Treatment plans are far more complex than diagnosis. Much of it is convincing the patient to make the right choice. An AI that can sympathize and educate is a long time coming.
From what I've heard, patients are often more aware of their own bodies than people give them credit for. Ie. If someone thinks they're dying, you should believe them.
Having another human there to reassure a scared patient that a treatment plan gives them their best chance is invaluable.
Also I'd be worried about pharmaceutical companies having an influence on the AI - as in making sure it doesn't know about generics and having it aggressively push their new drugs for any applicable cases. A human would be much more resilient to such machinations, on average.
Additionally, trying to elicit the information you need to make a diagnosis is much harder than non-clinicians realise. Present a series of yes no questions doesn't really capture all of the information that is required; I suspect this is partly why previous expert systems have failed thus far.
From the moment the patient walks into the room you are assessing their clothing, their colour, their gait, their physique, how they're breathing, their emotional state, their physical abilities.
When you are discussing their problem you are varying not only the content of the questions you ask but also the style and the wording to match the patient's level of education, understanding, local culture and terminology etc.
For certain conditions patient's will lead you completely astray if you take what they are saying as canon. There is an art to human conversation and the best doctors I know excel at this as well as having the depth and range of knowledge to synthesise a diagnosis.
After a history you will typically examine the patient. Listen to their heart and lungs, palpate their abdomen, look at their retina or their eardrum, etc.
The diagnosis is made by gathering all of this information together and then running a "pattern matching routine".
I have no doubt that machine learning software is coming for us, but the job that doctors too is often oversimplified by techies. Myself included. As a programmer and electronic engineer before starting medicine I thought that it would be trivial to replicate much of the work in software. It isn't. It will happen but after another couple of orders of magnitude improvement.
In the meantime, using machine learning for diagnostic aids is hugely valuable. The sort of things we are seeing at the moment are very much the "low hanging fruit" of medicine - that is, tasks for which the input, analysis and output is well defined and easy to feed into a computer algorithm. Examples include analysing histology specimens for cancer, analysing head CT scans for stroke, analysing cardiac monitoring for patients at high risk of heart attack, and so on. The "difficult" part is the data gathering.
As an aside, I do think the role of doctors has changed over time. Medicine has advanced enormously as a field over the last hundred years. Much of what used to be the role of the doctor is now the role of nurses, nurse specialists, nurse practitioners or even physician assistants. Yet still (in the UK at least) doctors have a never ending amount of work to do!
I'm very excited to be at the crossroads of medicine and technology and upbeat about what the future will bring for patient care.
