>In an old AI course the teacher talked about a medical diagnostic program that used relatively simple forward and backward chaining techniques. After years of development it did a pretty good job. I never understood why it wasn't used, except perhaps that most problems are relatively simple.
That would most likely be MYCIN.
Expert physicians would have a correct diagnosis rate of about 80%.
MYCIN had a correct diagnosis rate of 65%.
Non-expert physicians had a correct diagnosis rate near 40%.
Folks were more worried about liability, so they never released it.
If you ever used CLIPS (or jess), it will look very familiar, with rules like:
if (site culture is blood )
(gram organism is neg )
(morphl organism is rod )
(burn patient is serious)
(identity organism is pseudomonas )
which sort of translates into:
IF the culture came from blood,
AND the organism is gram-negative
AND the organism is rod shaped
AND the patient has serious burns
THEN it is 40% likely that the organism is pseudomonas.