When a clogged artery landed Peter Szolovits in the hospital for a coronary bypass operation in mid-October, he noticed a few incongruities other patients might not have. Machines that performed intertwined functions—dosing and delivering medication, for example—did not communicate with one another, and patient statistics detailed on paper were not in the hospital's electronic medical records.
As head of the Massachusetts Institute of Technology's Clinical Decision Making Group, which works to apply artificial intelligence (AI) to medicine, Szolovits knew that intelligent systems could optimize care by working together better to eliminate errors as well as avoid repetition of medical tests. Indeed, in the midst of the U.S. health care debate, some experts say that AI could lift some of the burden on physicians by helping them diagnose conditions and choose treatments.
Of course, the same claim echoed in the 1970s and 1980s, when a media blitz put medical AI on the cover of newsweeklies. Although it made inroads via various diagnostic programs such as INTERNIST and MYCIN, intelligent technologies did not revolutionize clinical care by saving lives, money and time.
Hurt by hype
One major problem was unrealistic expectations, remarks Edward Shortliffe, president of the American Medical Informatics Association. Integrating separate electronic medical records, for example, is complicated because the two sources may not share terminology and language. Usability was an issue, too: Early programs that helped physicians make diagnoses were inconveniently located outside patient rooms.
Today's AI researchers have taken such criticisms to heart and developed more appropriate software. One program that helps doctors make more accurate diagnoses was recently tested in a study conducted by investigators from the Mayo Clinic in Rochester, Minn. They entered lab test results and vital signs from 189 patients to train and test a program to assess whether subjects had a heart infection known as endocarditis. The infection and its complications kill 60 percent of the 29,000 people who develop it in the U.S. annually; tests for the condition are invasive, and can be painful and dangerous. But the software was able to definitively determine that half of the patients did not have the infection—eliminating the need for an unnecessary and risky procedure.
Real diagnoses with artificial networks
The software was based on an artificial neural network, a program that mimics the structure of biological brains and learns via adjustments in the strength of connections in its network. Researchers taught the software to recognize endocarditis by using information from medical records of patients once suspected of having the condition. The network learned to correlate each patient's unique symptoms with a diagnosis. "The network recognizes patterns," says M. Rizwan Sohail, an infectious disease expert at Mayo Clinic and lead author of the study, presented at the Interscience Conference on Antimicrobial Agents and Chemotherapy in San Francisco in September. "Just like humans, once we see the disease on a person a couple of times, we tend to associate symptoms with certain diseases."
A similar program to diagnose diseases that are more prevalent than endocarditis, such as pneumonia, would provide the greatest cost savings, and Sohail says that is one of the logical next steps in this research. "Diagnoses that are more common would show the value of the network by saving money and being helpful for the public and physicians," he says.
Even the most sophisticated programs, however, will likely need to be monitored by humans to put a diagnosis or treatment option in context. Today, as when AI programs such as INTERNIST and MYCIN were developed, "so-called expert systems have knowledge you might characterize as rather thin," says Patrick Winston, former president of the American Association for Artificial Intelligence. "They might know what to do, but they don't have any common sense. So if by some error the program suggested a barrel of penicillin, the program wouldn't know it was illogical."
Dubious doctors
And even if it works perfectly, a diagnostic AI system still might not be a major part of technology used by physicians in a clinic or hospital. Even though computers are ubiquitous in modern hospitals, novel technologies can initially slow work flow as people learn to use them. "We have to figure out how to motivate people to say, 'Yes, it's slowing me down, but in the long run it's going to make me more efficient or provide better care,'" M.I.T.'s Szolovits says.
Even if the systems flowed seamlessly into the existing practice of medicine, many experts suspect that physicians will be reluctant to use them. "Doctors don't want to lose control," says Vimla Patel, a cognitive informatics researcher at the University of Texas School of Health Information Sciences at Houston. Instead, for these programs to gain wider use they must supplement human expertise, not replace it. For instance, Patel says, "if a doctor has been working around the clock, or is aging, you want to have information that is easily accessible, retrievable and usable" to act as a backup.
By helping doctors keep track of patients' tests, make diagnoses and determine treatments, intelligent technologies could become invaluable in making health care more efficient. Combining the strengths of doctors and smart software could provide better care than either could alone, Szolovits says. "That's arguably a much better system for health care."
Reality Checkup: Medical Artificial Intelligence Still a Hard Sell in the Clinic
AI researchers now aim to improve the usability of programs aimed at assisting doctors in figuring out what is wrong with us