Standing on his magic carpet, Dishman shows me just how deeply the medical Numerati may eventually be able to peer into our lives simply by analyzing our steps as we grab a midnight snack or wash the dishes. He takes a couple of quick paces across the tiles. A video screen behind him displays his weight distribution with a trail of blue and red dots. “Now I’m putting more pressure on this foot,” he says, breaking into an exaggerated hobble. “It can tell I’m limping.” That might mean that he’s had a fall or maybe that his toenails need to be cut. (It sounds silly, but toenail care is something gerontologists keep a close eye on. Untrimmed nails can signal other problems, from immobility to depression to the onset of Alzheimer’s. And nail problems can lead to falls, a leading hazard for the elderly and a major focus for Dishman’s team. He says that falls in the United States lead to $100 billion in annual medical outlays.) He hops off the tiles and tells me to take a turn. I step on. The tiles are a little squishy, more like cardboard than normal linoleum. (I wonder if they would have absorbed the broken eggs and spilled cocoa that became so common in my parents’ kitchen.) The screen behind me displays what Dishman calls my postural sway. It looks like a blue Christmas tree. I see it tilting to the right and quickly make an adjustment. If I were one of the elders currently testing the magic carpet, the system would record that Christmas tree pattern, establishing it as my “baseline sway.” If the pattern changed, Dishman says, it could mean muscle loss or perhaps a side effect from a medicine. “You start capturing this data all the time, and you start to get some really nice trending information,” he says. “Basically, every time Mom walks, you want to compare it statistically to every other time she’s walked.” He lifts his voice a notch. “Wow, Dad’s stumbling more, especially in the morning. Why is that? Is it because of a medication he took overnight that’s not trailing off in the morning? Or is it the onset of some cognitive disease?”
The machine won’t be able to answer those diagnostic questions, at least not in the near future. It will simply issue alerts when it detects changes in patterns and perhaps urge the user to schedule a medical appointment. It will be up to doctors and nurses to follow up, figuring out why someone is limping or swaying differently at the kitchen sink. But in time, these systems will have enough feedback from thousands of users that they should be able to point people—either doctors or patients—to the most probable cause. In this way, they will work like the recommendation engines on Netflix or Amazon.com, which point people toward books or movies that are popular among customers with similar patterns. (Amazon and Netflix, of course, don’t always get it right, and neither will the analysis issuing from the magic carpet. It will only point caregivers toward statistically probable causes.)
Dishman’s team has installed magic carpets in the homes of people with neurological disorders or a history of falling. They’re starting in the kitchen but would like to extend the tiles into hallways, where they can capture more walking data. They’re also trying out two other technologies, a camera that monitors the entire body and a clip-on sensor, about an inch wide, that captures all sorts of data about movement and body tilt. Maybe, says Dishman, the magic carpet can report the data right to the user and act as a fitness coach. He says that the tiles, working with the home computer, will be able to “literally lead you though exercises.” And while this is happening, the machine—naturally—is busy collecting even more data. If the user is wearing one of those clip-ons, it might capture heart rate and customize the workout, just like a StairMaster at the gym. (I can’t imagine my own mother doing that one in her last years. But perhaps some sprier eighty-somethings would give it a go.)
Thirty-four-year-old Matthai Philipose is one of the quants in Intel’s Seattle labs. He and his team pick apart Dishman’s data. I call him up and ask how they use statistics to infer our health and behavior from footsteps on a kitchen floor. How far are we from impromptu machine-led calisthenics? He laughs. “These tools weren’t even around three or four years ago,” he says. To reach the kinds of sophisticated analyses I’m talking about, they’ll need to stitch together lots of smaller observations, each with its own range of probabilities. That’s what his team is working on now. Start, for example, with a toothbrush. In experiments, the Intel team has wired them with radio tags. These send an alert each time the toothbrush is moved. Can Philipose’s group infer that each time it’s moved, someone is brushing his teeth? Not necessarily. Someone might move the toothbrush when cleaning the sink. So the statisticians create a chart of toothbrush movement. Let’s say they see lots of activity in the morning and at bedtime. Together those two periods might represent 90 percent of toothbrush movement. From that, they can calculate a 90 percent probability that toothbrush movement involves teeth cleaning. (They could factor in time variables, but there’s more than enough complexity ahead, as we’ll see.) Next they move to the broom and the teakettle, and they ask the same questions. The goal is to build a statistical model for each of us that will infer from a series of observations what we’re most likely to be doing.
The toothbrush was easy. For the most part, it sticks to only one job. But consider the kettle. What are the chances that it’s being used for tea? Maybe a person uses it to make instant soup (which is more nutritious than tea but dangerously salty for people like my mother). How can the Intel team come up with a probability? One way, of course, would be to survey thousands of homes and ask people what they do with their kettles. That’s too much work. Philpose favors a simpler approach.
“Go to Google,” he says, “and type ‘making tea.’ How many Web pages does it locate?” (I do the search. It finds 261,000.) “Then you do another search, but add the word ‘kettle,’” he says. (This time it comes up with 29,500.) That gives scientists a rough conditional model. It assumes that of the incidents associated with “making tea,” a bit more than one of nine involves kettles. Like so many of the statistical assumptions, it starts as a crude estimate. But it’s a way to populate a monstrously big statistical table, one that comes up with the most likely behavior for thousands of scenarios. As more observations pour in from more sensors, the machine itself can adjust and refine the numbers. “Bootstrapping” is Philipose’s word for this variety of machine learning. “These kinds of models are good enough that they can start bootstrapping themselves,” he says. As they do, they’ll come up with better and better guesses about what we’re doing every minute of the day.
In these early stages, Philipose says, the Intel team is building statistical models around three groups of observations: morning and bedtime rituals, movement around the house, and nutrition. With these models in hand, they can start adapting the same statistical approaches the Numerati use to look for correlations among shoppers. In this case, do people who use the kettle or the microwave oven at lunchtime have high sodium levels in their blood? How about a person who forgets to brush his teeth more and more, and walks with slower steps through the house? Philipose’s numbers won’t tell the story by themselves, at least at this stage. But they should point doctors and nurses toward the people most likely to need help.
NOW, PICTURE an elderly couple. The husband speaks. The wife says, “What?” He repeats. She still can’t make it out. She crosses the room, turns an ear toward him, and what she finally hears is this: “Go get your ears checked.” So she does. And it turns out that her hearing isn’t the problem. It’s that her husband is speaking more softly—perhaps as a result of Parkinson’s disease. (And by the time his voice softens to this degree, the disease is advanced.) This is a key area for research because signs of Parkinson’s can show up in voice patterns and bodily movements as much as a decade before the disease is usually diagnosed. Early treatment, in exercises and medicines, can delay its development and lessen its impact. Dishman tells me that specialists studying the actor Michael J. Fox in his old TV shows can detect the onset of Parkinson’s years before Fox himself knew he had it. His gait grows shorter as the disease creeps up on him. His voice patterns change.
Very few of us have been creating weekly half-hour videos over the past 20 years that record our changing speech patterns and gestures. But with today’s technology, we’re in a position to turn the cameras, along with dozens of other sensors, on ourselves. And some of us will start doing just that. Think of all the people who pack their diets with cancer-fighting antioxidants and those who try to stave off heart disease (and risk pneumonia) by jogging relentlessly through the winter slush. Many of us are more than ready to take aggressive action when it comes to lengthening our lives. So it’s only natural that at least some of us will train a few sensors on ourselves and send the feeds to Numerati-powered consultants. Anyone offering this type of predictive service in the next few years, I should add, is likely to be a charlatan. For most diseases, the behavioral patterns are not yet established. But once analysts build up a decade or two of data, they’ll see the onset of disease early enough, hopefully, to nip it in the bud.
This power to predict is sure to raise a host of social and economic issues. Will those of us who resist using sensors be viewed as reckless, like those today who go for years without getting a medical checkup? Will governments demand a certain level of electronic reporting? Will insurance companies treat unmonitored customers as high risk, denying them coverage or saddling them with the same extortionate rates they levy today on teenage and drunk drivers? These are not issues yet, because the science is at an early stage. But Dishman’s team and others around the world are making progress every day.
For now, some of their most useful work is focused on helping people already struggling with diseases such as Parkinson’s. In these cases, Dishman says, a stream of data can help doctors refine medications. The status quo in many hospitals is to check patients once a year for 15 minutes to a half-hour and to give them prescriptions based on data gathered in that brief period. This regimen is especially ill suited to Parkinson’s, whose symptoms fluctuate wildly even over the course of a single day. “It’s a once-in-a-year shot in the dark,” Dishman says. “Think about it. You drive there, find a parking space. Your blood pressure is probably through the roof. Then you go into this very unnatural setting where they give you a series of diagnostic tests. And the nasty part of it is that if you’re having a particularly bad day, they’re going to increase your levodopa, the drug for Parkinson’s, which has a bunch of side effects.”
In clinical trials, Intel is installing five Parkinson’s tests in the homes of people with the disease. Some of these tests are familiar to patients. In one they press two piano-type keys as quickly as possible. Another has them placing tiny red and green pegs into what must be maddeningly small holes. Traditionally, a nurse with a stopwatch measures the time it takes their trembling fingers to finish the job. The Intel version does this electronically and even notes the patterns made as the user drags the peg across the surface of the box in search of the hole. Another device, which looks like a watch, measures the second-by-second shaking of the arm. In this early stage, Intel is simply gathering the data. But the next phase, says Dishman, will be to “close the loop,” giving the data to a doctor who can prescribe medicine on a day-to-day basis. He predicts that in time, computers will establish behavioral patterns and make the recommended prescriptions, first in the form of suggestions to the doctor and eventually directly to the patient.
While Dishman’s gadgets measure our behavior from outside our bodies, other researchers are busy developing sensors to report on the changing conditions inside. Teams of researchers at the Koch Cancer Institute at MIT are already testing implantable nanosensors in mice. These are built on a scale so infinitesimal that it’s hard to imagine. One building block of these sensors, a cone-shaped grouping of molecules called a carbon nanotube, is as small compared to a soccer ball as that ball is to the earth. Tyler Jacks, director of the Koch Institute, says these sensors can detect chemicals in the blood that indicate the growth of a tumor. This technology, potentially, would mean that cancer survivors would not have to wait nervously for their annual checkups to see if the cancer had metastasized (often to an untreatable stage). Instead they’d receive radio alerts immediately, perhaps straight to their cell phones. Doctors could then attack the nascent tumor. Eventually, Jacks envisions placing a host of microscopic sensors into all us, with tools to measure all kinds of conditions and alert us to trouble down the road. For these to work, the Numerati will have to develop statistical norms for hundreds of our biological patterns, from sodium and sugar levels and blood cell counts to the manufacture of all sorts of proteins. These will be our baselines, just like my tilting Christmas tree as I stood on the magic carpet. Developing the most precise models will be especially important as doctors embark on the next step: automatic treatment. “The next generation of embedded medical gizmos,” Jacks says, “will know what therapy you need and will deliver it.” Included in this micromedical cabinet, he predicts, will be an apothecary of so-called smart bombs, nanoparticles that can be dispatched for precision work, such as attacking cancer cells. It sounds promising, but as you might expect, these new sensors and medications will have to wind their way through years of development, trials, and regulatory approvals before they work their magic inside us. Some animals, though, don’t have to wait nearly so long.
THEY SAY it’s noninvasive. But I’d feel invaded,” Dan Andresen tells me, “if someone stuck a tool kit into my stomach.” He points to a big rust-colored steer named Norman, who’s wearing what looks like a white plastic Frisbee halfway down his left flank. This is actually a door, a fistula, which opens into the second of his four stomachs. It was surgically attached when Norman embarked on life as a beefy, three-quarter-ton lab rat on this research farm at Kansas State University. Fistulated cows are common in cattle country, where farmers and researchers like to keep an eye on digestive workings. But what passes through Norman’s door is most unusual.
Sometime later this spring morning, one of Andresen’s grad students will climb into Norman’s ring. He’ll unscrew the steer’s hatch and drop a black plastic packet, about the size of tennis ball, into the sloshing gallons of half-digested alfalfa. Inside that packet is a circuit board rigged with all sorts of technology. It has sensors to measure the temperature and barometric pressure inside the animal. It has a global positioning unit to track Norman’s steps in the unlikely event he bolts from this small corral and strolls down the hill through campus to the shady suburban streets of Manhattan, Kansas. Norman’s tool kit also includes a wireless transmitter with a small antenna and a memory chip big enough to log the animal’s movements and bodily functions. Much of this technology is a work in progress. But one day, when Norman wanders to the trough, the data will fly from his stomach to a wireless receiver, which will send it zipping straight into Andresen’s computer.
Dan Andresen, who teaches computer science at Kansas State, grew up on a cattle ranch in eastern Nebraska. About a decade ago, Andresen had lunch with Steve Warren, a software engineering professor who also spent his boyhood around cows. Warren had come to Kansas from the Sandia National Labs in New Mexico, where he’d been working on health monitors that people could strap onto their arm or around their chest. People, it turned out, weren’t the greatest test subjects. The devices were much bulkier back then. And humans often used their advanced brains and opposable thumbs to remove them. What Andresen and Warren needed was a duller and more pliant population. They had their idea before they even paid for their sandwiches.
Together they would build a computer network as vast as the Great Plains. It would stretch from the parched brown pastures of Kansas and the feedlots in Nebraska and Texas to slaughterhouses in Iowa and Minnesota. This network would not only track the health and movements of American cows, but it would also exist on or inside them, perhaps in packets behind their heads or encased in pellets they could swallow. Warren and Andresen planned eventually to put a wireless computer on half a million cows in Kansas—a state where the 7 million cattle outnumber people by nearly three to one. This would pr
oduce untold mountains of cow data. Heartbeats, head bobs, munch-munch, a siesta under a shade tree, a glug of water. Run that stream of data 24/7 and multiply it by a half-million, and it would create perhaps the most tedious reality show in the long history of agriculture. But the patterns in that data, analyzed mathematically, could point to all kinds of insights. Who knew what they might find? Perhaps they’d see fluctuations in cows’ temperatures before they got sick. Maybe they’d spot an epidemic working its way across the state. The key was this: instead of veterinarians checking up on cows every few months, computers would be reporting on them every single minute.
The two professors drew up a grant proposal and received funding from the National Science Foundation. Following the terrorist attacks of 2001, interest in the project grew. By tracking every animal from birth to the slaughterhouse, and even following its subsequent parts and byproducts as they were transported and sold, authorities could take a big step toward securing the nation’s food supply. In a sense, wiring the cows would be akin to equipping each animal with a recording machine, like the black boxes airplanes carry. If a diner in a restaurant anywhere in the world bit into an American steak and fell ill, the trail of information could help authorities trace the problem not just to a certain region or feedlot, but conceivably to an individual cow. They might see that on a certain day, while the cow was grazing in a certain Kansas pasture, its vital signs abruptly changed. That kind of detail could help solve the mystery.
The Numerati Page 16