Figure 1.

The proposed data-driven subtyping method. (A) Illustration of our LSTM recurrent neural network. The patient representation derived by recurrent hidden layer. Raw patient multi-source data are pre-processed by imputation. For each patient, the merged temporal records are set as input of LSTM corresponding each timestamp. The targets are a set of disjoint temporal records for each patient and obtained by the same pre-process method. There are two kinds of targets function for continuous and binary values separately. Representations generated by all the hidden states are used in patient subtyping. (B) Overall flow of the proposed LSTM-based method.