octors will one day be able to more accurately predict how long patients with fatal diseases will live. Medical systems will learn how to save money by skipping expensive and unnecessary tests. Radiologists will be replaced by computer algorithms.
These are just some of the realities patients and doctors should prepare for as “machine learning” enters the world of medicine, according to Dr. Ziad Obermeyer, an assistant professor at Harvard Medical School, and Dr. Ezekiel Emanuel of the University of Pennsylvania, who recently coauthored an article in the New England Journal of Medicine on the topic.
But what exactly is “machine learning”? And how will medical systems make use of it?
Obermeyer, who is also an emergency physician at Boston’s Brigham and Women’s Hospital, spoke with STAT to provide some answers. This discussion has been edited and condensed.
How is machine learning different than, say, artificial intelligence?
The traditional approach to solving problems with technology is to give the computer some rules and apply brute computing force. With machine learning, you don’t actually give machines rules. You give them data and ask them to learn the rules.
We can point this very powerful tool at a medical problem and say, “I’m going to show you a bunch of people who had heart attacks, and a bunch who didn’t. Go learn how to tell them apart.” Then, once the algorithm has seen a million patients and what happened to them, you can show it information about a new patient and let it predict whether he might be at imminent risk for a heart attack.
These algorithms are extraordinarily good at telling the difference. What we need to know more is, what are the rules the machine is learning, and how did it arrive at those rules? That’s sort of the next frontier of this.
You can’t see the logic the computer has used to reach a conclusion?
That’s the really weird thing. With machine learning, whether in medical or other settings, you actually just have predictions.
Will doctors be willing to accept the conclusions of an algorithm without understanding how it achieved those conclusions?
That sort of thinking isn’t foreign to medicine. With something like the development of the hip replacement, it was sort of this engineered product that came out of a deep understanding of the mechanics of the hip. But in the history of medicine, there’s also a lot of things that make less sense. Think about the discovery of steroids for immune suppression, where medicine begins with a very pragmatic observation of, “Oh, this thing works,” and then goes to work in trying to backfill our understanding of why it works. That’ll be the model for a lot of machine learning applications.
Your article suggests that of all medical jobs, radiology, in particular, may face the most profound changes as a result of machine learning. Is it still a smart career choice for med students?
In 20 years, radiologists won’t exist in anywhere near their current form. They might look more like cyborgs: supervising algorithms reading thousands of studies per minute and zooming in to inspect and adjudicate ambiguous cases; or they might transform into “diagnosticians” like Dr. House, that go out and have more contact with patients and integrate that into their diagnostic judgments.
Think about construction workers: They are still indispensable for construction — but they are doing very different jobs today than they were before mechanization 100 years ago. Bank tellers don’t hand out cash anymore, but they do handle far more complex transactions than they used to.
Technology doesn’t always eliminate jobs; sometimes it changes them, and sometimes it’s beyond recognition. So those that adapt, as I’d imagine highly educated doctors might do, can turn out to be big winners.
What will be the first place or manner in which patients notice the impact of machine learning, and when would you predict that?
One of the areas where we’re going to see the quickest transition is in making very personalized predictions for individual patients based on their unique histories and trajectories. In areas like end-of-life care, or optimizing the use of diagnostic tests to make diagnoses. We’ll see lot of these applications come on line very quickly in the next few years.
Right now we don’t have the studies yet to show that this directly improves care. Once they start coming out, though, I think there’s going to be an explosion in this kind of technology, where we take the information available — that’s produced as the exhaust of the clinical care we’ve delivered — and recycle that into very tailored predictions about risks and future outcomes for patients.
What’s the best example of that?
Better prognostic information about the end of life. “How much more time do I have?” is a very common question and it’s one that doctors are pretty ill-equipped to answer. We sort of give estimates that by most measures are way off from what patients end up experiencing, or doctors might not give any information, partly because they don’t want to, but also partly because they don’t know how to.
Predicting remaining life span for people is actually one of the easiest applications of machine learning. It requires a unique set of data where we have electronic records linked to information about when people died. But once we have that for enough people, you can come up with a very accurate predictor of someone’s likelihood of being alive one month out, for instance, or one year out.
That information is tremendously valuable to patients and doctors. Patients have a lot of things they need to plan around end of life, whether that’s advance directives or medical proxies, but also for doctors who need to know how to think about a treatment plan and diagnostic tests for the next few months.
What’s the biggest obstacle to achieving meaningful benefits of machine learning? And how do we get past that?
There are two. First, these data are extraordinarily messy when they come out of the electronic health record or the insurer’s database. And it’s also very hard to link them to other data. Let’s say you want to know when someone died. If they go home and die, or they die in another hospital that’s not part of your health system, you wouldn’t know they died. So the fragmentation not only within our health system but across the layers of our health system is a huge challenge. It’s solvable, but it takes time.
The bigger — not so much obstacle, but requirement — that hasn’t been met is that, in medicine, we actually have a playbook for changing medicine, where we do studies and we measure in a very rigorous way whether or not something works.
I think for all the enthusiasm for machine learning in clinical medicine to date, it hasn’t been matched by a burgeoning of activity in the rigorous testing of ideas and interventions. It’s all very well and good to say you’ve got an algorithm that’s good at predicting. Now let’s actually port them over to the real world in a safe and responsible and ethical way and see what happens.