Interpreting deep embeddings for disease progression clustering