A premature baby can be as small as the hand that cradles the head of a full-term infant. In a neonatal intensive-care unit, babies are often so covered with sensors that doctors and nurses struggle to find enough skin to place them on. A squadron of machines stands vigil around their tiny beds, monitoring heart rate and half a dozen other vital signs, in intervals that can be measured in thousandths of a second. All of this watchfulness is very expensive; a stay in a neonatal intensive-care ward can last months and cost hundreds of thousands of dollars.
Given that expense, and the constant danger these babies face as their underdeveloped lungs and immune systems struggle to cope with the world, the use we make of all this information is surprisingly primitive. Periodically, a nurse stops by, eyeballs what has happened since the last check, and makes a note in a chart. A doctor reviews the chart, and may scroll back through the readouts. But he or she has no easy way to view them all in one place. The machines do not talk to each other, or to anyone else; each electronic guardian stands its own lonely watch.
Technology analysts refer to the elements of systems like this as “data silos”—each data set stored by itself, never touching the others. Over the past few decades, many industries have started breaking data out of their bunkers and using powerful computers to cross-index them, revealing previously unsuspected patterns. In health care, however, data isolation is still the norm.
IBM hopes to change this. Pioneering technology now enables the company’s systems to store and analyze streaming data in real time, a task that was previously too big for ordinary computers to handle. In a small field trial at Toronto’s Hospital for Sick Children, IBM is using that technology to test a theory already suggested by some studies: that tiny changes in heart rate may indicate infections at least 12 hours before they would otherwise become apparent. At the moment, the machines are simply watching, storing data and their predictions, so that IBM can test whether its prognostication works. But if all goes well, in the summer of 2011 the machines will start relaying heart-rate changes to clinicians, who will then be able to start antibiotics earlier, before an infection rages out of control.
The new system would be a significant advance. Preemies are already vulnerable to lifelong complications ranging from vision problems to permanent brain damage. Infections can play a big role in those problems, and early detection offers a chance to stop bacteria before they can compromise organs or kill their victims. That should mean shorter intensive-care stays, smaller medical bills, and most important, a chance at a longer, healthier life.
But early treatment of infections is just the start. Researchers also hope that bringing together these streams of data will allow them to “mine” records for other potential early warnings—perhaps enabling them to detect the seizures that so often inflict brain damage on neonates. This sort of monitoring could be expanded to the many adults who also need watching, in intensive-care units and trauma centers everywhere.
Eventually, such systems might transform not just diagnosis, but the whole medical system. If we could develop more-comprehensive medical records, and collect that data in some central location, data mining might detect patterns in disease and treatment that we now discover only through painful trial and error. More than that, it could finally allow us to reach the holy grail of health-care wonks: paying for wellness rather than for doctors’ visits and procedures.
Technology’s champions have promised that sort of radical transformation before. But their plans have foundered in part because building a comprehensive system that can interact with so many different providers, from acupuncturists to pharmacists to heart surgeons, is hard. They have also often met fierce resistance from all the groups that are part of the current system: patients concerned about privacy; doctors concerned about autonomy; people at every level of the system concerned about expense. Building better computers, it turns out, is the easy part. You also have to change human behavior—and human beings are often quite happy the way they are.
When you interview experts on health-care IT, they inevitably agree on its backward state. Health care now accounts for roughly one-sixth of GDP, yet its IT infrastructure is barely in the 20th century, much less the 21st. Although most hospitals now have electronic medical-record systems, many physicians still do not, and those that do have not necessarily succeeded in integrating their systems with those of other providers—or their own workflow. Physicians will often jot down notes to be entered into the computer later, rather than altering their patient interactions so that they can talk and type at the same time. This behavior not only extends the workday, but also limits how useful the tools are. Though these systems could theoretically harness computing power to enhance diagnosis and patient discussion, many doctors use them as a poor substitute for a pen and paper.
Kavita Patel, the director of the Health Policy Program at the New America Foundation and a veteran of the Obama administration’s health-care-reform efforts, says that physicians are cautiously watching what might be done with more-detailed records; they’re used to operating as autonomous professionals, not closely monitored employees. Patel has been a practicing physician, and she points out that some of the concerns that doctors have, such as how electronic records might be used in malpractice cases, are valid. Inadvertent data loss, poor tracking of changes to the records, data-entry errors, and privacy breaches all raise potential liability issues for physicians. Also understandable is their fear that standards will be set in a way that turns their white coats into straitjackets. Over the past decade or so, insurers have become more aggressive about holding physicians to standards of care, not always with great results.
Take a woman with chronic urinary-tract infections or recurrent sinus problems. Ten years ago, physicians might have prescribed a prophylactic dose of Cipro; thanks to a shift toward rules-based medicine and tighter cost controls at insurance companies, they now tend to ask patients to come in for tests and then try a cheaper antibiotic. But since these extraordinarily painful infections often hit on a night, a weekend, or a business trip, the result can be a patient who ends up in the emergency room after hours of unnecessary agony. This outcome is more expensive, and worse for the patient: a lose-lose proposition. More-advanced data mining might let us set more-complex standards that take into account things like emergency-room visits. On the other hand, if data mining is done badly, it might simply lead to more crude rules with more unintended side effects.
Patel thinks that on balance, the results of expanded data monitoring will be good. So does David Cutler, a Harvard health-care economist who advised the Obama campaign. He believes that building a data mine could greatly improve how we treat patients. At the moment, randomized controlled trials are the gold standard of medical research. You divide patients into two groups, give one your new treatment, and see how those patients fare compared with a control group that is getting an old treatment, or no treatment at all. These studies have great rigor, but also great limitations: they cover a small number of patients, for a limited time, and they test only what the researchers are looking for. So even though drugs are subjected to the Food and Drug Administration’s scrutiny, horrible side effects may only become obvious later, like the heart problems that forced Merck to pull Vioxx from the market. If a complication appears in only one out of every 10,000 patients or takes years to develop, even a comprehensive trial can easily miss it.
Better data and better analytic tools could allow us to supplement such research with what scientists call “observational data”: say, watching all the patients who take a new drug to see what unexpected conditions pop up. Cutler thinks that mining could detect the subtle relationships that controlled trials often miss. “I’m pretty sure that if your mother has a stroke, the optimal treatment regime very much depends on whether or not your dad is still alive,” he points out. “That affects the kinds of things she can do, the kinds of help she can get. But right now we don’t have any way to look at those kinds of factors and design around them.”
The Obama administration is trying to rectify that. The stimulus and health-care-reform bills contained billions to set standards and help physicians acquire electronic medical-record systems; created a new center that will experiment with payment reforms in Medicare and Medicaid; and established a 15-member advisory board that has a mandate to tie Medicare and Medicaid payments more closely to research on comparative effectiveness. If done right, these steps should push us toward a future of better, more data-driven medical decisions. But they won’t be pushing us very fast. Patel says we’re “very far away” from seemingly minor innovations like letting your doctor know whether you actually filled the prescription she wrote. When it comes to health care, almost everyone seems to be locked inside a silo.
Chalapathy Neti, IBM’s lead researcher on health-care analytics, hopes that the sector’s historical resistance to change can be overcome with technology that combines better monitoring with better diagnostic tools. Simple electronic medical-record systems don’t necessarily offer physicians much; they are often more of a boon to billing departments than to doctors. But as part of IBM’s $100 million global health-care-technology initiative, Neti is overseeing multiple projects that are developing systems for data integration to help physicians improve their treatments. One of them is an HIV database that compares viral-DNA profiles across patients to predict which drug cocktails are likely to work best. Such advances seem to be moving analytics in the direction that Cutler hopes: more-customized medicine.
But Neti also thinks that more-advanced data-management systems, like the ones IBM hopes to sell, will be widely adopted if payment and incentive systems are better aligned with outcomes. It’s easy to justify the expense of such systems in a neonatal intensive-care unit, where the stays are so expensive and the benefits are so obvious. It’s harder for physicians working in small practices to justify even a basic medical-records package, which can cost $100,000 for limited benefits.
If we paid for outcomes rather than for treatments, Neti argues, we’d give providers a financial incentive to have the best possible information to treat their patients. But there’s something of a chicken-and-egg problem: to reward wellness, we first need more data. Patients are not widgets; they are complex and unique. We can’t simply pay more to the doctors with the healthiest patients, because that would just discourage doctors from treating very sick people. We need a way to sift through the data and control for all the complex factors that determine health, before we can start paying doctors based on how well they manage their patients’ care.
This is easy to say, but hard to do. Jim Manzi is the chairman of Applied Predictive Technologies, which helps companies in other industries analyze the kind of data-mining operations that health-care analysts are now eyeing. Manzi says that data mining is relatively simple when you’re, say, trying to analyze the risk of defects in the manufacture of a machine part. But it rapidly becomes complicated when you are trying to engineer a change in human behavior, especially when money is involved.
“When people’s compensation is on the line,” says Manzi, “they suddenly turn into Aristotle.” Many a company has tried to design its sales-force commissions around arcane data analyses that try to control for complex factors, like whether one salesperson’s territory has more customers than another’s. Suddenly, they find the salespeople are all crack statisticians who can explain where the model has gone wrong in the case of their territory. Many of the companies that press forward find themselves, after six or eight months, with a sales force in full revolt. “I’ve seen it many times. Very few data-mining systems survive first contact with reality.”
Based on his experience with other industries, Manzi has a few thoughts on where data mining is likely to succeed in health care—and where it will be an uphill battle. Chemical or biological reactions, he says, will probably be easier to target than behavioral changes. Acute conditions treated in a highly standardized hospital environment will be easier to target than chronic ones treated in the outside world. Patients with singular problems will be easier to target than patients with three—or eight—chronic conditions. Costs will be easier to target than effectiveness.
This suggests that efforts like IBM’s Canadian project are likely to succeed and spread: they offer doctors an obvious way to improve what they are already doing. But making fundamental changes in medical practice will be much more difficult. Fighting disease is relatively simple. Fighting patients, doctors, and all the other stakeholders in the current system may be beyond the powers of even the most advanced computing system.