IBM’s Watson Is Learning Its Way To Saving Lives

A few years ago, IBM’s new computer was a game-playing curiosity. Now Watson is poised to change the way human beings make decisions about medicine, finance, and work.

IBM’s Watson Is Learning Its Way To Saving Lives
Photo by Dan Winters

The woman was gravely ill. Her name was Ms. Yamato. Thirty-seven years old, born in Osaka, Japan, she had never smoked, and yet there it was anyway: a spot on her lung.


A doctor had already performed a bronchoscopy and had made the diagnosis of cancer. Then he referred the patient to Mark Kris, an oncologist at Memorial Sloan-Kettering Cancer Center in New York. Seated alongside me in his office on the Upper East Side of Manhattan, Kris is showing me Ms. Yamato’s electronic medical record on an iPad. “I’m preparing for the first visit,” he explains, swiping the screen to show what that entails. He’s interested in running at least two tests on the patient. The first is an MRI, to find out if the cancer has spread to her brain. The second involves a deeper diagnostic regimen. Lung cancer tumors are not all the same; there are thousands of variations. So a test that examines the mutations within a tumor will be crucial, he says. It so happens that cancer patients born in East Asia who have never smoked often have a particular mutation that responds well to a medication by the name of Erlotinib. That may be the case here. One can hope.

The woman is not real. She happens to be a character within an app that IBM has created for Watson, its new computer. Watson’s special talent, its reason for being, is a singular ability to grasp the intricacies of human language and answer exceedingly difficult questions. You may have heard about Watson already. Back in 2007, a group of computer engineers at IBM’s research labs in upstate New York began building the machine–named for IBM’s founder, Thomas J. Watson–with the goal of creating a question-and-answer technology that would be more authoritative and powerful than anything on the planet. The initial objective of the Watson group was simple: to win in the game show Jeopardy!, something Watson famously achieved in February 2011. Yet the group had a far more important goal: to turn Watson into a business, hopefully one of some scale. So starting in late 2009, a business development team at IBM began holding meetings outside the company in an effort to understand the ultimate worth of this new technology. No doubt it could be a business one day. But what kind of business?

“The first thing that hit us about Watson,” recalls John Kelly, IBM’s chief of research, “was that this thing could be applied almost anywhere.” Early on, IBM executives decided to focus on a field in which Watson could have a notable social impact while also proving its ability to master a complex body of knowledge. The team chose medicine. They believed Watson could help doctors make diagnoses and, even more important, select treatments. Specifically, they thought Watson could be the perfect tool to chart the complex decision trees that cancer specialists like Kris negotiate every day as they weigh treatment options that might involve radiation, surgery, and any of countless chemotherapy drugs. Watson can ingest more data in a day than any human could in a lifetime. It can read all of the world’s medical journals in less time than it takes a physician to drink a cup of coffee. All at once, it can peruse patient histories; keep an eye on the latest drug trials; stay apprised of the potency of new therapies; and hew closely to state-of-the-art guidelines that help doctors choose the best treatments. Watson never goes on vacation. And it never forgets a fact. On the contrary, it keeps learning.

Manoj Saxena is in charge of commercializing Watson for IBM. He calls it the most meaningful endeavor of his life. | Photo by Dan Winters

This fall, after six months of teaching their treatment guidelines to Watson, the doctors at Sloan-Kettering will begin testing the IBM machine on real patients. The Ms. Yamato app shows how it will work. After Kris inputs the results of her medical tests, Watson begins deliberating. “It’s going through its algorithms,” Kris says as we stare at the iPad. “It’s seeing where the data sends it today.” On the screen, a colorful globe spins. In a few seconds, Watson offers three possible courses of chemotherapy, charted as bars with varying levels of confidence–one choice above 90% and two above 80%. “Watson doesn’t give you the answer,” Kris says. “It gives you a range of answers.” Then it’s up to Kris to make the call. He regards the options on the screen and wonders how they might change if Ms. Yamato happened to develop a common symptom: hemoptysis, or coughing up blood.

“Let’s try that,” he says. He inputs the information and shows me the result approvingly. Watson has dropped one drug from the top chemo regimen. That’s just what Kris would have done.


To make sense of all this–that is, to gauge both the value of Watson to a hospital like Sloan-Kettering and its potential to change forever the worlds of medicine and business–you could follow two different paths. You might consider Watson’s evolutionary promise. Watson can almost certainly generate huge administrative benefits. Already, one large health insurer–Indiana-based Wellpoint–has begun using a Watson computer in its Virginia data center to speed along the authorization for medical procedures. Usually, authorizations are evaluated by a team of trained nurses and can sometimes take weeks to come through. Watsonizing the process would speed it up–a boon for a doctor like Kris, who now must wait while assistants exchange faxes with insurers before he can get clearance for any expensive tests.

Kris shows me what happens when Watson’s treatment plan calls for an MRI. A button pops up on his screen to ask for preauthorization. “I just click that,” he says, and it’s done instantly.

I ask him what if Watson’s request is denied.

Kris seems amused by the question. Watson has already consulted the latest medical literature, and it’s been trained by the best cancer doctors in the world. “Who is the authority that is going to trump that?” he asks. Insurers balk at paying for unnecessary procedures; Watson’s expert opinion essentially guarantees the necessity.

But the more intriguing path is the second one–a consideration of Watson’s potential to do something revolutionary. This is the trail that captivates Kris. Eventually, he thinks, Watson could provide any doctor anywhere with the world’s best second opinion. A physician in a community hospital in the Midwest, or at a remote medical center in China, could have instant access to everything that the medical field’s best oncologists–people like Kris and his colleagues at Sloan-Kettering–have taught Watson. What is more, Watson will be able to excavate facts beyond the ken of Sloan-Kettering’s current lineup of specialists. As Kris says, “We could ask Watson: What is the best treatment for this rare condition based on all of Sloan-Kettering’s records?” It could then go through several years of cancer cases looking for the most successful outcomes. In time, it could even look at hospital records from around the world. As Manoj Saxena, the IBM executive now in charge of commercializing Watson, tells me: “It’s like being able to take a knowledge worker–cancer specialist, nurse, bond trader, portfolio manager, whatever–and equip that person with the best knowledge, and have it available at their fingertips.” As Watson evolves, Saxena believes, these knowledge banks will significantly alter how, and how well, humans make decisions.


Within a few years, for instance, Watson may be reaching well beyond oncology to assist patients suffering from any chronic disease and help general practitioners make diagnoses in their offices. Ultimately, Saxena believes, Watson could play an essential role in the diagnosis and treatment of mental health; in the financial services industry, where Citibank is testing it now; and in education. It could become the world’s smartest dietitian.

Saxena now commands a team of about 200 people who are working to adapt Watson’s skills for various IBM clients. He and I are discussing his progress over lunch one day near IBM’s upstate New York headquarters when he leans back and tells me that after creating two successful tech startups, both of which he sold (the second to IBM), his current job is far and away the most meaningful endeavor of his life. Those startups, he confides, were exciting, important. “But this,” he says of the Watson rollout, “this is stuff that is going to change the course of history.”

Over the past year, IBM executives have come to believe that Watson represents the first machine of the third computer age, a category now referred to within the company as cognitive computing. As Kelly describes it, the first generation of computers were tabulating machines that added up figures. “The second generation,” he says, “were the programmable systems–the mainframe, the first IBM 360, PCs, all the computers we have today.” Now, Kelly believes, we’ve arrived at the cognitive moment–a moment of true artificial intelligence. These computers, such as Watson, can recognize important content within language, both written and spoken. They do not ask us to communicate with them in their coded language; they speak ours. And perhaps most important, they can learn, so they improve without constant human instruction.
Siri, on the iPhone, might be considered an elementary example. Watson is industrial strength. “Computers do numerical calculations, they move data around, and they’ve been doing that forever,” David Ferrucci, the IBM researcher who commanded the team that built the first Watson computer, tells me one day at IBM’s research labs. “When I think about Watson, it’s interpreting the information in human terms. It’s saying: What does this mean to me? And that’s a big deal.” Also significant is how Watson renders an answer. Unlike its responses in Jeopardy!, in the real world it will perform as it did for Kris at Sloan-Kettering–by giving not a single solution but a range of probable solutions, each backed up by Watson’s evidence and ranked by its level of confidence. In the lingo of computer science, that makes the machine probabilistic rather than deterministic. One might say this trait gives Watson a humanizing glow of humility and diminishes concerns that it marks a stride toward a computer-led dystopia. Watson, in IBM’s marketing schema, is here to help with our questions, rather than solve them. In the case of medicine, it–for Watson is not really a he–is here to support doctors, not replace them.

The Watson of today is not precisely the same machine that won in Jeopardy! IBM has fine-tuned its software and algorithms for medical applications (or, in the case of Citibank, financial services applications). Watson has shrunk, too, from a row of about a dozen server racks that would have filled a small bedroom to an assemblage about the size of a double-door refrigerator. But for all the concentrated power, it doesn’t look like anything special. Its sleek black servers are standard IBM Power 750s. You could wander around Watson and regard its blinking lights, as I did on a quiet midsummer afternoon at IBM’s research labs, and not think something unusual is happening inside it. But there is. The way Watson solves problems–or, rather, the way it looks for answers, simultaneously sending out thousands of inquiries in all directions and then scoring the evidence it collects–is different from how other computers work. One person at IBM likens Watson’s process to (1) gathering hundreds or thousands of possible solutions from a vast data bank, (2) pouring them into a giant funnel, (3) stirring with a dash of algorithms, and (4) letting only the best drip out of the bottom.

At the moment, a half-dozen Watsons are scattered around the country. Some are on the premises of IBM clients, as with the insurer Wellpoint, while others are cloud based, which is how hospitals such as Sloan-Kettering will access Watson. “Effectively, there’s no limit to how many Watsons there can be,” Bernie Meyerson, IBM’s VP of innovation, tells me. Watson is a creation of software, not hardware. “That’s the beauty of it,” he says.


Watson is different from big servers and mainframes in other ways, too. The best computers of today have the extraordinary processing power needed to create, say, complex supply chains for building a new automobile or planning a satellite launch. These machines are good at manipulating the vast amounts of clearly defined data–numbers and facts–known as structured information. But most of the world’s information is more ambiguous and less precise and lies beyond their reckoning. “We now have this proliferation of what we call Big Data,” Saxena, Watson’s business manager, tells me, referring to the flood of information created by our computers, our electronic sensors, and ourselves. “Ninety percent of the world’s information was created in the last two years,” he says. “But 80% of that 90% is unstructured or semistructured information, like doctor’s notes or product reviews on Amazon.” This near infinitude also includes tweets, blogs, emails–all the noise and scribble of modern life. So any company that aspired to manage the data of all the world’s businesses would today be able to analyze only a small part of it. Watson, though, is a genius at reading unstructured information. And it’s precisely this facility that explains why IBM sees such a rich business opportunity here.

It likewise explains why medicine is a logical first choice. While some health information is indeed structured–think of blood-pressure readings or cholesterol counts–the vast majority is unstructured. This cache includes textbooks, medical journals, patient records, and nurse and doctor evaluations. In fact, medicine embodies so much unstructured information that its proliferation has, by the account of many medical professionals, far outstripped the ability of doctors to keep up. Neither better training nor continuing education could ever wholly remedy this problem. When I meet with Herbert Chase, a professor of clinical medicine at Columbia University who consulted with IBM during the early stages of the Watson project, he says it is “not humanly possible” for a busy doctor to keep abreast of the current literature.

One result of information overload is a high rate of misdiagnosis and consequently incorrect treatment. By some estimates, Saxena tells me, 20% of initial diagnoses of cancer are eventually altered. “Imagine the implications of cancer care if there is a one in five chance that for the next six months whatever therapy they’re giving you is wrong,” he says.

Deciding on a course of treatment is even tougher than making a diagnosis. “It’s still possible for a doctor to know the ways that people get sick,” says Chase, who is also a kidney specialist. “But what is unmanageable, and what has been for decades, is knowing what the best option is today.” Some applications now available to doctors are meant to alleviate this problem; one popular web-based tool is named Isabel. But Watson, in Chase’s view, reaches a different level of sophistication. “I’ll give you an example of a test we thought up for Watson,” he tells me one day in his Manhattan office. “A patient was pregnant, had Lyme disease, and was also allergic to penicillin. And Watson came up with a drug. The first thing I thought was, Watson made a mistake. That drug can’t be given to someone allergic to penicillin.” But Chase was wrong, not Watson. “My knowledge was about five years old,” he says. “And in the past couple of years, all the muckety-mucks had reviewed all the studies and had concluded yes, you can give that drug to someone who’s allergic to penicillin.”

To Chase, this proves a point: If you’re a patient, you don’t want to believe your doctor doesn’t know everything. But he or she doesn’t, and can’t. At its best, the dispensation of treatment is inefficient today. “At its worst,” Chase says, “it’s subpar, incorrect, wrong therapy,” and doesn’t reach the standard of care to which his profession aspires. “As you can imagine,” he adds, “this is not something we like talking about.”


Last year, IBM turned 100 years old, which sets it apart from West Coast counterparts like Amazon, Apple, Google, HP, and Microsoft–all younger and ostensibly the tech world’s leading innovators. To delve into IBM’s recent research, though, is to wonder if our perception of technological leadership sometimes suffers from the distortions of branding and familiarity. We use iPhones and search engines and laser printers every day. But IBM’s technologies are lodged deeper within the infrastructure of daily life; you’re tapping into them whenever you send an email, for instance, or log on to a website. IBM has been granted more patents than any other company in the world for 19 years in a row. Yet since getting out of the laptop business in 2004, it has not produced a single product that it sells directly to the consumer.

To understand how Watson figures into the company’s culture of ideas, or to see how it represents the kind of large-scale innovation that arguably lies beyond the capabilities of any startup, it helps to understand what the company actually does these days. IBM has operations in 172 countries and an organizational chart that resembles a vast Soviet bureaucracy. It employs about 433,000 men and women. Though IBM still sells hardware–big mainframe computers, silicon chips, and supercomputers–mainly it makes money selling software and consulting services to businesses and governments. The company’s strategy has been validated of late by its performance: IBM’s stock price has been on an upward trek for the past five years, and its winning streak has attracted the likes of Warren Buffett, who last year decided the company merited an investment of $10.7 billion. Meanwhile, as one of the few global titans to invest staggering sums on R&D ($6 billion to $7 billion a year), IBM maintains one of the world’s last great industrial laboratories. At its main research center in Yorktown Heights, New York, a jet-age dream of glass curtain walls and rusticated stone designed by the Finnish-American architect Eero Saarinen, IBM employs the bulk of what is likely the world’s largest mathematics department, with 300 members. If you’re looking for a new PC design, you’re out of luck here. But if you’re shopping around for a new or better algorithm, IBM can build you one.
Not everyone is impressed by the direction of IBM’s management. A relentless focus on earnings and cost cutting has led to a significant offshoring of domestic jobs, and a vocal corps of disillusioned or laid-off IBMers regularly take to the web to lament that the company’s best days are behind it. IBM has also had its share of technological stumbles, apparently bungling several high-profile government contracts in recent years (in Texas and Indiana, for example) that left the company embroiled in disagreements with unhappy clients. And though these flare-ups may be uncommon, the company otherwise rarely quickens the pulse, with a long-standing reputation for being slow, steady, reliable, and maybe a little dull. IBM doesn’t have big growth spikes or ballyhooed product launches; rather, it has plodding, long-term client contracts built around its ability to help optimize, say, a company’s global IT services or a public utility’s electrical grid. The corporation moves along like a supertanker. “IBM’s annual revenue base is huge–$100 billion,” says Toni Sacconaghi, a technology analyst for Sanford C. Bernstein. “So to move the needle is tough. It’s hard to find big new products.”

The managers and engineers keep looking anyway. One way IBM tries to infuse the troops with a sense of mission is through its periodic attempts to create for itself a Grand Challenge, such as the construction of Deep Blue, a chess-playing computer, or, more recently, Watson. The Grand Challenges are focused and expensive efforts–IBM will not verify Watson’s cost, but estimates put the sum between $100 million and $1 billion–to push the company beyond the competition.

Watson’s origins can arguably be traced back some years to a more modest annual initiative IBM calls the Global Technology Outlook, or GTO. Anyone at IBM can contribute to the outlook, and most of the results are eventually made public. The GTO tries to identify future business opportunities by putting a spotlight on various technology trends. A while ago, the IBM outlook pointed to analytics as a potentially huge field. Not long after, then-CEO (and current chairman) Sam Palmisano green-lighted IBM’s acquisition of about $16 billion in smaller companies that had computer technologies to do this kind of work–essentially, to comb through vast stores of data, both structured and unstructured, and help extract nuggets from the global corporate babel.

Like Big Data or cloud computing, analytics is one of those contemporary catchphrases that everyone talks about but no one pauses to define. Bernie Meyerson, IBM’s VP of innovation, argues that the great promise of analytics is not just to spot trends or glean information for boosting sales but to use computers and software to change the future. “Analytics is the capability to see what no human can,” he says. Recently, at a public event, Meyerson was asked if IBM missed out by not building a tablet to compete with the iPad. He responded that as part of its Smarter Cities Initiative, IBM had just spent several years gathering all of the data on car transportation in Singapore; it then fed the data into a model it had built to predict the time and location of traffic jams. “We know from history what happens in Singapore if you slow the lights down in one direction by three seconds, and how to tweak the model so the jam never happens,” he told his questioner. “And so there will be a traffic jam that never occurs because we can predict what happens 20 minutes from now, because we can take enough Big Data and crunch it, and do analytics on it. So we’re predicting the future, and changing it. And you’re asking me if I’m worried about a tablet?”


Watson, too, fits into Meyerson’s conception of analytics, though it aims to change not the future of a traffic jam but of illness and investing. And by all indications, that tantalizing promise is not lost on the business community. “I have my shoulder against the door,” Saxena tells me. He means he is turning clients away–something I heard from several other sources, too–until IBM executives feel confident Watson has proved its credibility at places like Wellpoint and Sloan-Kettering. Saxena seems certain that Watson will be a multibillion-dollar business, though he will only go so far as to say that by 2015, IBM will have annual revenues of about $16 billion from its analytics portfolio, of which Watson will be a part. When I put the question of Watson’s potential to John Kelly, IBM’s chief of research, he says: “It’s like asking, at the very beginning, How big will the PC industry be?”

Kelly notes that the business model for Watson is still to be determined. He isn’t sure whether selling Watson as a computer or marketing it as a service will make the most sense. But he feels he has time to decide. None of IBM’s competitors, more than a year after the Jeopardy! victory, has announced a Q&A technology like Watson. “I think we have a huge lead,” Kelly tells me. “When people realize this is not a one-off game machine but a new era of computing, then you’ll see other companies tripling down to catch up.”

I asked a number of people, both within IBM and outside of it, whether other organizations could have built this machine first. The consensus was probably not. The reasons did not precisely connect to IBM’s technological capabilities–Google and Microsoft have plenty of computer prodigies in their ranks too. Rather, it was the combination of assets at IBM that made the difference. The company had its vast corporate lab, huge sums it was ready to invest, a profound expertise in hardware as well as software, and a collaborative culture that brought in lots of help from academia. And crucially, it had its business clients. In this respect, being a company that doesn’t cater to consumers has advantages. Watson is only as bright as its teachers. Without the staff at Sloan-Kettering, where doctors like Mark Kris teach it oncology, Watson would not be nearly so smart. In fact, it might be kinda dumb. Or it might get all sorts of things wrong, like Siri does, except you’ll be looking not for a pizza parlor but for a tumor.

From the start, the team that originally built Watson under David Ferrucci has worked out of a big room on the second floor of IBM’s Hawthorne Labs in Westchester County, New York. Hawthorne is a large glass cube of a building situated about 30 miles north of New York City. Inside the Watson work space are five fake wood-grained tables, each home to a group of computer engineers who sit around and alternately immerse themselves in their screens or break to discuss coding with a neighbor. The mood here is sober. The staffers bring water bottles, not junk food. These aren’t the unlined faces you’ll see at a startup. Indeed, Ferrucci, who sits off to the side, is a suburban dad who looks like he’d be just as comfortable standing in front of a grill with a basting brush as he is overseeing his team. The walls here are covered with huge whiteboards crammed with the hieroglyphics of computer science. Overhead lights cast the room in gloomy fluorescence. The place has the neglected feel of a finished basement in a 1970s-era subdivision.

In early fall, the Watson team, now about 45 strong, began moving its work to a gleaming new space in IBM’s main Yorktown Heights research laboratory–a promotion that reflects their importance as they support Saxena’s much larger business development group while simultaneously working on the next iteration of Watson, known as Watson 2.0. One of the team’s goals is to make Watson adaptable enough so that it doesn’t require several dozen people spending a year to get it ready for every new application, such as medicine or financial services. But a more immediate project is to help Watson through the U.S. Medical Licensing Examination, the complex test all med-school graduates must take before practicing. If it passes, says Ferrucci, “that doesn’t mean I can have a computer be a doctor.” But IBM would gain what he calls “a crisp metric” that proves Watson has a real proficiency in medicine. The credential would no doubt help Watson’s standing with health insurers, doctors, and patients, too. Passing the licensing exam is a difficult task–far harder than winning at Jeopardy!–but in early September, Ferrucci seemed pleased by the results. The computer is doing “interestingly well,” he said. He sounded confident that Dr. Watson will ace the test by year’s end.


Harder to intuit is how soon afterward Watson will infiltrate society. When I ask Jaime Carbonell, a computer science professor at Carnegie Mellon, he says he has no doubt the impact of Watson will be significant. “But I don’t think there will be one moment of, ‘Now we have it and yesterday we didn’t,'” Carbonell remarks. “It will take time to permeate. Like cell phones, which were big, clumsy things you could barely carry at first.” Was there a year, or month, or day, he asks, when cell phones began to change the world? “I can’t think of when that was,” he says. “But now we can’t do without them.”

Such is the course of technology: Electronic tools initially available only to the elite grow ever faster, smaller, cheaper. Kelly tells me he believes that eventually Watson will shrink to the size of a handheld device. Randy Katz, a computer science professor at UC Berkeley, sees a more approachable Watson, too. “Can the person in the street ask Watson a question now? No, he can’t,” says Katz. “But in five or 10 years, will there be systems like that–like Siri, but much better? I think the answer is yes.”

In many of my conversations at IBM, the talk often drifts to applications of Watson. All sorts of intriguing scenarios are presented to me–for instance, that Watson will soon analyze not just words but images, such as MRIs and EKGs. Or it will diagnose a spider bite on a child’s arm in a crop field in Africa, transmitted via smartphone by his worried father to a U.S. hospital. One afternoon, Saxena suggests this one: When you think you’re coming down with the flu, Watson will be able to discern, before you even arrive at the doctor’s office, that it might be a ragweed allergy, based on your medical record (you’ve had the same symptoms twice before at this time of year); your symptoms (gleaned from the insurance claim and diagnostic information in journals); and recent news (it just read an article in the Austin-American Statesman on a ragweed outbreak near your hometown).

It all sounds amazing. It’s also speculative. Watson has not yet saved a life or a dollar of medical costs, or added anything, really, to IBM’s bottom line. It has not yet faced its resistors–doctors who may find the technology objectionable and slow its adoption. It has not yet, as Saxena believes it will, changed the course of history. It has only won a television game show.

Still, Saxena predicts the computer will begin to scale up dramatically late next year. “By then,” he says, “we will have built the technology, demonstrated it, built the tooling and methods around it. We will have the recipe book, and then we’ll just push it out.” But he will only have reached the end of Watson’s beginning.