Learning to Exploit Invariances in Clinical Time-Series Data using Sequence Transformer Networks