Skip to main content

Advanced Search

Advanced Search

Current Filters

Filter your query

Publication Types



Newsletter Article


Q&A: Using Machine Learning to Sort Through Treatment Options

IMPORTED: __media_C062CB94E76D4CBCAF854DBDEFE6CC84.jpg

Watson, the IBM computer system that famously beat two human contestants on Jeopardy!, ventured into medicine in 2011 when IBM partnered with WellPoint, Inc., to make the process of approving medical tests and procedures for plan members more efficient. Watson, which can understand the nuances of English well enough to discern the meaning of "a house could burn up as it burns down," is also capable of finding meaning in information drawn from medical journals, books, blogs, and tweets as well as medical records and case reports. Quality Matters spoke to Stephen Gold, vice president of IBM Watson Solutions, about Watson's potential to help physicians make diagnoses and explore treatment options.

By Sarah Klein

Quality Matters: You've said what Google is to search Watson is to discovery. What do you mean by that and why is it important as Watson delves into medicine?

Gold: If you think about your own experience doing searches on the Internet, you get millions of results that you have to sort through to find the ones that are contextually relevant. Watson is a little more thoughtful. It's looking to understand the context of the question and bring back not an answer, but a set of responses that are aligned to the questions, supported by evidence, and weighted by confidence. The reason that's so important, especially in health care, is that there is often more than one possibility of what's wrong when you describe your symptoms to a doctor. The cause could be viral or bacterial and even if you know it's bacterial, you have to consider a host of other things: which bacteria, what comorbidity issues are present, what medications you're taking, your family history, and what's happening in the general populace at any given time. For instance, if you knew there was an early outbreak of pollen that year because everyone was blogging and tweeting about it, you may draw a different conclusion if your throat is scratchy.

Quality Matters: How is Watson able to make sense of such disparate information?

Gold: It's very different from the computers that have been around since the 1950s. Those are largely built on a programmatic approach. Data is structured fields—rows and columns, zeros and ones, and the systems are guided by rules and logic. These work well if all you need to do is retrieve all the customers whose names start with the letter "A" and made a purchase in the month of July. That's pretty straightforward. Watson on other hand isn't programmed; it's taught. We will give it a question and give it an answer and do that over and over so that it continually gets smarter. In essence, it uses machine learning and various algorithms to advance its understanding based on new information, iterations, and outcomes.

Quality Matters: How is it being used in medicine?

Gold: Its first application in health care focused on the exchange of information between a physician and a health plan—that is, the process by which a treatment or test is reviewed and determined to be appropriate or inappropriate. The exchange of that information is now slow, antiquated, and relatively cumbersome. A nurse has to review the request and it often take about 20 minutes to adjudicate whether it's appropriate. To do this job, we fed Watson 25,000 training cases and gave it the answers until he could make the determination almost instantaneously. That's a relatively simple example. A more complex one is the second application. We've partnered with Memorial Sloan Kettering Cancer Center, WellPoint, and other leading cancer institutes to train Watson to assist clinicians in identifying individualized treatment options for patients diagnosed with different types of cancer. Right now, there are 36 oncologists at Memorial Sloan Kettering that are teaching Watson about treatment options.

Quality Matters: But if the teachers are human, and therefore fallible, isn't there a good chance some of the lessons will be wrong?

Gold: Watson is not beholden to one source of data—it also picks up information from other clinicians, researchers, and educators who are working with Watson as well as information from medical journals, studies, reports, guidelines, historical records, and test results. By the same token, the recipient need not accept what resources Watson gives it. Let's say a doctor receives three treatment recommendations from Watson. Watson might say, "I'm 96 percent confident this would be the preferred course of treatment, 82 percent confident this alternate course would be good, and 32 confident in the third treatment." If the doctor prefers the third course, he could ask Watson what he's basing his decision on. It could be a clinical trial the doctor doesn't have confidence in and if so, the doctor can tell Watson to exclude that information. Now Watson says, "I'm still 82 percent on the second choice, but the first one has dropped to 80 percent." We call Watson a humble genius, because it's always willing to share how it reached its conclusions. It'll say here's the evidence, here's how it was used, and here's the confidence level.

Quality Matters: What's the plan for disseminating what Watson learns?

Gold: Memorial Sloan Kettering's vision is to take this to the masses by sharing the information with community cancer centers and other clinicians across the country. The hospitals and clinics that use it can also feed their treatment and outcomes information to Watson so it can learn from that. That is particularly valuable when one clinic sees only two or three cases of a particular cancer each year.

Quality Matters: What will Watson be studying next?

Gold: In a year, Watson will be dealing with the majority of cancers. We had assumed—inaccurately—that we would have to systematically train Watson in each area of cancer: lung, breast, and so on. What we found out when we moved to colon cancer is that Watson was learning on its own. It was basing some of what it knew about colon cancer on what it learned about lung cancer and breast cancer, so out of the gate it was much more accurate. That's true of humans. If you teach someone the basic elements of subject matter, then throw out a complex mathematical question, the person is going to go back to basics to make sense of it.

Quality Matters: How do you see this technology changing the day-to-day practice of medicine?

Gold: I think it will give providers and others including payers and pharmaceutical companies more dexterity and the opportunity to perform tasks at a higher order. Instead of spending time looking for data, the providers may be able to interact with patients more and they may also feel more empowered. I think it will also bring about some efficiencies. For instance, Watson could interact with patients who are calling the doctor's office or nurse help line for advice such as "Can I drink with this medicine?" or "I'm having this side effect. Is that normal?" Watson can give responses, explain where the information came from, and relay its level of confidence in the answers. The exchanges could be shared with a doctor or nurse as a safety check.

Publication Details