Med students, not computers. Image via Flickr
IBM late last week renewed its commitment to big data, the white-hot field of analytics and insight-mining. The company promised to invest $100 million in research tools, and threw its weight behind the Hadoop open-source technology used by Google, Yahoo and Facebook for managing globs of unstructured data.
During a break at the big-data confab at IBM Research in Yorktown Heights, N.Y., I had a chance to hear about another project: the education of IBM's Watson supercomputer as a medical student. Watson, you may recall, is the question-answering demolisher of former Jeopardy! champions Ken Jennings and Brad Rutter. But for more than year before the airing of the Jeopardy contest, IBM's Watson team had been prepping the underlying DeepQA software to take on the data problem of health care diagnosis. Answering the question "What's wrong with this patient given these set of symptoms and this family history?" turns out to be a natural fit for the computer.
IBM's collaborator on this project is Dr. Eliot Siegal, a senior radiologist and vice chair of informatics at the University of Maryland. Siegal saw several early attempts at software diagnosis fail in the 1970s and 1980s. Their approaches, mostly using artificial intelligence, were too brittle to make the kind of inferences that Watson is designed to make almost instantaneously, drawing on vast swathes of data to develop hypotheses and test them on the fly before delivering a diagnosis with some degree of confidence.
About 18 months ago Siegel's team in Maryland helped IBM identify which medical journals and textbooks were best to feed into the computer and which questions to start asking it. Watson read all of Medline, PubMed, dozens of textbooks and asked and answered every question on board exams. "It's all the information you'd need to be as good as the smartest second year med student," says Siegel. They tested it from time to time using the beloved clinicopathological (CPC) puzzlers that appear in each issue of the New England Journal of Medicine.(Columbia University's Medical School is also involved in the project.)
Just as Watson steadily improved at answering Jeopardy! questions until it reached the level of the most elite champions, Watson is climbing up the accuracy charts in making diagnoses correctly. I'll have more detail on that in an update.
The next, more difficult phase of the project is to load Watson up with anonymized patient records so it can marry what it knows about diagnostics with the procedures, treatments and outcomes that follow. Then doctors can query Watson and get a big assist in figuring out what to do (and what NOT to do) next. The project team is also eager to feed Watson with deeper, disease-specific data sets like the leukemia records at M.D. Anderson Medical Center in Houston or the Parkinson's records from Johns Hopkins. "Wouldn't it be great to distribute subspecialty expertise to the hinterlands where remote medical practices may lack the experience of seeing thousands of patients?" says Siegel.
Siegel says IBM-Maryland are three to five years from a real pilot test with doctors and widespread use of Watson as a diagnosis tool is more like 8 to 10 years out.
