Medical diagnosis is a trivial problem for machine learning to beat humans. Humans are terrible at this.
The problem is there is not even the concept of high frequency medical data. Imagine machine learning in quant stock trading with samples once a year. Of course it isn't going to work.
The problem is all in the externalities. Doctors don't want AI. They can see the automation path and their bank account change down the line quite clearly. Not to mention most doctors don't know anything about data science so how can you have any faith in the algorithm prediction? No one wants to be the test case and then get a law suit. "I was just following what the computer said was correct". Wrong answer, pay up.
The real irony to me is in a 100 years people look at the current medical system as complete quakery. Literally have everything right now to build medical super intelligence but stuck with the human doctors doing the exact same things from 50 years ago.