Seven Questions With IBM’s Manoj Saxena About Watson and Cancer

It’s been nearly a year since a talking computer stunned humanity by beating the world’s best players at the TV game show “Jeopardy.”

And while it was something of a publicity stunt to put a sophisticated and specialized IBM computer in people’s living rooms, the fact remains that Watson is, well, a pretty sophisticated and specialized computer.

Since schooling humanity at “Jeopardy” — which was the subject of a book — Watson went on to get a real job working for the health insurance company Wellpoint.

Now IBM has decided it is ready to tackle something a little more involved. Watson is about to go to medical school, and will even study a specialty: Oncology. Sometime this year, after studying and even taking exams to prove what it has learned, Watson will be assigned to assist human physicians in the treatment of breast, lung and colon cancer.

If this seems like kind of a big deal, it is. Watson won’t be the first computer to serve as a reference tool, helping doctors do their jobs. But then there has never been a computer quite like Watson, which can learn so readily from natural language — and play TV game shows and win.

Last week, I talked with Manoj Saxena, general manager of the Watson program at IBM, to talk about what Watson will — and won’t — be doing in helping doctors treat humans with cancer, and what that might mean for the future of medicine.

My first question was about what Watson has been doing since its big win:

AllThingsD: So, Manoj, last I knew, Watson had been working for Wellpoint, which is a large health insurer. What exactly has it been doing?

Saxena: Let me bring you up to speed. In August, we announced the first commercial relationship of Watson with Wellpoint, one of the nation’s largest health insurers. They have 35 million customers in 14 different states. One out of nine Americans are covered by them. The first area was around utilization or approval. Let’s say you or I call up a clinic or hospital saying we have flu-like symptoms. Where Watson would come in is on the approval process, saying we’re covered. Then Watson looks at the history that the hospital has in its records. It might say that it’s early December, and I come in at this time every year saying the same thing; and the last two times it was a ragweed allergy, not the flu. And the medical journals say there’s a connection between ragweed and fever that looks like the flu. And by the way, the newspaper says there was an outbreak of ragweed in Central Texas. And then, in addition to treating for flu, also look for allergies. So Watson is considering the medical record; the patient history that the insurance company has; and third, the medical journal and news information about what may be causing a certain thing. So that’s what it’s doing with Wellpoint so far.

How then do you make the pivot to working with cancer?

We’ve installed another adviser — these solutions are called Watson Advisers. This one is called Watson Oncology Adviser, and this is a big one. As you may remember, medical information is doubling every five years. Doctors tell us that they are spending only five hours per month going through new information in medical literature. On one hand, you have all this medical information coming out. We’ve decided to focus first on breast, lung and colon cancers as the three to apply Watson to. And Cedars-Sinai has partnered with Wellpoint to help come up with the right cancer solutions. And the point is to build the expertise within Watson to help treat cancer.

So Watson won’t be directly involved with the treatment, but rather to build up its own knowledge base?

Watson doesn’t make the decisions. It’s a physician’s assistant. But before it becomes that, it has a lot to learn. Out of the box, Watson has the knowledge of a first-year medical resident. That is where it’s at today. With Cedars-Sinai and Wellpoint, we’re going to teach it all about cancer during the next six months. We’re going to show it actual cases that were solved in the past. And over time, we’ll tweak and teach it, using things we already know.

Is there a human analog to this process?

A good human analog is how we learn. As children, our teachers and parents sit with us and ask questions to understand how well we learned from what we read. And then, later, we learn by doing. This will address the first two phases. Watson will read on its own, and then oncologists are going to ask questions of Watson to understand how well he has learned and then understood. And then once we feel comfortable that it has learned enough, then we will let it begin working as a physician’s assistant, and then it will go from there.

Since, in the end, there are humans being treated, do you have to get any kind of regulatory approval to do this?

No. It’s very similar to how doctors refer to medical journals. Doctors might turn to Google or something like that to look up info from their medical journals. That doesn’t require any approvals. Someone else asked me what happens if Watson suggests a particular treatment, the doctor accepts it, and the patient dies. Or what happens when Watson suggests something and the doctor doesn’t take his advice. Our view is that it’s the same as looking up textbooks and information. The physician is the one who makes the final decision.

And that will always be the case?

That will always be the case, yes. We are far, far away from computers doing medical treatments. I don’t even see it in the forseeable future.

How do you actually go about feeding information to Watson? How does it learn?

It’s a good question. There are four different types of information that’s fed into Watson. At the base of the pyramid, it’s general information like Wikipedia and Google and general information like that. And a lot of that is general knowledge; and a lot of that is already in place, because we needed that to play Jeopardy! Then the second layer is the medical textbooks and medical journals and vocabularies, and those are fed in as natural-language information. It can be any scanned information or text information because Watson understands natural language. So that information is the second part. It can process text and tables, but it can’t process pictures and videos, but we’re working on that. And then there’s the actual test cases, the information on people with 30 years of cancer treatment history. We feed that into what are called “answer keys.” The fourth layer are new domain-specific information models that are specific protocols and procedures that the health insurance companies will want to feed into Watson.

Where do you draw the line? There is an accepted mainstream body of knowledge and accepted treatments for different cancers, and then there are newer things that may be controversial for some reason.

The way we approach it is in two parts. One is the body of knowledge that is already known. But it does not get applied and in context, and often doctors don’t have access to it in context. There are things like cancer treatment guidelines and well-understood things about radiation and effects on different cancers. Call them the known treatment pathways. The second are the emerging treatment pathways, particularly in the area of genomics. That is the one that can get added on. It’s the one our partners are looking at. In about a decade, most cancer treatments are going to shift to genomics-based treatments, rather than chemotherapy-based treatments. There’s a deluge of information about converting the knowledge about DNA into biological knowledge, and then converting that into treatment knowledge. That is the second part of what we’ll be doing.

Do you have other diseases that you think Watson can help treat in the future?

Yes. Diabetes and cardiology, heart problems are next on the horizon. We’ll also be applying Watson in financial services.