Watson the supercomputer was last seen on "Jeopardy!" in 2011, crushing two very human former champions and taking home the $1 million prize. Now IBM's whiz kid (whose winnings went to charity) has moved beyond trivia to an even tougher question: Can a really smart computer become the ultimate physician's assistant and vastly improve the quality of health care?
Mark Kris, chief of thoracic oncology at Memorial Sloan-Kettering Cancer Center, thinks so, and he demonstrates how using the case of a virtual patient, Lin Yamato, a 37-year-old woman who never smoked but who recently received a diagnosis of advanced lung cancer. Even before Yamato visits her cancer specialist, Watson goes to work, culling details of her test results and medical history from her electronic medical record and analyzing them in the context of the 2 million pages of knowledge it's already been fed: journal articles, clinical trial results, treatment guidelines, textbooks and the histories of hundreds of de-identified Sloan-Kettering patients.
Based on the evidence, Watson suggests more tests – a molecular analysis of the tumor and an MRI of the brain – and brings up the relevant references. And the computer offers a preliminary assessment of three treatment plans. At this stage, there's no clear leader; the computer's confidence in each plan hovers around the 30 percent mark. It says it needs more information.
That judgment is the crux of Watson's role. The computer "doesn't tell you what to do. It tells you that a treatment is at the top of the list" for a particular patient, says Kris. What makes Watson unique, besides its power (on "Jeopardy!" it sifted through more than 66 million pages of data per second and has only gotten faster since), is its ability to absorb text that isn't in a standardized format, such as a doctor's case notes, and pull out the most relevant ideas. And it learns from experience, says Marty Kohn, IBM's chief medical scientist; the more cases Watson sees, and the more interaction it has with users who point out additional sources Watson should consider, for instance, the better it gets. "It's like teaching someone to be a doctor," says Kris.
Watson's foray into oncology is only the first baby step toward applying "big data" to thorny medical problems. By one estimate, health information – electronic health records, insurance claims, images such as CT scans, vital signs of people being remotely monitored by hospitals or smartphone, gene sequencing results – will grow to the equivalent of about 500 billion four-drawer file cabinets by 2020, from a mere 10 billion in 2011. High-powered computers and new algorithms have the potential to allow physicians and researchers to combine and decipher all that information and see what connections pop. Ideally, medicine would then be able to better track and predict the spread of disease, and diagnose, treat and prevent it, all while improving safety and lowering costs.
"You can look back and see what happened historically. You can do a real-time analysis of that data, and you can look forward: How can I use this to predict what will happen?" says C. Martin Harris, chief information officer of the Cleveland Clinic, which is working with Watson to improve medical education. At the moment, Watson is "learning" the same kind of information as trainee physicians, with the goal of taking the same licensure exam they take. Once up to snuff, the computer will turn the tables and help the students.
There's clearly great potential in crunching all this data, but also many obstacles. For starters, this kind of computer power can require big upfront investments. The Icahn School of Medicine at Mount Sinai in New York spent $3 million on Minerva, its supercomputer to be used for genomic and other research, for example.
Another challenge: Before meaningful interpretation of the information can begin, all sorts of disparate data sources must be brought together, no small feat in an industry where patient privacy and security are paramount and the culture is oriented toward not sharing data. Efforts to overcome that tradition are underway. The American Society of Clinical Oncology, for example, is developing a database intended to pool clinical information on hundreds of thousands of cancer patients for use by oncologists. And several electronic health record vendors recently announced an agreement to ease the flow of data between their systems.
Back at Sloan-Kettering, the Watson initiative aims to help physicians arrive at the most effective possible cancer care. Taking into account Yamato's preferences (she has young kids and so prefers a drug regimen that is less time-intensive, and also would like to avoid losing her hair) and the results of the tests (a clear MRI and the identification of the mutation driving her tumor), Watson's confidence in the options changes. After the revelation that she's been coughing up blood, it changes again. Now Watson can identify the best option, a regimen of three particular drugs. It can also suggest clinical trials for which Yamato may be eligible.
Let's be clear: Watson will not be present at your next physical. Training the computer on each disease currently takes many months, though the process is getting faster as the computer learns more. And teaching Watson to make sense of doctor lingo is an ongoing challenge. Time stamps on faxes can throw off the computer, and words can have multiple meanings. "T2" can refer to a stage of cancer, the second thoracic vertebra or one component of an MRI scan, says Kris. It's early days yet for Watson.
But all the "ifs" aside, the potential to spread the specialized knowledge of a major cancer center far and wide is "tremendous" once more advanced commercial versions are widely available, says Kris. The first iteration of an oncology tool, cloud-based and accessible via computer or tablet, is being tried out at two medical practices in Maine and New York; it won't yet have the full capabilities being tested on Yamato. Insurer WellPoint will ultimately market the product.
At some point, too, Watson will assist doctors with diagnosis. Imagine a primary care doctor seeing a patient with a befuddling constellation of symptoms: double vision, dry mouth, slurred speech. A supercomputer could help winnow the list of possible diagnoses and remind her of some she might otherwise overlook (like botulism).
Beyond diagnosis and treatment, plenty of other challenges might be addressed by new kinds of number-crunching. Take disease surveillance. If a kid comes in complaining of a sore throat, "what I'd really like to know is if strep throat is going around," says Kenneth Mandl, a researcher and physician in the Boston Children's Hospital Informatics Program.
In a study published in 2011, Mandl and his colleagues showed that real-time regional patient data from outlets of the retail chain MinuteClinic gave a good indication whether the next person to walk in was more likely to have strep or just a virus. That could cut down on unnecessary antibiotic prescriptions and ensure that people who do have strep get tested and treated.
And there are entirely new data streams to wrangle, thanks to people's constant connection to the Internet, social media and their beloved mobile devices. A platform created by the Cambridge, Mass.-based startup Ginger.io allows researchers and patients to use smartphone data to passively track (and ideally gain insights from) the movements and behaviors of patients with conditions such as diabetes and inflammatory bowel disease. If someone is repeatedly unlocking his phone at 3 a.m., say, it means he's not sleeping and may not be feeling well. The same may be true if he usually moves around a lot during the day but suddenly sticks close to home.
It becomes "a lot easier to actually track symptoms if we can rely on the digital exhaust from your cellphone," says Michael Seid, a professor at Cincinnati Children's Hospital Medical Center who is using the Ginger.io app to improve the care of young IBD patients. By correlating their activity levels and call patterns with reported symptoms, doctors might be able to send an alert when a medication change or a visit seems to be in order. They might even be able to predict an IBD flare-up days in advance and prevent it.
Using past patterns of data as a crystal ball has already showed promise in other settings. A study published in 2009 found that a computer model could rely on previous entries in a patient's medical record, such as substance abuse or poisoning incidents, to predict the risk of a future diagnosis of domestic abuse. IBM and Excel Medical Electronics are working with the UCLA Department of Neurosurgery to study whether real-time analysis of subtle changes in a host of different vital signs of traumatic brain injury patients can predict (and preclude) a dangerous increase in brain pressure.
Big data is also being used to tackle behavior change. On behalf of clients such as big insurance plans and employers, Eliza Corp. uses automated phone, email and text messages to connect with people and encourage them to take their prescribed drugs, get screened for cancer, or engage in some other health-promoting behavior. By drawing on its more than 900 million previous interactions, Eliza can tell what works, and what doesn't, says Alexandra Drane, founder and chief visionary officer. Reminding women about breast cancer screening by saying that their mammogram machine "missed" them, for example, is far more effective at bringing them in than a lecture on early detection.
Now that people are eagerly sharingtheir health experiences online, might that information be advantageously mined, too? Patients have such unparalleled access to medical information and to each other that their collective experience can be as valuable as, or even more valuable than, that of a single practitioner, argues Herbert Chase, professor of clinical medicine in biomedical informatics at Columbia University's College of Physicians and Surgeons and a member of Watson's advisory board.
Chase offers a personal example: He started taking Lipitor and experienced insomnia, but found no mention of that side effect in the literature. Patients on blogs were all over it. If a supercomputer could somehow seek out and collect that information, experts say, it would offer yet another perspective.
Dave deBronkart, for one, would welcome it. He blogs about participatory medicine as e-Patient Dave, and when he faced a late-stage kidney cancer diagnosis in 2007, read that median survival was 24 weeks. For patients like him, the lag between finishing a study and seeing it published can be deadly, he says, so what's needed is to unleash Watson and its ilk to harvest this early, more tentative knowledge. He envisions a tool that reveals a potentially life-saving tidbit shared by a researcher "last Tuesday in Budapest, on slide 27 of his presentation."
Elementary? Maybe not. But surely possible.