Our technique is based on the assumption that accurate and early prediction of sepsis requires to be able to predict the evolution of the vital signs. This idea was translated in the use of a recurrent neural network, a Long Short-term memory (LSTM) network, which was trained to accomplish two tasks: the prediction of (i) sepsis and (ii) the vital signs at time t+6. This LSTM can be interpreted as a State-Space model, where the hidden states (the cells of the LSTM) can detect the outcome, but also can help to predict the evolution of the system (That is where the idea diverge from state-space models, as the cells cannot predict their own evolution but that of the vital signs). We assume that the use of this auxiliary task allows for a better training of the network given the low prevalence of sepsis. This LSTM is fed with the eight vital signs as inputs, and is composed of five layers of 128 cells, followed by layers of fully connected networks interleaved with Batch Normalization and Relu activation layers. The current best network achieved a utility score of 0.31 during a cross-fold validation on the training set. We will be working on optimizing the network architecture and optimizing its different hyper-parameters. A next step will also consist in incorporating the information provided by the demographic parameters, but also more importantly the laboratory values; as these parameters have not yet been considered in our technique.