Brian Buntz

May 20, 2014

4 Min Read
5 Ways Watson Has Evolved Since Winning Jeopardy

IBM's Watson artificial intelligence platform has evolved significantly since it handily defeated human opponents on the game show Jeopardy three years ago. Shortly after winning, IBM began to retool the platform for the medical field where it can be used to improve the accuracy of clinical decision making and ultimately be integrated with medical devices.

To learn more about how IBM has tailored the technology for the medical market, we spoke with Eric Brown, Director of Watson Technologies at the IBM T.J. Watson Research Center in Yorktown Heights, NY. Brown will be elaborating on the power of IBM Watson in medicine in a keynote address on June 10 at MD&M East.   

Here are five things IBM has been doing with Watson since the Jeopardy days:

1. Getting Watson into the Real World

The Watson platform that won Jeopardy was developed by a research branch of IBM, Brown says. "The first step to explore commercializing Watson was to engage our professional software and engineering part of the company and work on creating a hardened product-ready version of the software," Brown says. "All of our work in the first year after Jeopardy was just transitioning the technology from a research environment to a commercial development environment addressing things like performance, speed, scale, efficiency, and starting to build out some of the administrative management tools that you need to support the technology."

2. Teaching Watson Medical Jargon

IBM immediately identified medicine as the next domain for Watson after the software won jeopardy. The first challenge in applying the platform to medicine was boosting its natural language processing abilities to be able to understand the nuances of medical language. Doing so meant that Watson would have to understand both common expressions for medical problems as well as their clinical equivalent. "A patient might complain of chills and have a fever and think they are suffering from a cold. The medical professional might use a more technical term--coryza--to describe that," Brown says. Not only must Watson understand that "coryza" and "cold" are potentially synonymous, being able to disambiguate the word "cold" depending on context is important. "Am I talking about suffering from the common cold or am I talking about the sensation of being cold, which might be associated with having chills?"  

See Eric Brown, PhD, director of IBM's Watson Technologies, deliver a keynote address at MD&M East, June 9-12, 2014 in New York City.

3. Boosting Watson's Understanding of Anatomy

Dealing with anatomical and spatial reasoning is another important consideration. "A patient complaining of lower back pain could actually be experience kidney problems," Brown says. "There are a variety of ways of talking about a location in the body.

4. Tackling Complex Problems

The original Jeopardy was focused on question answering. "The Jeopardy clues are these fairly complex, natural language questions ultimately designed to identify a single unambiguously complex answer," Brown says. "When you look at a problem in the medical domain, it is actually possible that there are multiple correct answers. Some might be more correct than others. But the input is usually ill defined and very vague."

One of things IBM has been developing since then is the ability to analyze complex problem scenarios and then automatically break them down into smaller pieces, into sub-questions and to connect these different pieces together in a graph format. "It ends up being similar to the way the humans reason complex problems," Brown says. "We might diagram out the individual symptoms, what we call 'factors.' We will then try to connect those symptoms to intermediate diganoses or other relevant pieces of information that can be connected together to allow you to infer other things and ultimately get you to a final answer."  

IBM calls this functionality WatsonPaths. "It gives you a graphical representation of essentially the thought process of how you take a complex problem, break it down to relevant pieces to potentially get a final answer typically supported by textual information," Brown explains."Part of this WatsonPaths approach is the ability to demonstrate the evidence used to support this whole graph and trace the paths that get you to the answer. That can allow a human to more quickly understand what is going on."

5. Making Watson Interactive   

IBM envisions that Watson will be able to interact with end users in a way similar to a human advisor would. "The end user can look at some of the options that Watson is considering and actually provide feedback or corrections and say: 'this part is relevant. But I know that this other part is irrelevant, so stop considering that,'" Brown says. As the clinicians review the evidence, they also might trigger other things to explore or other ways to phrase the question. "Sometimes the problem isn't so much coming up with the right answer, it is coming up with the right question in the first place," Brown says. "Having a more interactive and visual way to work through these problems can trigger different lines of attack for the human problem solving process."

Brian Buntz is the editor-in-chief of MPMN. Follow him on Twitter at @brian_buntz and Google+.

Sign up for the QMED & MD+DI Daily newsletter.

You May Also Like