MYCIN

What it was

MYCIN was an early expert system built to identify the bacteria responsible for severe infections such as bacteremia and meningitis, and to recommend antibiotics with dosages adjusted to the patient's body weight^[1]. It was a stand-alone system running on a DEC KI10 PDP-10, at a time before personal computers, and a user had to type in all of the patient information^[1].

Where it came from

The program was developed over five or six years in the early 1970s at Stanford University, and there is no single release date for it^[1]. It was written in Lisp as Edward Shortliffe's doctoral dissertation, under the direction of Bruce G. Buchanan, Stanley Cohen and others, and it came out of the Stanford Heuristic Programming Project^[1]. A Stanford account describes it as developed at the Stanford Medical School in the mid-1970s by Edward H. Shortliffe, a physician and computer scientist, as a collaborative endeavor between the medical school and the Heuristic Programming Project^[3]. A report from that project dated October 26, 1976 lists Buchanan as co-principal investigator of the ARPA contract and the MYCIN grant, and Cohen as co-principal investigator of MYCIN^[2].

How it worked

MYCIN reasoned by backward chaining: it began from the question it was trying to settle and worked back to the pieces of information it needed. Its architecture was split deliberately in two, a fairly simple inference engine and a knowledge base of about 600 rules, with the domain knowledge kept clearly separate from the inference procedures^[1]. A Stanford article gives the knowledge base as 500 rules in IF-THEN form, and says the program examined evidence for and against approximately 100 diagnoses^[3]. To collect that evidence it queried the physician through a long series of simple yes/no or textual questions^[1].

The output was a ranked list of possible culprit bacteria together with the system's confidence in each, the reasoning behind each diagnosis, and a recommended course of drug treatment^[1]. Uncertainty was handled with a certainty factor model that combined evidence weights, with the properties that −1 means false, +1 means true and 0 means uncertain^[1]. The 1976 report adds that diagnostic and therapy strategies were represented in a system of meta-rules^[2].

Explaining itself

A distinctive part of the design was that the program could answer a physician's queries about why it asked a particular question, how it had reached a conclusion, and why it had not considered certain factors^[1]. Typing RULE produced a human-language explanation of the last rule executed, while WHY triggered a back-trace of the inference process^[3]. The 1976 project report describes MYCIN as holding consultations in English with physicians about a patient's infection, leading to a diagnosis and a therapy recommendation, and as answering questions in English about its own line of reasoning^[2].

Evaluation

In an evaluation at the Stanford Medical School, MYCIN received an acceptability rating of 65%, comparable to the 42.5% to 62.5% ratings of five faculty members^[1].

Why it stayed a research system

MYCIN was never used in practice^[1]. That was not because of any weakness in its performance but because of the state of system integration technologies at the time^[1]. It remained a stand-alone program on a large machine, fed by hand-typed patient data, rather than a component of the hospital's own information systems^[1].

Its place in the field

What MYCIN demonstrated was a way of organising expert knowledge: several hundred conditional rules kept apart from a general-purpose reasoning program, so that the knowledge could in principle be inspected and explained without rewriting the inference machinery^[1][3]. Equally important was the emphasis on justification. A system that only printed a recommendation would have been hard for a clinician to trust; one that could say which rules it had used, and what it had ruled out, made its judgement open to examination^[1][3]. Those two features — a separated rule base and an explanation facility — are the parts of MYCIN that later work on expert systems kept returning to, even though the program itself never reached patients.