Making deep neural networks work for medical audio: representation, compression and domain adaptation