I agree. Compassion is something that is hard to encode with a deep learning approach. However, having worked with genes and genetic mutations I can conclude that very few doctors can have a relatively complex map of nucleotide polymorphisms (mutations) in their head and decide in a matter of minutes if patient is at risk or not. The same applies to more superficial types of analyses with medical claims or lab tests.
Good algorithms and good data are a powerful match capable of putting doctors out of the loop sooner than we think.
Regulations and responsibilities are another matter, beyond the unreasonable effectiveness of artificial intelligence in more and more domains.