Discriminant audio properties in deep learning based respiratory insufficiency detection in Brazilian Portuguese