PUFF: an expert system in a hospital's daily work
In September 1982 Stanford issued a report on PUFF, a program that interprets lung function tests at Pacific Medical Center in San Francisco. It had been working there since 1979, gave interpretations for about ten patients a day and had handled over 4,000 cases; about 85% of its reports were accepted without change. On 144 new cases it agreed with the physician who had taught it in 96% and with an independent physician in 89%.
Why it matters
An expert system built on a shell taken from MYCIN became a daily tool of a hospital: a physiologist checks the program's interpretation and edits it if needed, and the report goes for a physician's signature into the patient's record. In the independent test it agreed with the outside physician on 89% of diagnoses, where the two physicians agreed with each other on 92%.
The first version was built with EMYCIN, a generalization of MYCIN, on the SUMEX machine at Stanford: 64 rules and 59 clinical parameters. For the hospital the rules were translated into procedures in BASIC on a PDP-11 minicomputer; that version has 400 rules. The knowledge base was tuned on 107 cases until the expert agreed with its interpretations, then frozen and tested on 144 new ones, each also read separately by two pulmonary physiologists: the one who worked with the builders and a physician from another medical centre. Agreement was counted as close: degrees of severity differing by at most one. The authors are Janice Aikins, John Kunz and Edward Shortliffe of Stanford and Robert Fallat of Pacific Medical Center. What the record does not claim: a month in 1979, since the document gives only the year; savings in time or money, which are not given. The 1982 report's figures are recognised text; where they differ from the reprint of the journal version (the production version's number of parameters, two table cells), the record follows the reprint.