You can't quite say that a computer will consistently outdiagnose a doctor. But it will give him some helpful hints.
DOCTORS make mistakes. In 12 cases at LDS Hospital in Salt Lake City in 1987, they forgot to change antibiotics when treating infections after lab results indicated the doctors"best-- guess" therapy was inappropriate.
What saved the hospital from a malpractice claim, and sped up the healing of the 12 patients, was a computer program called Help. Originally developed at LDS by researchers affiliated with the University of Utah and marketed by 3M, Help takes patient information such as lab results and drugs taken, analyzes it and then issues alerts if anything seems wrong. Help, a massive program representing 20 years of development and about 6 million lines of computer code, isn't so arrogant as to think it can outsmart the average doctor. It just delivers helpful hints, much like a PC spelling checker.
Some of this M.D. software is darn good. The Apache program, which evaluates an intensive care patient's chance of survival, has matched wits with doctors and computed the odds more accurately. Another system seems better than physicians in detecting acute heart attacks.
What the computer lacks in inductive quality it can sometimes make up in quantity. "The amount of information physicians have is bewildering," says Mark Kelley, vice dean for clinical affairs at the University of Pennsylvania Medical Center. "The human mind can only handle three to four variables at one time." A typical hospital carries an arsenal of 2,000 drugs. How could a human being know whether any given combination of drugs is dangerous? TDS Healthcare Systems in Atlanta sells a program that, among other things, warns of adverse drug reactions.
As development costs get spread over more users and as hardware prices crash, medical diagnosis becomes much cheaper. Help, for example, sells for as little as $1 million, including Tandem mainframe and terminals. Individual physicians can license diagnostic re software for less than $1,000 a year.
What's holding many hospitals and doctors back from the computer age is collecting the information, typing it in and having the computer make sense of it. Few hospitals even approach total computerization of everything from patient records to prescriptions to lab work to treatment orders. Physicians balk at entering patient notes into a computer; until new technologies like voice recognition and pen computing fulfill their promise, data entry will continue to take precious time.
What the problem boils down to is that computer and software marketers have a lot of missionary work to do. They must persuade doctors and hospitals that their services are worth the cost and effort. In some cases, the payoff goes beyond the possible benefit to one patient; the system is part of a hospital-wide program that also does accounting or review of equipment utilization. The Apache system, sold by Washington, D.C.-based Apache Medical Systems, can be used to evaluate a patient as well as manage the intensive care unit, which accounts for about a third or more of hospital costs. Price: $150,000 and up (software and hardware).
This is one kind of medical expenditure that just might lower patient bills rather than raise them. At the Regenstrief Institute in Indianapolis, Clement McDonald developed Gofer, a program that is supposed to cut back on unnecessary lab tests simply by letting doctors know how likely it is that a given test is going to come back with an uneventful result.
[Photograph]
The Help system in the intensive care unit at LDS Hospital in Salt Lake City
The computer doesn't tell the doctor what to do. It just flashes if it thinks he might have made a mistake.
For garden-variety problems that physicians are used to seeing, diagnostic software doesn't add much. The payoff comes when a physician hasn't yet heard of a disease or doesn't recognize it because it's either very rare or the symptoms are somewhat uncharacteristic (see box).
Doctors don't have to turn treatment decisions over to a computer program to get value from it. If nothing else, the doctor can use the program the way he would use a textbook-look under eosinophilia (an abnormally high count of a certain kind of white blood cell), combined with swelling around the eyes, and you get trichinosis as a possible cause. The computer beats the index at the back of a textbook because it digests several symptoms at once. Massachusetts General Hospital's DXplain covers 2,500 diseases. "It's just a different way of accessing information," says Octo Barnett, who headed the program's development team.
DXplain, like many diagnostic systems, builds on a simple law of probability, Bayes' Theorem, to make intelligent guesses as to possible diseases and to rank them in descending order of likelihood. Bayes' Theorem tells you the chance that the patient has disease X if a symptom is Y, when you know how often X and Y coincide and how often Y occurs in the absence of disease X.
Simple as this little 18th-century theorem is, it can produce some startling conclusions. Here's a classic example of its application, taught to virtually every medical student these days. Imagine that a dread disease occurs in 0.5% of the population and that a diagnostic test is 98% accurate-that is, if a patient has the disease, there's a 98% chance that the test will come back positive, and if he doesn't have it, there's a 98% chance the test will come back negative. A person is selected at random and tested; the test comes back positive. What is the probability that he in fact has the disease?
Only about 20%. Why? Because, if everyone in the U.S. were tested, not quite 2% of them would have false positive tests and a bit less than 0.5% would have accurate positive tests. If all you know about the patient is that the test is positive, you have to conclude that the odds are 4-to-1 against the disease.
Experienced doctors build this logic into their daily practice. They say to themselves: Don't jump to conclusions. Machines pick up this wisdom by having statistics fed to them.
However sound the theorems behind them, diagnostic programs can be, just like doctors, simply wrong. Charles Safran, a physician and Harvard Medical School professor, kept one patient in the hospital for an extra week for tests-all of which turned out negative-because a program was fixated on the hypothesis-incorrect, as it turned out-that the patient had tuberculosis.
[Sidebar]
A medical detective story
THE STAKES were high. A 50-year-old school janitor was trying to collect thousands of dollars in workers' compensation. The janitor, a Korean male, suffered from tingling and numbness in his feet and legs. He had gone to see a doctor who specialized in "clinical ecology," and that doctor said the janitor was ill because he used toxic insecticides at work.
Stuart Rusnak, a Honolulu allergist and immunologist, was called in to study the case. After reviewing 30 years of the janitor's medical records, Rusnak entered all of the symptoms he found in the QMR (Quick Medical Ref erence) program sold by Camdat. The program suggested he consider two other possibilities: diabetes and alcoholism. Lab tests ruled out the first. But answers on yearly medical questionnaires and other notes by the janitor's doctor suggested the janitor drank heavily.
QMR, says Rusnak, pointed him in a direction he hadn't yet considered. The workers' comp claim was turned down.
QMR can also suggest possibilities when all the obvious diagnoses are ruled out. Barry Blumenfeld, one of the founders of Camdat and an internist, recalls using QMR to help him figure out what was wrong with a patient with acquired immune deficiency syndrome and severe diarrhea. Blumenfeld tested the patient, a 34-yearold homosexual male, for all the probable culprits: parasites or bacteria. The lab tests found nothing.
What then? Blumenfeld turned to QMR and entered the patient's two main findings: ADS and acute diarrhea. The program came up with a half dozen or so possibilities already ruled out by the tests. But the program also listed a diagnosis Blumenfeld hadn't originally considered: non-Hodgkin's lymphoma. That's what ailed the patient.
While QMR and similar software can help physicians remember a diagnosis they've momentarily forgotten, don't expect computers to replace doctors anytime soon, says Randolph Miller, a University of Pittsburgh physician who helped create QMR. There's still a lot more art than science in medicine. The programs are most useful in simply speeding up the doctor's diagnosis. -R.A.
