MYCIN

MYCIN was an early expert system built at Stanford University in the first half of the 1970s to help physicians diagnose bacterial infections and choose antibiotic therapy. It grew out of the doctoral work of Edward H. Shortliffe, who developed the program with the computer scientist Bruce G. Buchanan and the physician and geneticist Stanley N. Cohen as advisors ^[1][2]. The name comes from the "-mycin" suffix shared by many antibiotics.

The program worked by holding a structured consultation: it asked the clinician a series of questions about a patient (symptoms, test results, the site of infection), reasoned over a knowledge base of roughly 600 rules, and then recommended one or more antibiotics with doses adjusted for the patient's body weight ^[1]. MYCIN never entered routine clinical use, but it became one of the most studied programs in the history of artificial intelligence. Its rule-based design, its handling of uncertain evidence through "certainty factors," and its ability to explain its own reasoning shaped a generation of symbolic AI work, and its rules were later stripped out to create EMYCIN, a reusable shell for building other expert systems ^[1].

Background

MYCIN belongs to the lineage of Stanford's Heuristic Programming Project, the research group that had already built DENDRAL, a program that inferred the molecular structure of organic compounds from mass spectrometry data. DENDRAL, begun in 1964 by Edward Feigenbaum, Joshua Lederberg, Carl Djerassi, and Buchanan, is usually described as the first expert system, and several later systems were derived from it, MYCIN among them ^[3]. The central lesson the group drew from DENDRAL was that narrow, expert-level performance came less from clever general reasoning than from encoding large amounts of specialist knowledge and letting a relatively simple inference procedure run over it. That idea, sometimes called the knowledge-is-power hypothesis, is the foundation MYCIN was built on.

The medical problem MYCIN addressed was real and pressing. Physicians treating severe infections such as bacteremia (bacteria in the bloodstream) and meningitis often had to start antibiotics before laboratory cultures came back, sometimes a day or two later. Choosing the wrong drug, or the wrong dose, carried serious consequences, and many clinicians were not infectious-disease specialists. A program that could reason from the available clinical picture to a defensible therapy recommendation had obvious value, and infection was a domain where expertise was both scarce and highly structured, which made it a good target for the knowledge-based approach.

How MYCIN worked

MYCIN was written in Lisp and ran on a DEC PDP-10 mainframe, using Stanford's shared research computing facility ^[1]. Its knowledge was represented as production rules: conditional statements of the form "if these conditions hold, then conclude this, with a given degree of certainty." A representative rule might state that if the site of a culture is blood, the organism's morphology is rod-shaped, and the patient is a compromised host, then there is suggestive evidence that the organism is a particular species.

Backward chaining

The program's control strategy was backward chaining, run by a separate inference engine that was kept distinct from the rules themselves ^[4]. Rather than starting from the facts and grinding forward to every possible conclusion, MYCIN started from a goal (for example, "determine the identity of the infecting organism") and worked backward, looking for rules whose conclusions bore on that goal and then trying to establish their premises. When establishing a premise required information the system did not have, it asked the user a question. This goal-directed questioning is why a MYCIN consultation felt like an interview: the program only asked about things that were relevant to a hypothesis it was currently pursuing. Separating the inference engine from the domain rules was itself an influential design choice, because it meant the same reasoning machinery could in principle drive a different rule base.

Certainty factors

Medical reasoning rarely deals in certainties, and MYCIN's designers needed a way to combine evidence that was merely suggestive. They introduced certainty factors, numbers attached to rules and conclusions that expressed a degree of confirmation or disconfirmation on a scale that ran from strong disbelief to strong belief. The model, described by Shortliffe and Buchanan as a model of inexact reasoning, provided rules for combining the certainty factors of several pieces of evidence into an overall measure of confidence ^[5]. It was deliberately not a strict application of probability theory. The developers argued that clinicians did not have, and could not easily supply, the large tables of conditional probabilities that a fully Bayesian approach would demand, so they built a lighter-weight calculus that a physician could reason about directly. The certainty-factor model was later criticized on theoretical grounds and is one of the reasons probabilistic methods returned to favor, but at the time it let MYCIN function usefully under uncertainty.

Explanation

MYCIN could be asked to justify itself. During or after a consultation, a user could type "why" to learn why the program was asking a particular question, or "how" to see the chain of rules behind a conclusion. Because the reasoning was an explicit trace of fired rules, the system could reconstruct and display that trace in something close to English ^[4]. This transparency was a major selling point for a medical audience that was, reasonably, unwilling to accept therapy advice from a black box, and it set an early standard for explainable reasoning that remains a live concern in AI.

Evaluation

The question of whether MYCIN actually gave good advice was tested more rigorously than was common for AI programs of the era. In a formal evaluation, MYCIN's recommendations on a set of cases were compared against the judgments of human specialists, with the evaluators not told which recommendations came from the machine and which from people. In one such study MYCIN's therapy recommendations received an acceptability rating of about 65 percent, comparable to the 42.5 to 62.5 percent ratings earned by five human faculty members on the same cases; the evaluation used ten test cases judged by infectious-disease experts ^[1]. In other words, the program performed at roughly the level of the human experts and, on this measure, was not clearly worse than any of them.

That result was striking for the time and is often cited as evidence that a knowledge-based program could reach expert-level competence in a narrow domain. It should be read with care, though. The evaluation covered a small number of cases in a tightly bounded slice of medicine, and acceptability ratings from reviewers are not the same as measured patient outcomes. The finding supported the claim that the approach worked, not that MYCIN was ready to treat patients.

Why it was never used clinically

Despite its performance, MYCIN was never deployed in actual patient care ^[1]. The reasons were only partly technical.

Scroll sideways for more →

These obstacles were characteristic of the first wave of medical AI: the reasoning could be made to work in the laboratory, but fitting it into the workflow, liability structures, and infrastructure of real hospitals was a separate and much harder problem. Many of the same issues, especially responsibility for automated advice and integration with clinical systems, still shape debates about AI in healthcare today.

EMYCIN and legacy

One of MYCIN's most durable contributions was indirect. Because the inference engine had been kept separate from the medical rules, researchers realized that the reasoning machinery could be reused for entirely different problems by supplying a new rule base. Emptying MYCIN of its infectious-disease knowledge produced EMYCIN, short for "essential MYCIN" (sometimes read as "empty MYCIN"), a domain-independent shell for building rule-based expert systems ^[1]. EMYCIN is an early example of the expert-system shell, a category of software that became a commercial product line during the expert-system boom of the 1980s. PUFF, a system for interpreting pulmonary function tests, was among the applications built on the EMYCIN framework.

The MYCIN work was documented at length in two books that became standard references. Shortliffe's dissertation was published in 1976 as Computer-Based Medical Consultations: MYCIN, and a later retrospective edited by Buchanan and Shortliffe, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, appeared in 1984 and collected the project's papers on rule representation, certainty factors, explanation, and evaluation ^[2][5]. Shortliffe himself earned a PhD in medical information sciences from Stanford in 1975 and an MD there in 1976, and received the ACM Grace Murray Hopper Award in 1976 for the work ^[2].

MYCIN's influence on later AI is easy to trace. It helped establish the standard architecture of an expert system, a knowledge base of rules plus a general inference engine plus an explanation facility, and it demonstrated that this architecture could match human specialists in a real domain. That success fueled the commercial enthusiasm for expert systems in the 1980s, and the subsequent gap between what those systems promised and what they delivered contributed to the funding pullback later called the AI winter. The knowledge-acquisition problem MYCIN exposed, the sheer difficulty and cost of extracting expert knowledge and hand-coding it into rules, became the central bottleneck of the whole approach and is one reason later AI shifted toward systems that learn from data rather than rules written by hand ^[4]. Even so, the concerns MYCIN raised first, transparent reasoning, calibrated uncertainty, and accountability for automated medical advice, are once again central as machine-learning systems enter clinical settings.

See also

• Expert system
• DENDRAL
• Symbolic AI
• Inference engine
• AI in healthcare
• History of artificial intelligence

References

1. ^"Mycin." Wikipedia. en.wikipedia.org/...Mycin
2. ^"Edward Shortliffe." Wikipedia. en.wikipedia.org/...Edward_Shortliffe
3. ^"Dendral." Wikipedia. en.wikipedia.org/...Dendral
4. ^"Expert system." Wikipedia. en.wikipedia.org/...Expert_system
5. ^Buchanan, Bruce G., and Edward H. Shortliffe, eds. Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project. Addison-Wesley, 1984 (out of print; chapters freely available). people.dbmi.columbia.edu/...MYCIN%20Book