Rise of the Robots: Technology and the Threat of a Jobless Future (23 page)

Chapter 6

THE HEALTH CARE CHALLENGE

In May 2012, a fifty-five-year-old man checked into a clinic at the University of Marburg in Germany. The patient suffered from fever, an inflamed esophagus, low thyroid hormone levels, and failing vision. He had visited a series of doctors, all of whom were baffled by his condition. By the time he arrived at the Marburg clinic, he was nearly blind and was on the verge of heart failure. Months earlier, and a continent away, a very similar medical mystery had culminated with a fifty-nine-year-old woman receiving a heart transplant at the University of Colorado Medical Center in Denver.

The answer to both mysteries turned out to be the same: cobalt poisoning.
¹
Both patients had previously received artificial hips made from metal. The metal implants had abraded over time, releasing cobalt particles and exposing the patients to chronic toxicity. In a remarkable coincidence, papers describing the two cases were published independently in two leading medical journals on nearly the same day in February 2014. The report published by the German doctors came with a fascinating twist: whereas the American team had resorted to surgery, the German team had managed to solve the
mystery not because of their training but because one of the doctors had seen a February 2011 episode of the television show
House.
In the episode, the show’s protagonist, Dr. Gregory House, is faced with the same problem and makes an ingenious diagnosis: cobalt poisoning resulting from a metal prosthetic hip replacement.

The fact that two teams of doctors can struggle to make the same diagnosis—and that they can do so even when the answer to the mystery has been broadcast to millions of prime-time television viewers—is a testament to the extent to which medical knowledge and diagnostic skill are compartmentalized in the brains of individual physicians, even in an age when the Internet has enabled an unprecedented degree of collaboration and access to information. As a result, the fundamental process that doctors use to diagnose and treat illnesses has remained, in important ways, relatively unchanged. Upending that traditional approach to problem solving, and unleashing all the information trapped in individual minds or published in obscure medical journals, likely represents one of the most important potential benefits of artificial intelligence and big data as applied to medicine.

In general, the advances in information technology that are disrupting other areas of the economy have so far made relatively few inroads into the health care sector. Especially hard to find is any evidence that technology is resulting in meaningful improvements in overall efficiency. In 1960, health care represented less than 6 percent of the US economy.
²
By 2013 it had nearly tripled, having grown to nearly 18 percent, and per capita health care spending in the United States had soared to a level roughly double that of most other industrialized countries. One of the greatest risks going forward is that technology will continue to impact asymmetrically, driving down wages or creating unemployment across most of the economy, even as the cost of health care continues to climb. The danger, in a sense, is not too many health care robots but too few. If technology fails to rise to the health care challenge, the result is likely to be a soaring,
and ultimately unsustainable, burden on both individual households and the economy as a whole.

Artificial Intelligence in Medicine

The total amount of information that could potentially be useful to a physician attempting to diagnose a particular patient’s condition or design an optimal treatment strategy is staggering. Physicians are faced with a continuous torrent of new discoveries, innovative treatments, and clinical study evaluations published in medical and scientific journals throughout the world. For example, MEDLINE, an online database maintained by the US National Library of Medicine, indexes over 5,600 separate journals—each of which might publish anywhere from dozens to hundreds of distinct research papers every year. In addition, there are millions of medical records, patient histories, and case studies that might offer important insights. According to one estimate, the total volume of all this data doubles roughly every five years.
³
It would be impossible for any human being to assimilate more than a tiny fraction of the relevant information even within highly specific areas of medical practice.

As we saw in
Chapter 4
, medicine is one of the primary areas where IBM foresees its Watson technology having a transformative impact. IBM’s system is capable of churning through vast troves of information in disparate formats and then almost instantly constructing inferences that might elude even the most attentive human researcher. It’s easy to imagine a near-term future where such a diagnostic tool is considered indispensable, at least for physicians confronting especially challenging cases.

The MD Anderson Cancer Center at the University of Texas handles over 100,000 patients at its Houston hospital each year and is generally regarded as the best cancer treatment facility in the United States. In 2011, IBM’s Watson team began working with MD Anderson’s doctors to build a customized version of the system geared
toward assisting oncologists working with leukemia cases. The goal is to create an interactive adviser capable of recommending the best evidence-based treatment options, matching patients with clinical drug trials, and highlighting possible dangers or side effects that might threaten specific patients. Initial progress on the project proved to be somewhat slower than the team expected, largely because of the challenges associated with designing algorithms capable of taking on the complexities of cancer diagnosis and treatment. Cancer, it turns out, is tougher than
Jeopardy!
Nonetheless, by January 2014, the
Wall Street Journal
reported that the Watson-based leukemia system at MD Anderson was “back on track” toward becoming operational.
⁴
Researchers hope to expand the system to handle other kinds of cancer within roughly two years. It’s very likely that the lessons IBM takes away from this pilot program will enable the company to streamline future implementations of the Watson technology.

Once the system is operating smoothly, the MD Anderson staff plans to make it available via the Internet so that it can become a powerful resource for doctors everywhere. According to Dr. Courtney DiNardo, a leukemia expert, the Watson technology has the “potential to democratize cancer care” by allowing any physician to “access the latest scientific knowledge and MD Anderson’s expertise.” “For physicians who aren’t leukemia experts,” she added, the system “can function as an expert second opinion, allowing them to access the same knowledge and information” relied on by the nation’s top cancer treatment center. DiNardo also believes that, beyond offering advice for specific patients, the system “will provide an unparalleled research platform that can be used to generate questions, explore hypotheses and provide answers to critical research questions.”
⁵

Watson is currently the most ambitious and prominent application of artificial intelligence to medicine, but there are other important success stories as well. In 2009, researchers at the Mayo Clinic in Rochester, Minnesota, built an artificial neural network designed to
diagnose cases of endocarditis—an inflammation of the inner layer of the heart. Endocarditis normally requires that a probe be inserted into the patient’s esophagus in order to determine whether or not the inflammation is caused by a potentially deadly infection—a procedure that is uncomfortable, expensive, and itself carries risks for the patient. The Mayo doctors instead trained a neural network to make the diagnosis based on routine tests and observable symptoms alone, without the need for the invasive technique. A study involving 189 patients found that the system was accurate more than 99 percent of the time and successfully saved over half of the patients from having to needlessly undergo the invasive diagnostic procedure.
⁶

One of the most important benefits of artificial intelligence in medicine is likely to be the avoidance of potentially fatal errors in both diagnosis and treatment. In November 1994, Betsy Lehman, a thirty-nine-year-old mother of two and a widely read columnist who wrote about health-related issues for the
Boston Globe,
was scheduled to begin her third round of chemotherapy as she continued her battle against breast cancer. Lehman was admitted to the Dana-Farber Cancer Institute in Boston, which, like MD Anderson, is regarded as one of the country’s preeminent cancer centers. The treatment plan called for Lehman to be given a powerful dose of cyclophosphamide—a highly toxic drug intended to wipe out her cancer cells. The research fellow who wrote the medication order made a simple numerical error, which meant that the total dosage Lehman received was about four times what the treatment plan actually called for. Lehman died from the overdose on December 3, 1994.
⁷

Lehman was just one of as many as 98,000 patients who die in the United States each year as a direct result of preventable medical errors.
⁸
A 2006 report by the US Institute of Medicine estimated that at least 1.5 million Americans are harmed by medication errors alone, and that such mistakes result in more than $3.5 billion in additional annual treatment costs.
⁹
An AI system with access to detailed patient
histories, as well as information about medications, including their associated toxicity and side effects, would potentially be able to prevent errors even in very complex situations involving the interaction of multiple drugs. Such a system could act as an interactive adviser to doctors and nurses, offering instantaneous verification of both safety and effectiveness before medication is administered, and—especially in situations where hospital staff are tired or distracted—it would be very likely to save both lives and needless discomfort and expense.

Once medical applications of artificial intelligence evolve to the point where the systems can act as true advisers capable of providing consistently high-quality second opinions, the technology could also help rein in the high costs associated with malpractice liability. Many physicians feel the need to practice “defensive medicine” and order every conceivable test in an attempt to protect themselves against potential lawsuits. A documented second opinion from an AI system versed in best practice standards could offer doctors a “safe harbor” defense against such claims. The result might be less spending on needless medical tests and scans as well as lower malpractice insurance premiums.
*

Looking even further ahead, we can easily imagine artificial intelligence having a genuinely transformative impact on the way medical services are delivered. Once machines demonstrate that they can offer accurate diagnosis and effective treatment, perhaps it will not be necessary for a physician to directly oversee every encounter with every patient.

In an op-ed I wrote for the
Washington Post,
shortly after Watson’s 2011 triumph at playing
Jeopardy!,
I suggested that there may eventually be an opportunity to create a new class of medical professionals: persons educated with perhaps a four-year college or master’s degree, and who are trained primarily to interact with and examine patients—and then to convey that information into a standardized diagnostic and treatment system.
¹⁰
These new, lower-cost practitioners would be able to take on many routine cases, and could be deployed to help manage the dramatically growing number of patients with chronic conditions such as obesity and diabetes.

Physicians groups would, of course, be likely to oppose the influx of these less-educated competitors.
*
However, the reality is that the vast majority of medical school graduates are not especially interested in entering family practice, and they are even less excited about serving rural areas of the country. Various studies predict a shortage of up to 200,000 doctors within the next fifteen years as older doctors retire, the Affordable Care Act plan brings as many as 32 million new patients into the health insurance system, and an aging population requires more care.
¹¹
The shortage will be most acute among primary-care physicians as medical school graduates, typically burdened by onerous levels of student debt, choose overwhelmingly to enter more lucrative specialties.

These new practitioners, trained to utilize a standardized AI system that encapsulates much of the knowledge that doctors acquire during the course of nearly a decade of intensive training, could handle routine cases, while referring patients who require more specialized care to physicians. College graduates would benefit significantly from the availability of a compelling new career path, especially as intelligent software increasingly erodes opportunities in other sectors of the job market.

In some areas of medicine, particularly those that don’t require direct interaction with patients, advances in AI are poised to drive dramatic productivity increases and perhaps eventually full automation. Radiologists, for example, are trained to interpret the images that result from various medical scans. Image processing and recognition technology is advancing rapidly and may soon be able to usurp the radiologist’s traditional role. Software can already recognize people in photos posted on Facebook and even help identify potential terrorists in airports. In September 2012, the FDA approved an automated ultrasound system for screening women for breast cancer. The device, designed by U-Systems, Inc., is designed to help identify cancer in the roughly 40 percent of women whose dense breast tissue can render standard mammogram technology ineffective. Radiologists still need to interpret the images, but doing so now takes only about three minutes. That compares with twenty to thirty minutes for images produced using standard handheld ultrasound technology.
¹²