repeats the low-dimensional representation in time and concatenates it with the original input’s time features. The same GRU block architecture as for the encoder is used to calculate the temporal dynamics of the time series. Weights are not shared between the two GRU blocks in the model. As a final step, a dense layer is applied to each time step to reconstruct the original features. The model is trained to predict its original input.

Back to article page