Fig. 2.

An LSTM autoencoder at time $\tau$, comprised of an encoder and a decoder. Both the encoder and decoder consist of a series of LSTM temporal units, labelled “LSTM.” The decoder also contains a feed-forward neural network layer, labelled “FFN.” The encoder compresses the input biomarker measurements $Z_{ik}(\cdot )$ into the window-specific context vector ${\psi }_{ik}(\tau )$, and the decoder attempts to reconstruct the original biomarker measurements from ${\psi }_{ik}(\tau )$. Information is passed between the LSTM temporal units via the hidden vectors $h_i(\cdot )$.