Stories
Stories
Big Blue’s Big Bet
Topics: Innovation-Technological InnovationHealth-Health Care and TreatmentTechnology-Search TechnologyInformation-Information Management
Big Blue’s Big Bet
Topics: Innovation-Technological InnovationHealth-Health Care and TreatmentTechnology-Search TechnologyInformation-Information Management
Big Blue’s Big Bet
One day not long ago, a Japanese woman in her 60s walked into a hospital in Tokyo, worried she might have cancer. Doctors ran tests and ultimately diagnosed her with acute myeloid leukemia, a blood cancer in which abnormal white blood cells grow quickly. They comforted her by saying that the chemotherapy they were prescribing would attack the abnormal cells. And it did.
But her recovery from post-remission therapy was slow—too slow in fact. Doctors began to wonder if she had a different type of leukemia. They ran more tests but saw no sign of one.
The hospital was affiliated with the University of Tokyo’s Institute of Medical Science, which had partnered with IBM Watson, a cloud-based cognitive-computing system that attempted to complement the expertise of the doctors treating the woman. Watson had already ingested millions of oncology papers and volumes of leukemia data from research institutes around the world. Now, the doctors in Tokyo fed Watson the woman’s genetic information, hoping it would find a gene mutation that had led to her cancer. Watson took the woman’s data and cross-checked it against the genetic databases of the research institutions.
After 10 minutes, Watson noticed that the woman had not one but roughly 1,000 gene mutations. Many of these, however, seemed to carry hereditary marks unrelated to her cancer. It would have taken doctors at least two weeks to sift through these mutations, searching for the ones that were diagnostically significant. But Watson was able to highlight within moments which mutated genes had likely developed into cancer.
Based on Watson’s analysis, the treating oncologists believed the woman had a rare secondary leukemia caused by myelodysplastic syndrome, a group of diseases in which the bone marrow makes few healthy blood cells. The doctors changed the woman’s therapy plan, and her health improved considerably. She was discharged from the hospital weeks later. “We would have arrived at the same conclusion by manually going through the data,” the lead doctor told the Japan Times, “but Watson’s speed is crucial in the treatment of leukemia, which progresses rapidly and can cause complications.”
This is the beauty of Watson, which began as a television stunt and has now become a major driver of innovation in fields as diverse as real estate law and songwriting. But it is in the health sector where Watson’s impact is most tangible, where its benefits are literally the difference between life and death. “What I focus on,” says Watson VP and CFO Deana Korby (MBA 2003), “is making sure that we’re smart in the things we’re attacking, and to try to solve real-world problems that make an impact.” Watson General Manager David Kenny (MBA 1986) adds: “The tech sector tends to reward short-term thinking.” But Watson is different, he says: A product that was “decades” in the making, and one that will reimagine health and other fields for decades more.
Revenues at IBM have been in the red since 2012, yet analysts have pointed to Watson as the engine that can revitalize the company—as the potential through-line that will touch nothing less than everything we do in the next 30 years. So, yes, Watson is a big deal.
Perhaps the best place to start to understand it all is in 2007, in the Semantic Analysis and Integration Department at IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York. There, a team of roughly 25 IBMers were researching “natural-language processing,” which is the ability of computer systems to learn from a language as it is spoken, in all its idioms, colloquialisms, and technical-ese, and without having to be explicitly programmed to learn these things. The IBM computers answered questions with 35 percent accuracy, and among the institutions that researched natural-language processing, “35 percent was nothing to be ashamed of,” Dave Ferrucci, the leader of the IBM team, later said.
Still, IBM’s advancements were small. The company had a sense of Watson’s potential to dramatically change how business is done, but had somehow fallen into the business of publishing technical papers. “I didn’t want to spend the next 5 to 10 years doing these little increments and never knowing whether, with the right effort, we could solve a big challenge,” Ferrucci told HBS Professor of Management Practice Willy Shih in a subsequent case study. So his IBM team pitched their superiors on the company building a computer that could compete on Jeopardy!. It was bold, even risky; Jeopardy! was the ur–game show of idioms and double meanings. Building software for natural-language processing meant coding a small set of question and answer types, and then having the computer choose the right answer. But in Jeopardy! there were roughly 2,500 types of questions, and within each type, thousands of different words or phrases that might hold the clue to the answer.
“Not knowing where to start, we started everywhere,” Ferrucci said. He and his team used high-end computer power to run many different probabilistic software approaches simultaneously. They built the system so that every component produced both a proposed answer and a confidence in that answer. They built error-analysis tools, too, that helped the computer system learn from past mistakes, and all the while they fed it books, dictionaries, the Bible, encyclopedias, everything. They tested it, endlessly, reinforcing the methods that had led to correct answers and amending the mistakes, which meant tweaking the algorithms informing the computer system. Soon, IBM was flying in former Jeopardy! contestants to compete against Watson, its “brain” the compendium of 100 algorithms working in parallel against 200 million pages of text in 500 gigabytes of data.
When it finally played master Jeopardy! winner Ken Jennings in 2011, Watson beat the all-time champion by over $50,000. Jennings was reduced to ending the show by acknowledging “our new computer overlords.”
At IBM, the mood was euphoric, but Ferrucci had always had an ulterior motive for developing Watson. “I actually was pre-med in college and aspiring to be a doctor,” he recalls in a recent interview. “But I hated that I had to memorize everything, and I thought, ‘Could I get a computer to do all this stuff so I don’t have to?’ ” That notion had furthered his interest in artificial intelligence, which had led to his career at IBM, and now, with Watson’s success, he saw an opportunity to merge his passions. If the natural-language processing underlying Watson could master Jeopardy!, what else could it do?
With Jeopardy!, Watson had mastered trivia, become a better Wikipedia. It had done this by gaining what was in essence multiple advanced degrees in the humanities, in the form of all the books—really, all the libraries—it had scanned and stored.
He wasn’t the only one to ask that question. At the time, Sean Hogan (MBA 1993), a VP in IBM’s health care unit, was tasked with finding the major trends in the health and life sciences sectors that could lead to billion-dollar businesses. “We saw an opportunity in Watson,” says Hogan, who came to the company in 2001 after his IT and consultancy startup, Mainspring, was acquired by IBM. Regardless of specialty, doctors complained of rushing through their days, too over- scheduled to fully comprehend patients’ lengthy and complex medical records, each often hundreds of pages long. And physicians certainly didn’t have time to read all the articles in medical journals that might help them make better diagnoses and prognoses, not with the proliferation of academic journals reportedly publishing articles at a rate of once every 26 seconds. Doctors themselves began to email and phone IBM after Jeopardy!, asking if Watson might pick up some of the burden of their professional lives. “We even went on a tour of hosting various dinners with prominent physicians,” Hogan says, all with a single aim: “We were looking for dramatic improvements in productivity.” There were so many issues to address in medicine, that IBM knew it had to rethink its omnibus IBM Watson Unit. “So we formed a unit called Watson Health,” Hogan says.
With Jeopardy!, Watson had mastered trivia, become a better Wikipedia. It had done this by gaining what was in essence multiple advanced degrees in the humanities, in the form of all the books—really, all the libraries—it had scanned and stored. For it to augment doctors’ skills, Hogan and others knew that Watson would need to go to med school. So IBM partnered with the Cleveland Clinic and New York’s Memorial Sloan Kettering Cancer Center, and set about teaching Watson how to be a doctor. Watson scanned and stored medical journals and textbooks, but it also began to problem-solve. At the Cleveland Clinic, for instance, which is affiliated with Case Western Reserve University, Watson worked alongside students who received clinical cases from their professors. Watson and the students began to attack these cases by working together. The goal was threefold, says Dr. James Young, executive dean of the Cleveland Clinic Lerner College of Medicine: Watson was to help doctors get better at diagnosing illnesses and diseases; Watson was to utilize knowledge and wisdom more efficiently; and Watson was to translate all that into outcomes that reduced morbidity.
The reality is that subtle biases or a lack of full information lead doctors to the wrong decision, which in turn leads to bad health outcomes, including death. To highlight the fallibility of the human mind, Young’s Cleveland Clinic colleague Dr. Neil Mehta, assistant dean of education technology, likes to tell a story. “We had physicians review medical charts in our electronic medical records,” he says. These were veteran doctors, with decades of experience, asked to review charts and come up with a problem list: All the issues that might be facing the patient. The doctors had up to one hour to pore over these records. Watson then scanned the same charts, 27 in all, representing 27 patients, but was only given five minutes with each chart. The doctors and Watson then reviewed each other’s work. In almost every case, Watson found a problem that the doctor had overlooked. “We have so much data now for every patient,” Mehta says—up to 1,000 pages of records, for someone who’d been coming to the Cleveland Clinic for just two years—“and physicians have maybe 20 to 40 minutes to review the records of one patient.” Watson could consume a vast quantity of files—reading 200 million pages of text in three seconds—and augment a doctor’s expertise, informed by all it reads and never forgets.
Over the years, that human fallibility has led not only to misdiagnoses but to misspent dollars as well, which in turn has contributed to rising health care costs. According to a study published by the RAND Corporation, health care spending from 1999 to 2009 wiped out the economic gains of the typical American middle-class family. “If we solve our health care spending, practically all our fiscal problems go away,” Victor Fuchs, the so-called dean of health economics, was quoted as saying in the study.
Sean Hogan read that RAND report. What struck him as he studied the business of health care further, as he and other IBMers talked with physicians like those at the Cleveland Clinic, is that about a third of the dollars spent were wasted, “because care was redundant, unnecessary, or potentially harmful,” he says. “And we knew Watson could help.”
After years of testing, refining, and more testing, Watson was about to graduate from med school.
Around this same time, IBM began to throw big money behind Watson. In 2014, the company announced a $1 billion investment in the Watson Unit, which would advance its cognitive computing technologies in not only medicine but law and finance and government services and culinary services and on and on and on. When a computer system has the processing power of Watson, and can understand language as it’s spoken, as a person can, “There is no shortage of problems Watson can solve,” says CFO Deana Korby. “If anything, we just have to pick the right problems.” Adds GM David Kenny: “We think there’s a couple of trillion dollars of inefficiencies in the marketplace.” And IBM is attempting to leverage its expertise to get its share of that, he says.
Doing so, analysts say, will be the foundation for IBM deep into the 21st century. “Transformation and transition is something IBM does very well,” says HBS’s Willy Shih of the 105-year-old company, citing its start in butcher scales and time clocks and movement through virtually every aspect of technological development of the 20th century. While IBM doesn’t break out Watson’s revenue, the unit falls within IBM’s “strategic imperatives”: the company’s analytics, cloud-computing, mobile, security, and social businesses. Even though IBM has seen revenues decline in 18-straight quarters, the strategic imperatives unit is growing, with revenues up 12 percent from 2015. Cloud-computing in particular is rising, up 2.4 percent from the second quarter of 2016 to the third, to $8.75 billion during the quarter. Overall, the unit now accounts for 38 percent of the company’s business, according to its financial statements, with IBM spending more than $5 billion in acquisitions in the first half of 2016 to nourish cloud businesses like Watson. CEO Ginni Rometty says hundreds of millions of users interact with Watson today, across 45 countries and 20 industries. For Kenny, the genius of Watson is not that it tries to be one service for every sector, but that it specializes, just as the experts in various fields specialize. So today there is Watson for insurance underwriters, Watson for lawyers, Watson for chefs, Watson for investment bankers, and Watson for Top 40 songwriters. “Our ability to augment expert knowledge”—to even cater to it—“is unique,” Kenny says.
This specialization appears most notably in health care, a sector where one-third of dollars spent are wasted and where spending itself has outpaced the nation’s GDP in all but two years since 1950. Four years after Watson went to med school, there is Watson for radiology, Watson for oncology, Watson for diabetes, and Watson for primary care. Watson serves “thousands” of clients in the medical sector, Hogan says, and that number will only increase as hospitals learn all it can do.
Once they learn, they tend to incorporate Watson quickly. For instance, at the University of North Carolina’s teaching hospital, Watson moved from a novelty assisting doctors to the standard of care in six months, says Dr. Norman Sharpless, director of the UNC Lineberger Comprehensive Cancer Center. Besides reading all the medical journals, Watson could inform doctors of when certain drug trials began and when others ended; it could also sequence a genome and find important gene mutations faster and more accurately than physicians. “Today, there’s just too much information for us to process, and Watson doesn’t miss anything,” says Sharpless. “It became almost an ethical question: ‘Why shouldn’t we do this?’ ”
At UNC, and medical centers throughout the world, the interface of Watson is fairly simple. Watson lives in IBM’s cloud-computing system but appears as an icon on a physician’s computer screen. If a doctor asks Watson to assist in, say, a diagnosis of a patient, the doctor types in all the symptoms, and Watson weighs those symptoms against the medical literature and perhaps even the patient’s genes. Watson then lists a series of diagnoses, ranked in accordance with its confidence: The more likely the diagnosis, the higher it appears. The diagnosing itself is left to the doctor, and if he or she wants to understand why Watson ranked one diagnosis higher than others, he can click on a hyperlink that moves to a corpus of information—the medical journals, the gene readouts and medical imagery—on which Watson based its probabilistic recommendation. “When you have an array of images and the additional information that can swing your confidence in a certain scenario,” Hogan says, “it just makes for better decision-making.”
Subtle biases or a lack of full information lead doctors to the wrong decision, which in turn leads to bad health outcomes, including death.
And what doctors see, or rather what they will experience, will change as the technology evolves. There’s an anecdote IBMers share at Watson’s gleaming new headquarters on Astor Place in the East Village. One day, a boy named Kevin walked into his pediatrician’s office in New York complaining of a sore bump on his neck. He was crying, rubbing his red eyes constantly, and had a fever of 102. Kevin’s pediatrician thought she knew what might be making him sick, but she couldn’t be sure, so she sent him to the emergency room. There, unfortunately, every test came back negative, and Tylenol didn’t reduce the swelling as the doctors had hoped. As one day in the hospital became two and then three and four, doctors tried antibiotics, but to no avail. On the sixth day of his stay, Kevin was no better than when he was admitted.
Then Kevin’s nurse noticed something: Not only was Kevin’s lymph node still swollen, his eyes were still red. Everyone had assumed his eyes were red because he had been crying. But the nurse wondered if Kevin might have Kawasaki disease, a rare childhood illness that can harm the coronary arteries, and whose symptoms include a high fever, swollen lymph nodes, and red eyes. He was treated accordingly, his swelling decreased, and soon after Kevin was discharged.
For a couple of years now, IBMers have told this story as a way to highlight Watson’s strengths, that if Watson were utilized it would have shown the doctors the probability of Kawasaki disease as the culprit much earlier. But soon that story will be out of date or rather, won’t capture everything. One day, Watson will do more than just suggest likely illnesses, says David Kenny. It’ll get more granular and actually prompt doctors: Did you look for red eyes? Watson will ask. This is the breakthrough—a computer system smart enough to ask humans questions, instead of just providing probabilistic answers. “The technology is absolutely moving in that direction,” says Kenny. “Every day. We’re extrapolating from what we know into what we don’t know.”
In other words, the machine is starting to think like the wisest among us, humbly aware of all it does not yet comprehend—but one day might.
Paul Kix is a deputy editor at ESPN The Magazine. HarperCollins will publish his first book, about a French resistance fighter, later this year.
Post a Comment
Related Stories
-
- 01 Mar 2024
- HBS Alumni Bulletin
The Internet's Next Frontier
Re: Liang Wu (MBA 2018); Scott Duke Kominers (Sarofim-Rock Professor of Business Administration); Shai Benjamin Bernstein (MBA Class of 1960 Professor of Business Administration); By: Jennifer Gillespie -
- 01 Mar 2024
- HBS Alumni Bulletin
Leading Businesses in a Web3 World
-
- 01 Mar 2024
- HBS Alumni Bulletin
What is Web3?
-
- 22 Feb 2024
- HBS Clubs
GCC Crossroads Aims for a Bright Future; Seattle Alumni Talk Leadership in Tech; Italy Preps for European Alumni Summit
By: Margie Kelley