Calling "Dr. SUMEX"

By Richard Stengel

The diagnostician's new colleague needs no coffee breaks

The white-coated doctor is hunched over his patient's bed. He is running the curious combination of symptoms through his mind, but he cannot quite get the overall picture to add up. "I'd like to call in a colleague to recommend a diagnosis," he tells the patient. The consultant he has in mind has decades of medical experience, an infallible memory for past cases and an encyclopedic knowledge of symptoms and diseases. The consultant, whose only shortcoming is a rather cold bedside manner, is a computer.

Ten years ago, given the relatively rudimentary state of computer programming and the entrenched skepticism of the medical profession, this scene could only have been imagined. Although this technology is still years away from wide utilization, today it is already in use in a few hospitals. In the pulmonary lab at Pacific Medical Center in San Francisco, after a patient exhales into an instrument that gauges lung functions, a computer program takes over. Named PUFF, it analyzes 250 factors that determine pulmonary dysfunction, then within 90 seconds issues a printout that may, in its own words, "indicate" or "suggest" what is wrong with the patient. Dr. Robert Fallal, the hospital's chief of pulmonary medicine, claims that in 85% of the cases, PUFF makes recommendations a doctor can use. "It is still rigid, and it has a problem with mixed diagnoses," says Fallat, "but it gives you time to concentrate on the remaining, challenging 15%."

At the Stanford Oncology Day Care Clinic in Palo Alto, a computer program called ONCOCIN watches over 30 patients suffering from Hodgkin's disease and non-Hodgkin's lymphoma. It studies their test results, then proposes a complex treatment program called a protocol, which includes lab tests, X-ray studies and subtle changes in chemotherapy. Says ONCOCIN's chief programmer, Carli Scott: "We're not taking decisions away from the doctors, but helping with their calculations."

PUFF and ONCOCIN are brainchildren of SUMEX-AIM (Stanford University Medical Experimental Computer-Artificial Intelligence in Medicine). SUMEX is an interlocking of two sophisticated computer facilities at Stanford in California and Rutgers University in New Jersey, funded by the National Institutes of Health. Half the transcontinental system's capacity is devoted to development of an artificial intelligence for diagnostic medicine. At the moment, SUMEX's two bi-coastal computers also link 20 research projects in the U.S., Europe, Japan and Australia. Says Saul Amarel, chairman of the Rutgers department of computer science: "The purpose of SUMEX is to help doctors match signs and symptoms and get not only some probable diagnoses, but information on drugs and their side effects."

SUMEX, say proponents like Dr. Edward Shortliffe, who holds both an M.D. and a Ph.D. in medical information sciences from Stanford, is the product of a new kind of programming that imitates a physician's thought processes. In contrast to electronic libraries (computers that are little more than sophisticated adding machines or memory banks) the SUMEX program works with "uncertainty" factors. These yield recommendations to physicians that avoid simplistic black-and-white terms in favor of various shades of gray. The program is designed to take into account all the available evidence that might either buttress or call into doubt a given medical hypothesis. This enables a physician to go back to the computer and question it as to how it weighed the relevant information and arrived at its recommendations.

To attain this subtlety in computer reasoning and logic, SUMEX programmers spent hundreds of hours with physicians in an effort to understand the thought process that led them to certain diagnoses. For PUFF, programmers observed Dr. Fallat for three months as he analyzed some 100 pulmonary cases. The resulting program, says its model, in effect represents "the cloning of Bob Fallat." M.I.T. Computer Expert Peter Szolovits and his colleagues examined 15 doctors' thought processes over a year to program his Digital Equipment PDP-10 computer. "We don't believe we understand their analytic behavior until we write a computer program that reproduces it," he says.

Diagnosis, of course, is far from an exact science. Respected specialists will examine the same set of symptoms and arrive at different conclusions. Most medical decisions are educated guesswork; nevertheless, the computers are already functioning well. A report in the Journal of the American Medical Association states that one SUMEX program performed at a level comparable to that of five medical experts at Stanford. William Baker, NIH administrator for the SUMEX project, says that a computer system at the University of Pittsburgh called CADUCEUS is so sophisticated that it "would be a board-certified internist if it were human." Pittsburgh researchers administered one part of an internal-medicine board exam to "Dr. CADUCEUS." It passed.

Some physicians remain cautious about the ultimate potential of computer colleagues. Says Stanford Oncology Fellow Bob Carlson: "My reaction is mixed. The system is very good for data gathering, but for its recommendations of drugs and treatments, it's in its infancy." But knowledgeable physicians predict that consultation by diagnostic computers will soon be widespread. Sums up Szolovits: "The computer doesn't get tired, and it doesn't forget things." Or, one assumes, play golf every Wednesday afternoon.

--By Richard Stengel. reported by Dick Thompson/San Francisco

With reporting by Dick Thompson/San Francisco

This file is automatically generated by a robot program, so viewer discretion is required.