I thought expert systems were supposed to replace diagnosis, like three decades ago.
Most notably, he suggested we separate diagnosticians and prognosticators from practicians. The former do their best to guess what you have, send you to the right kind of practician, and estimate your possible outcomes (with probability distribution and Bayesian goodness and all). Then whatever practician you chose do their best to treat you.
Separating prognosis from practice allows us to have a system where prognosticators are rewarded for their accuracy and calibration, and practicians are rewarded for their outcomes relative to prognosis. Being good doesn't mean having better outcomes, being good means having better outcomes than predicted.
I'm not sure how well this system would work in practice, and there may be a huge problem with eliminating the short feedback loops we can have between diagnosis and practice when they're the same person. What should be a nice separation of concern may turn into a communication nightmare.
Nevertheless, it is possible that specialised diagnosticians may be better than any expert system.