The Stethoscope of the Future
Machine Learning at the Point of Care
- By Austin Price
- 7 min. read ▪ Published
Two hundred years ago, a French physician invented the stethoscope.
Inspired by a hearing aid known as the ear trumpet, René Laennec fashioned a small funnel to the end of a wooden tube to direct soundwaves from inside his patients’ bodies to his ear. Before that, doctors used percussive methods, or placed their ear directly on patients’ bodies, to observe internal sounds. Laennec wasn’t alone in recognizing the limits of such practices, so he developed a way to help. Today, doctors seldom practice without the symbolic tool around their necks.
“There are all these precedents for fundamental changes in this field brought about by new technology,” says Dr. Ziad Obermeyer, an associate professor of Health Policy and Management at the School of Public Health. After the stethoscope came microscopy, then radiography and x-rays. Health professionals quickly adopted these tools into common medical practice, which tells us one thing: as healthcare has become increasingly complicated; man left to his own sensory devices can’t keep up. We’ve developed tools because we require them.
Obermeyer joined the faculty at the School in the summer of 2018 after years of research into machine learning as a way to better understand patients’ histories, environments, and levels of health risks. Like the stethoscope, machine learning can improve diagnoses and limit physician error.
In other words, the next fundamental tool in healthcare, he says, is the algorithm. “Big data will transform medicine,” he wrote definitively in the New England Journal of Medicine.
Machine learning—the process of running data through automated algorithms to produce a result or prediction—has become more and more commonplace in our everyday, twenty-first-century lives. Netflix recommends movies based on our preferences. Amazon shows us advertisements based on our purchase history. A self-driving car can approach a crosswalk and—after computing the world around it through light detection and ranging (LIDAR)—stop for a pedestrian.
Doctors are faced with difficult decisions, under a lot of uncertainty. They can become acutely aware of how errors in judgement can lead to really bad outcomes that affect people’s lives.
“It’s just algorithms layered on algorithms,” says Obermeyer. “And each of those algorithms parse out this super complex task into what is fundamentally a prediction.”
As an emergency physician in Boston, Obermeyer recognized that machine learning could also be applied to the increasingly complicated field of medicine. “Doctors are faced with difficult decisions, under a lot of uncertainty,” he says. “They can become acutely aware of how errors in judgement can lead to really bad outcomes that affect people’s lives.”
After his residency, Obermeyer joined the faculty at Harvard Medical School and began working with computer scientists and economists like Sendhil Mullainathan at the University of Chicago and Ben Handel at UC Berkeley to develop algorithms and data sets that could make certain predictions about patient health risks. He has focused particularly on the unfortunate number of patients who die after being discharged from the emergency room, despite no risk of life-limiting illness reported in the patients’ claims. Many of those patients, he has found, die of heart disease. In the meantime, doctors are often accused of doing way too many tests in the ER, says Obermeyer, while their diagnoses can still overlook life-saving information. Machine learning, his research has shown, can mitigate trial and error from the emergency room, helping doctors make predictions on risks of abnormal heart rhythms, even for patients without clear risk factors.
The use of machine learning in healthcare doesn’t have to be limited to the hospital setting. As researchers find ways to collect data in developing countries, machine learning could better inform global health policy. Obermeyer gives the example of maternal health. “There’s a lot of complex decision-making around childbirth,” he says. “Some women are high-risk, others are not.”
The first step would be to matriculate data, which, even in developing countries that may lack electronic health records, Obermeyer says, is getting more and more feasible. He points to Joshua Blumenstock at the UC Berkeley School of Information, who uses cell-phone and satellite data to map poverty in Rwanda. This type of data could help healthcare professionals sort through the complexities that burden them so they can better deliver appropriate care to the people that need it the most.
No matter the healthcare issue, Obermeyer sums up machine learning’s role: “Complex data in, important decision out.”
Obermeyer and his team haven’t yet implemented machine learning in an emergency room setting. Medicine, he reminds us, is of course different from other uses of algorithms like Netflix or Amazon. The stakes are higher. Randomized trials take careful planning. There are statistical problems that Obermeyer and his colleagues are currently addressing. A 2016 ProPublica investigation showed one algorithm, for example, that had a racial bias when computing judicial decisions. “These are the things you need to be careful about when designing algorithms, well before you deploy them,” says Obermeyer.
Recently relocated to the Bay Area, Obermeyer and his team are looking for partners to begin putting his research into practice. “We’re now at the point where we are sufficiently confident that the work we’re doing is ready to test,” he says. “We’re actively starting to explore to try to find partners to do that.”
Our computer screens won’t replace doctors, Obermeyer clarifies. Although airline pilots use machine learning on commercial flights more and more, “I don’t think that even today anyone believes that we’re going to replace pilots with algorithms,” says Obermeyer.
But the use of algorithms has already fundamentally guided much of our twenty-first-century technology. Machine learning has already become ubiquitous in other sciences. Astronomers no longer depend on the naked eye to chart the night sky for example. The future of medicine, Obermeyer believes, is no different.
Much of Obermeyer’s talk of the future builds on his perspective of the past. Before medical school, he earned a MPhil in science history from Cambridge.
Think of the printing press, Obermeyer says. In the 15th century, Johannes Gutenberg invented a way to use movable type to mass produce printed text. The most obvious port of this technology was the first printed Bible—followed by a complete fundamental change in the spread of information and a subsequent revolution of thought. “Of course, no one saw the fundamental transformation that the printing press was going to produce in society,” says Obermeyer. “It’s like that with any transformative new technology. We just can’t imagine what may happen.”