So you look, and it's
a talk with a title and a question mark
at the end of it, and you have a right to assume
that, by the end of the talk, the speaker will
at least propose an answer to the question. I'm actually going to put
you out of your suspense immediately. No. All by itself, AI is not
going to fix the US healthcare system. It might need help from
those people in Washington. All right, so I should have an
obligatory conflict-of-interest disclosure because
NIH insists upon it. In addition to my
role at MIT, I'm the CTO of a
healthcare AI company. Let's start looking at
health care around the world. We'll first look at economics. These numbers-- maybe you've
seen them; maybe you haven't. But it's kind of surprising
to think that, in the US, we spend twice as much
per person on health care as they do in Germany, way
more than in Canada, Japan. But maybe it's OK. Maybe it's because it makes
us all so much healthier than these other people. Well, apparently not. It does not seem to make
us all that much healthier. In fact, as you can see,
we're not doing very well on the world stage. I've only shown you
four countries here. I could have shown
you another 10, and the trend would be
pretty much the same. What's going on here? Is it that we have
crummy doctors in this country or maybe
venal hospital administrators? No, at its best,
medicine in the US is as good as it gets
anywhere in the world. The problem is that we have
a healthcare system that means many of our
citizens and residents don't actually get
access to this good care. So what should a
healthcare system do? It should provide access
to the right treatment, the right provider,
at the right time, at a cost society
considers appropriate. What could be easier? Well, actually almost
anything could be easier. But predictive models can
make a difference here. They really can help. So I'm going to only look
at one of these problems today because I only
have eight minutes. Let's look at finding
the right provider. Could be easy. You just open up
a local magazine. It tells you who
the best doctor is. And you go to him or her. That would be-- I hope this doctor
is not in the room. I'm not intending
to be offensive. But it would be a
bad idea to do this. The problem is that
the typical approaches to choosing the
right provider assume that there is such a
thing as the person who is the right provider
for everybody. So it's the CMS quality
stars, consumer ratings, reputational rankings, volume. None of these things actually
lead to better health care outcomes. And there's been many
articles showing that. The issue is that
different providers do well for different kinds of patients. So what we really
need is not a ranking of providers but an app that
matches patients and providers. Now, I'm sure this
seems simple to you. How many of you have
used one of these apps? You don't have to
raise your hand. How many of you know somebody
who has used one of these apps? Probably almost everybody. Unfortunately,
these apps are not very good at choosing doctors. So what we want is the best
provider for each patient. So what we do is we
use machine learning to build models
of every provider in the country that tells
you what kinds of patients that provider does well with. So here we have two
fictional patients, John Doe and Jane Doe, presenting
with the same symptoms but having very different
demographic information and, more importantly, very
different health care history. If we run a model to choose
providers for those patients, we get very different answers. So what the model is
doing is predicting the rate of adverse outcomes--
for example hospital admissions, visits to
the emergency department. And what you can see the
model is doing for these two individuals predicts
very different rates of adverse outcomes, depending
upon which physician they see. And it's a different
choice of, quote, "best physician" for
each of these two people. All right, this is just a model. We all know we can build
models that make predictions. And the real question
comes when you take these predictions
in the field, do they actually lead to
good outcomes, improvements? We looked at a
couple of examples, actually more than a couple. Let's start with orthopedics. We did a study of 4,000 patients
who received hip replacement surgery in Chicago. They were all Medicare patients. We trained the model
on two years of data and then tested it on the next
year and then compared it, the model one, which
is on the right, to a bunch of different
conventional methods. The best conventional
method was volume-based. Unsurprisingly, hip surgeons
who do many surgeries a year are better at it
than surgeons who do a handful of surgeries a year. So all else fails, go to
someone who does a lot of it. And you can see, if we
look at 90-day admission, there's a 13%
improvement if you go to somebody who does
a lot of hip surgeries and an 8% improvement in visits
to the emergency department the next 90 days. And a little bit
surprisingly, a slight cost increase over 90 days. Maybe they're just more
expensive surgeons. On the other hand,
if you look at what happens with the
machine learning model, there's a much more dramatic
relative improvement, 36% improvement in 90-day
admissions, 23% in ED visits, and 12% in total cost of care. Maybe it only makes a
difference for hip replacement. Well, we looked at a much larger
study of a million Medicare patients for a one-year
follow up on visits to a variety of
different specialties. And what we're
looking at here is the reduction in either
emergency department visits or hospitalizations
per 100 member years. These are members of Medicare. So you can see, for
cardiac surgery, going to the right surgeon
roughly results in one fewer trip to the hospital per year,
which is a huge difference. Everything is pretty much
an improvement except EMT, and that's flat. So indeed, choosing
the right doctor makes an enormous difference. To wrap up, healthcare
and medicine are not the same thing. Healthcare should
deliver high-quality care to the population at a cost
that society can sustain. But even though we're trying to
deliver it to the population, we have to make the
decisions not based on averages but on individuals. And to do that at
scale, we really need to deploy AI-based models. Thank you. [APPLAUSE]
