Over 80% of the job could be done by properly designed algorithms - and 80% would be a conservative guess. The "interesting" cases are very rare, and mostly consist in easy-to-do diagnosis of unfrequent illnesses (I once diagnosed an insulinoma! hurrah!)
Just look at the growing number of tasks given to non-MD- say RN. It's a trend that will go on.
Human knowledge is nothing but the application of algorithms - deciding which symptoms are important, which are to be discarded, which require more investigation, etc.
Education is building algorithms. Experience is machine learning. Wisdom is statistics. Instinct is prioritization.
The faster we can implement all that in software, the better.
I think that even rare disease diagnosis can be done better by algorithms by using Isabel healthcare's diagnostic system[1] or Watson(in the future).
The place where a human will contribute the most is the interpersonal stuff:motivating people, digging data from people(although computers compete there[2]) and administering placebo.
And of course doing the physical stuff.
But that seems very different from the job of a doctor and closer to the job of a nurse.
[1]http://www.isabelhealthcare.com/home/default [2]There was some research that showed that people give more personal information to a computed form than via face to face chat with a doc.
Of course, it should never be top of the list - unless it becomes a possibility given the accumulation of negative results of other tests for more common diseases, or the positive results of DNA sequencing.
Unfortunately, that's not the way it will happen with an human : rare almost means "ruled out" simply because "poorly taught" or even unknown, and for the best of us "get a second opinion with xxx who specializes in this very disease".
DNA sequencing is at most used for a couple of diseases per specialty, and data mining/fishing expeditions are frowned upon (maybe rightly so given the human bias of trying to find meaning and correlation in what may be unrelated).
I strongly believe the beginning of true e-medicine will make us look at the current medicine the way we look back at blood letting, holy waters, the four humors imbalances etc.
I see many topics where a proper mix of IT and medicine could make a real difference - for example, improving WorldVista, adding IA agents to mine the data and suggest things.
Then I see the sorry thing that calls itself Health-IT and which is just a new way to put government stimulus money into dull uses and overpriced software.
I'm the founder, lead developer and clinical researcher behind StethoCloud - http://www.stethocloud.com/
I find being a clinician helps me understand the problem when it comes to the diagnosis of pneumonia in children. I understand the scope of the problem, its current diagnostic hurdles and as a hacker, I know how to build the app and cloud backend for the project.
Having exposure and knowledge in other fields of computer science, mathematics and engineering helps too. I'm glad I took the Stanford AI and Machine Learning courses that gave me a good understanding of what machine learning techniques are capable of doing given sufficient data to train on.
I would say having knowledge of those fields was important in order for me to articulate the problems clinicians face with the diagnosis of pneumonia in children to data scientists in a way that they could understand how their skills could be helpful and thus convince them to help out with this project.
Fundamentally, the brain processes information to make decisions, which machines can always in principle be designed to do. Thus as long is medicine is fundamentally about making diagnostic and treatment decisions, it can be solved by machines.