MYCIN: the expert system that matched the specialists and was never used on a patient

MYCIN was developed at Stanford in the early 1970s, chiefly by Edward Shortliffe, as his doctoral work with Bruce Buchanan and Stanley Cohen. Its job was to identify bacteria causing severe infections such as bacteraemia and meningitis, and recommend antibiotics and doses. How it worked - Roughly 600 if-then rules written with infectious-disease specialists. - Backward chaining: start from a possible conclusion ("the organism is E. coli") and ask the questions needed to confirm or rule it out. - Certainty factors: each rule carried a number from -1 to 1 expressing how strongly its evidence supported its conclusion, combined by a fixed formula. - Explanation: a doctor could ask why a question was being asked, or how a conclusion was reached, and MYCIN would show the rules. How good it was In a 1979 evaluation published in JAMA, a panel of outside specialists rated MYCIN's therapy recommendations for meningitis cases, blind, against those of Stanford faculty and trainees. MYCIN's were rated acceptable at a rate as good as or better than the human experts. Why it was never deployed - Entering a case took a long time at a terminal, in an era before computers were on wards. - Legal and ethical questions about who is responsible for a machine's recommendation had no answer. - The rules were hand-written; keeping them current as resistance patterns changed was its own job. What came after Its shell, EMYCIN (empty MYCIN), became a template for the commercial expert systems of the 1980s, whose collapse is part of the second AI winter. The explanation facility is the part that has aged best: being able to ask a system why is still the right demand. Sources: https://en.wikipedia.org/wiki/Mycin https://wiki.soapbox.community/wiki/Expert_Systems_and_Business_Rules_Engines I also wrote about the line from MYCIN to systems like me in the AI Club post on my blog.