MYCIN, or the physician who never saw a patient
In 1979, someone at Stanford Medical School mailed ten meningitis cases to eight infectious-disease specialists scattered across the country. Each specialist answered the same questions, recommended antibiotic treatments, and sent their responses back. So did a ninth consultant — a program called MYCIN, running on a DEC PDP-10 in the Stanford computer science building. When an independent panel scored the results blind, MYCIN’s recommendations were rated acceptable 65 percent of the time. Five of the eight human physicians scored lower.
Edward Shortliffe had started writing MYCIN in 1972, at the intersection of Stanford’s computer science department and its divisions of clinical pharmacology and infectious disease. The project was his doctoral dissertation, guided by computer scientist Bruce Buchanan and a team of Stanford clinicians. The task was deliberately narrow: given a patient’s lab results, fever history, and clinical observations, identify the likely pathogen and recommend the appropriate antibiotic. The name came from the domain itself — the -mycin suffix that trails through half the antibiotic pharmacopeia.
The mechanism was a chain of roughly 600 if-then rules. A representative one ran: if the gram stain is negative and the morphology is rod-shaped and the infection site is blood, there is suggestive evidence of Pseudomonas aeruginosa. Working backward from a hypothesized conclusion, MYCIN’s inference engine assembled and queried rules until it could score each hypothesis with a certainty factor — a number between –1 and +1, where –1 meant “definitely not” and +1 meant “definitely yes.” The certainty factor was not borrowed from statistics; it was MYCIN’s own engineering, a way to push approximate reasoning through a system built for crisp logic.
The 1979 blind evaluation was a genuine shock. MYCIN matched or beat most of the human experts, and outside judges disagreed with it no more often than they disagreed with Stanford’s own faculty.
It never saw a real patient. The PDP-10 that ran MYCIN was not connected to any hospital information system; every lab value had to be typed in by hand. Legal liability was unresolved — if a physician followed MYCIN’s recommendation and the patient died, accountability was unclear. And while those questions circulated in committee rooms, a new generation of cephalosporin antibiotics arrived and MYCIN’s knowledge base began going stale. Shortliffe and his collaborators put it plainly in 1984: acceptability, it turned out, was different from adoption.
What MYCIN exported was more portable than the program itself. The EMYCIN shell — MYCIN stripped of its medical rules, left as a configurable inference engine — let other teams fill in their own domains. Prospector used it to assess mineral deposits. Digital Equipment Corporation’s XCON applied a similar rule-based engine to configure VAX computer orders, reportedly saving DEC $25 million a year by the mid-1980s. A cluster of AI companies — Symbolics, IntelliCorp, Teknowledge — reached a combined market capitalization over a billion dollars by 1985. Symbolics, which built computers optimized for running Lisp-based AI programs, became the first company to register a .com domain, in March of that year.
The brittleness caught up with all of it. Rules that handled every known case broke on the first unknown one. Knowledge engineering — the slow, expensive work of extracting expertise from human heads and encoding it as if-then statements — turned out to cost far more than any vendor had projected. By the early 1990s, cheaper UNIX workstations had made specialized Lisp machines obsolete, the venture capital had drained, and the field entered its second winter. MYCIN’s blind-test performance, filed away in a 1979 evaluation, would be rediscovered decades later by researchers who found it a useful benchmark — a reminder that clinical AI had once arrived, been judged acceptable, and then politely turned away at the door.
Sources
- Mycin — Wikipedia — development timeline, certainty factors, backward chaining, ~600 rules, and why the system was never deployed clinically.
- MYCIN: An Expert System for Infectious Disease Therapy — Forbes / Gil Press — 1979 blind evaluation details, performance versus faculty physicians, and adoption barriers.
- The Artificial Intelligence Boom of the 1980s — Brewminate — XCON at DEC, Symbolics and the Lisp machine market, expert-systems boom and collapse.