Continuous monitoring and prediction of Length of Stay (LoS) for critically ill patients admitted to Intensive Care Units (ICU) can help anticipate on their needs, thereby reducing risks of adverse events occurring, efficiently allocate resources and improve care. In this work, four different sequence-to-sequence deep learning models (Long-Short Term Memory, Gated Recurrent Units, Temporal Convolution Network and the Transformer network) were implemented to predict hourly remaining LoS, using 271 input features extracted from 20,489 ICU stays. Input sequences for model training were created using two approaches: (i) by stacking data one hour at a time and (ii) by using a sliding window of length $l$ shifted one hour at a time. To enhance model explainability, both the models’ latent space visualization and feature importance ranking were used. Compared against baseline models, the sequential models reported the best performance, reducing the prediction error by almost 3 days. Moreover, the sliding window approach showed improvement both in terms of performance and training time whilst requiring less resources. The latent space visualization of each model using t-SNE revealed that, with only 6 hours of data, the models are able to learn early clinical patterns that separate short ($0<\text{LoS}\leq3$ days) from medium ($3<\text{LoS}<7$ days) and long ($7\leq\text{LoS}<+\infty$ days) stays, and, the models are able to capture existing correlations between long LoS and risk of mortality, the latter not used during modeling. The benefit of this approach in a true hospital setting, besides LoS prediction, is that early clustering enables the early identification of long stays with shown existing risk of death, to whom prompt allocation of focus and adequate resources can be delivered.