What was MYCIN, and how did a 1970s program choose antibiotics as well as specialists?

MYCIN was an expert system built at Stanford in the 1970s that advised doctors on serious bacterial infections, first of the blood and later meningitis, using several hundred if-then rules written with infectious-disease specialists. In a blinded 1979 study, outside experts rated its antibiotic choices acceptable more often than those of Stanford faculty doctors, yet it was never put into routine clinical use.

What was MYCIN, and how did a 1970s program choose antibiotics as…

Imagine a doctor facing a patient with a dangerous infection. The lab has grown bacteria from a blood sample, but identifying them precisely will take a day or two. Treatment cannot wait. Which antibiotic should the doctor start now? In the early 1970s, a team at Stanford set out to capture how infectious-disease specialists answer that question, and to put their reasoning into a program any physician could consult. The result was MYCIN.

MYCIN became the textbook example of an expert system: a program that stores a specialist's knowledge as rules and applies them to a specific case. Its name is simply the common ending of many antibiotic names, chosen at the suggestion of one of the doctors on the team.

Who built it, and when

MYCIN grew out of DENDRAL, an earlier Stanford program that worked out the structure of chemical compounds from laboratory data using rules supplied by chemists. Bruce Buchanan, a computer scientist on the DENDRAL team, wanted to try the same approach on a medical problem.

The catalyst was Edward Shortliffe, a medical student who had discovered AI in a computer science course. He was working as a research assistant for Stanley Cohen, chief of clinical pharmacology at Stanford's medical school, on a drug-interaction warning program. Their talks turned to a program that would check antibiotic prescriptions, and then to something more ambitious: a computer consultant modeled on an infectious-disease specialist. Thomas Merigan, head of Stanford's infectious disease division, lent the expertise of himself and of physician Stanton Axline. The work became Shortliffe's doctoral dissertation.

Dates vary a little by source. The team's first grant application, in October 1973, said the collaboration had been running for about a year and a half, which places the start around early 1972. Overall the core development took five or six years in the early and mid-1970s. The program was written in Interlisp, a version of the Lisp language, and it ran on a large shared computer that users reached through terminals, including over the ARPANET.

How MYCIN reasoned

MYCIN's knowledge lived in about 600 rules (the count grew over time). Each rule had the form "IF these facts hold, THEN conclude this, with a given strength." The team described the specialist's task as four questions: Is the infection significant enough to treat? What organism is probably causing it? Which drugs could work against it? And which drug or combination is best for this patient?

To answer, MYCIN used backward chaining. Instead of starting from the data and seeing what follows, it started from a goal, such as "identify the organism," found rules that could conclude it, and then asked about whatever those rules needed. That is why a session felt like an interview: the program asked the physician one question after another, each one chosen because a rule needed the answer.

Medicine is rarely certain, so each conclusion carried a certainty factor, a number from −1 (definitely false) to +1 (definitely true). When two rules supported the same conclusion, MYCIN combined their certainty factors so the total grew without passing 1. For two positive values X and Y, the formula is X + Y − XY.

Here is a worked example of how evidence piles up:

Evidence	Certainty factor	Running total
Rule 1 suggests organism E. coli	0.8	0.8
Rule 2 also suggests E. coli	0.5	0.8 + 0.5 − (0.8 × 0.5) = 0.9
Rule 3 also suggests E. coli	0.3	0.9 + 0.3 − (0.9 × 0.3) = 0.93

Each new piece of support raises confidence, but by less and less. Negative factors, from rules that argue against a conclusion, pull the total down.

MYCIN could also explain itself. Asked why it wanted a piece of information, or how it reached a conclusion, it listed the rules involved. That feature mattered: the doctors on the team could read a chain of rules, spot a mistake and correct it.

Try it: Pick an everyday decision, such as whether to take an umbrella. Write three if-then rules with certainty factors (for example, "IF the sky is gray THEN rain, 0.4"), then combine them with X + Y − XY and see how your confidence changes.

How well did it work?

The most cited test, published in the Journal of the American Medical Association in 1979, used ten difficult cases of meningitis, an infection of the membranes around the brain. Ten "prescribers" recommended treatment for each case: MYCIN, five Stanford faculty specialists, a fellow, a resident, a medical student and the doctors who had actually treated the patient. Eight infectious-disease experts from other institutions then rated all 100 recommendations without knowing which one came from a computer.

The experts judged 65 percent of MYCIN's recommendations acceptable. The five faculty members scored between 42.5 and 62.5 percent. MYCIN also never missed a treatable organism. The authors themselves pointed out that ten cases is a small sample, and that the experts often disagreed with each other, which says as much about how hard these decisions are as about the program.

Why it mattered

MYCIN showed that the knowledge of a narrow specialty could be written down as rules and applied with expert-level results. A shell based on it, called EMYCIN, kept the reasoning engine but left out the medical rules, so others could fill it with knowledge from different fields. That idea, separating the reasoning engine from the knowledge base, helped fuel the expert-systems boom of the 1980s, alongside commercial systems such as DEC's XCON. MYCIN also exposed the field's chief bottleneck: getting knowledge out of experts' heads and into rules took enormous time.

In 2019, much of the excitement about AI in medicine centers on systems that learn from large collections of examples rather than from hand-written rules. MYCIN is a useful reminder that a medical program can perform well on paper and still face a long road to the bedside.

Why was MYCIN never used to treat real patients?

It was not because it gave bad advice. The obstacles were practical. MYCIN was a stand-alone program: a doctor had to sit at a terminal and type answers to a long series of questions, re-entering information that already existed in lab and pharmacy records. In the 1970s there was no easy way to connect it to those systems, and there were no personal computers on doctors' desks.

There were also open questions that the 1979 study named directly: Would doctors accept advice from a computer? Would it improve patient care? What would it cost, and who would be legally responsible if its advice was wrong? Those questions are still with us whenever an AI tool is proposed for a hospital, which is one reason MYCIN remains worth studying.