Medicus ex machina: In this screenshot, a prototype of IBM’s Watson suggests cefuroxime as the correct antibiotic to give a patient with Lyme disease. Watson’s use of statistical correlations is apparent in the last suggestion. “Steere” is not an antibiotic; Allen Steere is the doctor who identified Lyme disease in the 1970s.
Studies have found that doctors misdiagnose patients 10 to 15 percent of the time, sometimes with fatal consequences. Herbert Chase, a professor of clinical medicine at Columbia University who is working with IBM to turn Watson into a diagnostic advisor, is still haunted by a case from 35 years ago. He encountered a patient with rickets, which causes bones to soften. He and his colleagues were mystified about the cause, because the patient didn’t have vitamin D deficiency, the usual culprit. It took four months before Chase came across a reference in the medical literature to a rarer cause of rickets, one related to a kidney malfunction. The patient survived, but she suffered more than she should have. “I’ve spent my whole life in medicine grappling with how hard it is to find information that is known,” he says.
A family medical drama motivated Jason Maude to launch Isabel Healthcare, which runs an online diagnostic-support service used by about 30 U.S. hospitals and medical practices. In 1999, Maude’s three-year-old daughter, Isabel, was being treated for chicken pox but ended up suffering a multisystem organ failure. Doctors were stumped about how to treat it; finally they realized she had necrotizing fasciitis, a severe bacterial infection. She barely survived, and Maude quit his job in finance to launch a service that could help doctors spot mysterious illnesses sooner, improving a patient’s odds and decreasing the need for unnecessary tests. Today Isabel is available for individual doctors for $375 per year; a community hospital would pay about $50,000 annually, and a large medical system as much as $400,000.
The system does some of what IBM envisions Watson will do. A doctor can log in to Isabel over the Internet and enter a patient’s age, gender, and symptoms, using either medical terminology (“hand edema,” for example) or everyday language (“swelling of the hand”). Isabel will display a list of likely diagnoses, with links to supporting evidence in medical textbooks and journal articles. As more symptoms are added, Isabel recalculates the likelihood of each diagnosis. For example, entering that a 30- to 39-year-old man has swollen hands and fatigue leads Isabel to suggest hypothyroidism as the likeliest cause, followed by less probable ones from kidney ailments to carpal tunnel syndrome.
Both Isabel and another diagnostic tool called SimulConsult, which specializes in helping doctors diagnose metabolic and genetic disorders, use information painstakingly organized by teams of medical specialists. But Watson won’t be loaded that way, which could be a strength or a weakness. Before IBM put the computing engine on Jeopardy!, its creators didn’t feed Watson all the possible clues and answers that might come up on the show. Instead, they encoded rules about English grammar into the software, then fed the program with unstructured information—massive amounts of text—and let Watson make its own connections between words and phrases that tended to appear together. From these statistical correlations, Watson inferred facts. No one specifically told the program that Theodore Roosevelt negotiated the Treaty of Portsmouth, but Watson could guess after seeing countless pages about the Russo-Japanese War.