AUREL va are dt The melding of artificial intelligence and medicine “One of the lessons learned in knowledge engineering is just how complicated’ is human knowledge” Artificial intelligence in medicine — computer prog- trams that alterupl lo acl as medical experts or consul. tants — is passing Hough the laboratory, door and is, ready to begin appearing in the clinics. Doctors’ offices are likely to be uext. Since Stanford is one of the world's leading centers for Al {artificial intelligence) research, a fot of those faboratory doors and clinics are here, And some of the first patients whose treatment is influenced by these prograins also are here. At the Medical Center's outpatient oncology clinic patients with Hodgkin's disease, non-Hodgkin's fym- phomas, and breast cancer, have been treated with the help of a progiam called ONCOCIN. ‘The program keeps wack of the patient's progress, advises on tabora- tory tests, and recommends drug Ulerapies. Since almost no two cases of these diseases are exactly alike, and since many of the therapentic drugs are toxic, physicians in the clinic, even chose reluctant to use computers, generally accept the help. ‘The researchers are now adapting the program so it will run on the desk-top computers that doctors’ offices can afford. A progiam called PUFF, derived from another Stan- ford AI programs, has been actually interpreting phy- siological tests of tung patients at the Pacific Medical Center in San Francisco. A program to help physicians appraise professional journal reports is under de- velopment. Stanford, of course, is not the otliy place where re-* searchers are lying to use Al tyols to help physicians and medical researchers. The national interest in biomedical applications of Al became so intense in the carly 1970s that the National Enstitutes of Health several years ago agreed to fund a national computer resource, a large machine that many researchers front around the country could access. They put iton the ground flooe of the Stanford Medical Center, and itis known as SUMEX-AIM (Stanford University Medical Experimental Computer-Artilicial Intelligence in Medicine). SUMEX-AIM is a collection of scientific worksta- tions tied to a Digital Equipment Carp. 2060 and VAX computers. Uhiee to four hundred researchers avound the country are plugged into (he SUMEX com- puters. The leading edge of this research at Stanford is at the Knowledge Systems Laboratory (KSL), an tnterdis- ciplinary set of projects tying the Computer Science Department to the School of Meclicine. More than [00 researchers, including five principal investigators, and scores of graduate students work in the five sublabs of the KSL. du addition, researchers in other ficlds, such as Oleg Jardetzky, of the Magnetic Resonance Laboratory, and Ghasles Yanofsky in bialo- gy, are working on specialized progiams. From the beginning, a goal of Al researchers has been to encapsulate the knowledge and problem. solving skills Of experts into a computer application, so-called expert systems. ICwould be useful, the resear- chers believed, to try to put the reasoning and know- ledge of humau subject experts into the computer so non-experts could draw upon dheir skills. The first expert system was produced at Stanford in the fate 150s. Called DENDRAL, the program was developed by Nobel laureate Joshua Lederberg, bruce Buchanan and Edward Feigenbaum, professors of computer science. LLattempted to capture Lederberg’s expertise ty analyzing of ganic compounds from mass spectroscopy One ol the most successful Al programs ever, DE- NDRAL can be found in many organic chemistry labs, issued under a Stanford license. DENDRAL. is based on a core concept in artificial intelligence, the use ol Aeuristics, somelines known as the art of good guessing. Eleuvistics are similar to the mental processes of a ress player. The chess expert ignores the almost-infinite oumber of possible moves ina game, concentrating only on those relevant to the parbcular situation. DENDRAL cannotconsider every possible molecule in doing its analysis. Using heuristics, however, the progtam considers only those likely to be the answer. The next logical step was a program for clinical medicine, designed as part of a dissertation by Edward Shortlilfe, now associate professor of medicine and computer science, Shotliffe isan M.D, witha passion for computers. Along with Stanley Cohen, Bruce Buchanan, and Stanton Axtine, Shortliffe developed MYCIN. MYCIN, which uses heuristics, gives advice on the yi a Edward Feigenbaum use of antibiotics for the trealinent of intectious dis-|. eases. ‘The method used to put the expertise of Cohen and| Axline into the comprter shaws how Uese programs are developed — and bow difficult it can be to build them. The process is called “kriowledge engineering, mt and it tins outta be one of tie most complicated parts t of building AL programs. “We sat around, went over patient cases, and tied to understand how Axline and Cohen would decide how to treat those cases,” explained Sharuiife, who was a second-year medical student at the time, “We'd stop them — those of us who knew only alittle medicine and were more computer scientisis — interrupt and ask, “Well, why de you say hac” “We met once a week lor an hour aud a half. Axline would bring in an interesting case, a chart, and Cohen would ask him questions. Axline would answer them fiom the chart. Cohen would be problem-solving, trying to figure out what he'd do For that case, and we would ty to understand why Cohen was asking the! questions he was. “We'd write down the rules that he told us, that Axtine would help refine with him, and in the interim week, I'd put those rules into this developing and emerging computer system,” Shordiffe said. “And then we'd all have a good laugh the following , week when I showed them how the computer had tried to handle the same case. What you did (was) discover . the great simplifications they made in explaining the . 1D OBSERVER rules fromthe previous week, where they'd go wrong if you Wied to run it on any kind of different case.” And, “every once ina while,” Shortliffe said, “some issues would arise which caused a major change in the underlying structure of this developing program. Sud- denly it wasn't just a matter of wilting the rules; you bad to make major program changes.” : The program, like most AJ programs, is writlen ina dificult, non-numesic computer language called LISP; the data and the commands are so integrated that by altering one, you may automatically alter the other. One of the lessons learned in knowledge engineer- ing is just how complicated is human knowledge. “the compnier, for instance, was quite capable of asking whether a patient was pregnant, ever ifthe patient was male. We know only females get pregnant, but the programmers had to remember to add that ttle bit of commonsense knowledge to an intinsically ignorant machine. Nonetheless, MYCIN worked. More important, the program could be exported to other fields. The logical part of the program, what came to be called the “infer- ence engine,” seemed tg be relatively universal. By extracting it and simply plugging in new data from a dillerent domain, We MYCIN program could work for other fields just as it die for infectious diseases. ‘The inference engine was called Essential MYCIN or EMY- GIN. PUFF at the Pacilic Medical Center is based on EMYCIN, MYCIN also led to ONCOGIN, the cancer treat- nent program under development by Shortliffe and Lawrence Fagan. ONCOGIN nionitors the patient's condition by asking the physician questions aboul the patient, the treatinent given so far, and the results of tests. ‘The nages the patient's dig therapy and is se of the complexity of cancer chemother- apy. The computer also has the ability to retain in its memory far more (eatment details than an oncologist can remember. It is also constantly being updated with the newest research: results. Hf, for instance, a Lest shows that a patient's white blood cell countis decreasing, ONCOCIN may suggest ways of changing drug tcalinent. As with all such programs, the physician can acceptor reject the advice — the uldimate responsibility remains with the doctor. (Thatis nota minor issue. Besides the natural reluct- ance of a physictan to re qtish control, particularly to a inachine, the question arises: what happens iT the machine is wrong? Who gets sued for malpractice? ‘Vhese are challenging legal issues ot yet tested in thie courts.) ONCOCIN has a secondary program that explains the basis for its decision. Lf the physician asks why a recommendation is made, ONCOCIN reviews ils reasoning and documents the information upon which the advice was given. The physician then has a better iclea how much weight to give the suggestion — and, in soine cases, may learn: something be or she did not know, or remember something Forgotten. ONGOGCIN is built on a series of [F-THEN state- ments; if a particular situation applies, Hen the compu- ter should conclude something specific or recommend a particular course of action. However, “although ONGOCIN may appear to be very literal. . itis impor- tant Lo remember that we don't tell the computer what to recommend for every conceivable situation,” said Larry Fagan, project director and senior research associate. “Instead, we supply it with knowledge and instruc- tion on how to put bits of knowledge together. The computer then integrates incoming information on its ‘own. In-very large prograins such as ONCOCIN, the result is often unexpected. How well do these machines do? Ina test matching ONGOCIN with human experts, the progrant shows excellent performance of the computer relative to physicians Uealing cancer patients at Stanford,” Fagatr said. (Lhe study was described in the December 1985 issue of The Annals of Internal Medirine.) Another large expert system, CADUCEUS, writen at the University of Pittsburgh, was lested using cases from The New England Journal of Medicine, Its perform- ance was then compared to that of a group of physi- cians. Over a wide range of diagnoses, the mar hine was found lo be more accurate than an average physician, roughly comparable to the teams of physicians who cared for the patients, and almost as good as expert hysicians asked lo review the cases in retrospect. Although CADUCEUS can handle 600 different di- aghoses, Ute scientists who programmed it, like those in all aspects of AL in medicine, do not believe the program can ever replace the physician. Among other things, says Pittsburgh's Jack Myers, CADUCEUS lacks imagination. It cannot cope with a disease it is unprogrammed to diagnose, whereas the bright human physician will recognize something new. All the programs, including ONCOGIN, lack immagina- tion. ONCOCIN has been used experimentally in the Stanford oncology clinic since 1981, but the large com- puler prototype was recently removed in anticipalion of the new version that will run on small machines. The new interface with Uke physician will resemble the Apple Macintosh with graphics and “windows” (sections of the screen that show specific Functions). The physician will move a mouse to control the action on the screen, {The screen will look just like the kind of paper chart the physician is familiar with — a deliberate design feature. “Itcan do everything the physician's paper record can do, butis also able to do things that only computers bebtitniy boo Edward Shortliffe can do — for example, using continuous forms on the screen thal can be lengihened as necessary,” Fagan says. A separate research program is being developed to use more general strategies, Fagan said. This wifl en- able the physician to cope with more complex cases. “For the simple case it comes out with a simple auswer; for the complex case it comes oul wilh more geueral or fuzzier answers. “This corresponds to two cases. One is the regular case which the physicians generally handle them- selves," Fagan said. “The other corresponds to the place where they would call in the expert for special- wed advice.” Naving ONCOCIN ona large research computer is of litte practical value toa local physician; the goal is to put ONCOCIN on a workstation, a inicrocompuler- sized machine costing between $40,000 and $20,000, which puts the machinery within react of many doc- tors’ offices. ‘Phe Medical Computer Science Group has prop- osed distributing ONCOUIN experimentally in col- laboration with the Northern California Oncology Group so the programs can be tested on workstations away froru Stanford. ONCOCEN is perhaps the largest of the medical Al projects going on at Stanford. However, there are several other related projects. One program, called PATILEINDER, is an attempt to advise pathologists on the proper classification of lymph node abnormalities viewed under the snicro- scope. This program is being built in collaboration with the University of Southern California, REFEREE, a program being written by Bruce Buchanan of KSL and William Brown and Daniel Feldman of the Medical Center, is an attempt to try to build an adviser for interpreding papers from the medical literature. REFEREE, will give the physician a feel for the reliability of the data in Ue reported study. PROTEAN, developed by Buchanan and Barbara Hayes-Roth of KSL, along with about 10 students and nuclear magnetic resonance expert Oleg Jardetzky. will help researchers determine the structures of pro. tein molecules. In the meantime, SUMEX has been changing. Anum ber ol research projects from around the country have chosen to move olf the large DEC machines at the Medical Genter and use workstations or large micro computers. SUMEX is now part of the Symbolics Systems Re search Group, the sublab in the KSL thatis responsibk for the computing “environment” for the laboraters litercotingly, aithough the medical programs, i all the other AJ programs at Stanford, frequently hav practical applications, the researchers are driven b the means, not the end. ‘Prying to unravel Oleg detzky's world in nuclear magnetic resonance mit prove helpful to Jardetzky, but it is more helplul to th Al tesearchers, who learn by doing. “L's fine line, because itlooks like we're building a application for Jardetzky. That we hope will be ar sult, but it's nol anything whose success we cout fuarantee, certainly not when we started. It's true ¢ every project we do,” Buchanan says. On the other hand, of course, it would be nice it! programs work, he adds. That helps get enthusias: collaborators. Allthis could tead to a scenario out of science fictior A doctor inthe Sierras has a patient whose disease ! cannot handle. Ele needs help. The help comes fro his olfice computer. Housed in the metal box, ont! silicon chips, behind the video screen is an inanime expert. The computer asks questions, The physict: answers them at the keyboard. ‘The computer dels into its vast data bank, makes the judgment worthy an expert, aud prints out its best aclvice. , Ed Feigenbaum, the co-author of DENDRAL, t lieves chat will begin the real computer revolutig : —Joel Shur