From Jeopardy! to Medical Diagnosis
By Robert J. Murphy
Vol. 12 No. 8 P. 24
IBM’s Watson uses Nuance language processing technology in a collaboration seeking a diagnostic computer.
Over the past four years, IBM scientists and engineers have developed a computer system that’s designed to understand and interact with natural human language and has the capacity to access and prioritize vast amounts of data as a basis for comprehending questions and providing informed answers.
The Armonk, N.Y., company introduced the Watson computer—named for the firm’s founder, Thomas J. Watson—in a place sure to attract attention as well as showcase its analytic capabilities: the TV quiz show Jeopardy! It was an opportune venue—a complex question-and-answer game in which contestants must draw on immense stores of knowledge and information to consider many possibilities before selecting their answer.
In a man-vs.-machine contest reminiscent of the 1997 chess match in which IBM’s Deep Blue computer defeated world champion Garry Kasparov, Watson likewise vanquished former Jeopardy! champions Ken Jennings and Brad Rutter in a three-game match.
IBM had more in mind than playing games. Surely a system with Watson’s information processing, analytic—what they call “Deep Question Answering”—and natural language capacities must have multiple applications in a wide range of human endeavors.
Those possibilities were clearly on the mind of Eliot Siegel, MD, following the Deep Blue chess victory 14 years ago. “What are you going to do for an encore?” Siegel asked IBM’s scientists and engineers during a visit to Armonk.
Many years later, Siegel, a professor and vice chairman of radiology at the University of Maryland’s School of Medicine and chief of imaging at the VA Maryland Health Care System, recalls the exchange fondly: “And then I asked them if they had thought about doing medical applications because it just seemed perfect for diagnosis. And they said they thought that was a really cool idea, and they said they would love to work on that.”
From that exchange came a grant for the University of Maryland to begin work on medical applications of what would become Watson. It’s also emblematic of the collaborative nature of the project, which involves not only Maryland but also Columbia University, Albany’s College of Nanoscale Science and Engineering, and Nuance Communications, which contributed its voice recognition and clinical language understanding technology. The hope is that a version of Watson designed to aid in medical diagnoses will be available for testing in the next 12 to18 months.
Natural Language Processing
Given that Watson’s genesis had to do with natural human language processing, teaming with Nuance, which is based in Burlington, Mass., made a synergistic fit. “We partnered with them to take the Deep Question Answering technology, the Watson engine, and apply that in the healthcare setting,” says Nick van Terheyden, MD, chief medical information officer at Nuance. “We’re using several facets that Nuance brings—obviously the speech and the analysis of speech—but actually a much richer set of knowledge-based elements. We have a clinical language understanding component that allows us to understand what clinicians say, not just in terms but actual meaning.”
The clinical language input adds to the Watson database. “We’re then able to apply that to the Watson engine and bring even more value to that large referenced knowledge base,” van Terheyden says. “What Watson does is it consumes large amounts of information in an automated way and then presents it as a framework that we can ask questions of it. And that really typifies the clinical setting, this massive amount of clinical data which we as clinicians try to consume.”
So we have a case of mutual expression and consumption in which both sides benefit. Think of it as having perhaps the world’s most knowledgeable medical consultant right at your shoulder, ready to give as well as take.
In 2011, it’s a good bet that there is more information on the planet than ever before. Some of it is organized and accessible.
It’s anyone’s guess where the divide is when it comes to medical information. Surely the Internet has helped bring online massive amounts of information, mostly in journals but to some in degree books as well. Yet too much information can be little more helpful than a dearth. Type in a commonly used keyword and you’re awash in a deluge of text with no clue of where to begin. The Cochrane Library database itself includes some 645,000 titles.
“As a physician, I can go online,” says Martin Kohn, MD, chief medical scientist for care delivery systems at IBM Research. “I could go into a search engine and put in certain specific search words, and I might get a list of 500 articles or textbooks that were valid for the keyword search. Then I would have to go through those and maybe I’d go through the abstracts and, say, 20 seem like they may be relevant. And then I’d read those articles and add to them ones that are actually more relevant.”
As a former emergency physician, Kohn knows that such time may not always be available to the physician.
Watson seeks to offer a far more efficient and timely approach. Numerous collaborators are in an ongoing process of building it, with input from journals, books, lectures, and ultimately patient-provider interactions. The fact that Watson is the ultimate speed-reader—100 million pages per second—helps accelerate the process. But it must also be a discriminating effort, for example, including pediatric articles from the The New England Journal of Medicine while excluding children’s articles from Highlights magazine.
van Terheyden has given this some thought. “IBM and Nuance, as we progress with this project identifying the appropriate sources and the way we can access [them], it’s clear that there are many resources available to us,” van Terheyden says. “There are a couple questions. One is, ‘Which ones should we access?’ Should we give greater credence to, say, the Oxford Textbook of Clinical Medicine over WikiMed?”
The key to making Watson’s medical database efficient and user friendly is that it prioritizes a search based on the amount of detail entered. Here again, it’s an interactive process between user and machine.
“Let’s say I’m interacting with a patient, and I obtain some information from this interaction,” Kohn says. “The patient is complaining of chest pain or has this finding, that finding. We get some lab tests and Watson gets this information, and Watson will understand the natural language that it absorbs from this encounter. It will go out into the literature and read 100 million pages a second and come back to me with direct suggestions—suggestions for diagnosis, suggestions for treatment, and a list like a differential diagnosis, with assigned confidence levels based on the evidence it has found in the world.”
The Watson-based database may be beneficial not only in the clinic but for those designing clinical trials of pharmaceuticals and devices as well as in teaching settings, whether hospital based or in the classroom. Postmarketing reports of drug-related adverse events can be processed more quickly in such a database rather than through a cumbersome bureaucracy.
Diagnosis and Treatment
A medical diagnosis comes down to an aggregation of data, including a patient’s history, family history, allergies, lifestyle, symptoms, and signs. The most clinically useful patient history is a matter of assigning a degree of severity to each abnormal finding and designating a level of relevance and confidence of each as it pertains to a suspected diagnosis. The best doctors do this intuitively; Watson does it systematically and automatically.
“Let’s say it gives you five possible diagnoses for why a patient complains of chest pain,” Kohn says. “It’ll come back to me and say, ‘I need some more information to help adjust my confidence levels in these [more likely] diagnoses, so tell me more about the patient’s chest pain. Is it sharp? Is it burning? Is it constant?’ Being able to know about the nature of the chest pain may be more important than knowing that the patient has chest pain.”
It’s a matter of boiling down many possibilities to a few. “There would be an interactive process,” Kohn says. “And it may come down to, I’ve got four diagnoses left with various confidence levels. Watson says there is a test that, based on the literature, will provide the information that is most needed. And since Watson knows when it needs to learn more, it can go into the literature. And if the answer’s not there, it must be something I must get from the patient. So I’ll have to come back [to the patient] and ask for more information or suggest a test because the literature tells me that this test is the next best thing to do.”
The IBM technicians and doctors and the project’s consulting physicians place their faith in evidence-based medicine. That is, most diagnostic and therapeutic conclusions can be based on information from well-designed clinical studies published in peer-reviewed journals.
Watson’s consult doesn’t stop there. Once a diagnosis has been determined, Watson can help determine the most effective and least burdensome treatment. To that end, it will take into account specific elements of a patient’s history or the eccentricities of the condition at hand, which may call for some modification of the standard treatment.
“What it will do is make a recommendation for treatment,” Siegel says. “And then it will refine that. So it will ask, ‘Is the patient pregnant? Does the patient have any allergies? Have there been previous episodes of this condition?’ When it makes treatment recommendations, it will refine those based on that information.”
To Err Is Human
It’s hardly a well-guarded secret that each year many medical errors occur in multiple healthcare settings. Machines can malfunction as well, but early evidence from the Watson project suggests that its setup reduces the incidence and severity of iatrogenic mishaps. “The issue of medical errors is a big one, in medical training and healthcare in general,” says Sara Brenner, MD, MPH, assistant vice president for NanoHealth Initiatives at the College of Nanoscale Science and Engineering. “With technology, we have an opportunity to really reduce our errors.”
And the Answer Is…
There’s little reason to conjure dystopian scenarios of robots on the loose and running the clinic, far more clever and decisive than any mere human. Nor does the horizon hold any vistas of former physicians selling ice cream on the boardwalk or mopping men’s rooms, having been replaced by a machine. Physicians who have been involved in Watson’s early development sound as enthusiastic about the project. For all its sophistication and mind-boggling computational capacities, Watson is nothing more than a clinical tool to help physicians do a better job.
— Robert J. Murphy is a freelance writer based in Philadelphia.