One of the things I'm most concerned about is actually this question of fairness in machine learning.

[music and background talking]

My name is Irene Chen.

I am a third-year PhD student at MIT in Electrical Engineering and Computer Science.

My advisor is Professor David Sontag in the Clinical Machine Learning Group.

Healthcare in particular has a long history

of having disparate impact on different groups.

So with machine learning we have access to a lot more people's data,

people have more voices, people get more represented in the machine learning models

that we could then potentially roll out.

[music]

I am working on heart failure specifically.

Heart failure is a chronic condition

where after diagnosis people live on average about five years.

We want to be able to determine

what kinds of treatments will have better outcomes than other treatments

based on who you are your age your past history, your lab results, all of that.

And then there's also different types of heart failure out there

so we're trying to figure out,

"Are the types that we think are out there even the same types that there should be?

or it be slightly different classifications?"

We have really close collaborations with hospitals in the downtown Boston area.

And they have allowed us to have access to a lot of their electronic health records.

So every time a patient comes in with heart failure diagnosis

we can actually pull all the other times they've come to the hospital

and maybe they didn't know that they had heart failure

and we can see, "Oh, three years ago you had this abnormal lab result

"and we maybe didn't think it was anything but now, knowing what we know now

we can pull it all together and build this patient trajectory.

And then using your patient trajectory and then all these other patient trajectories,

when a new person comes in, we can build potentially a patient trajectory for them as well."

[music]

Health care machine learning has been going on for a long time.

It is only recently though that we have access to all this data

because a lot of hospitals are now digitized.

So with this increased amount of data

we can now start to take some of the techniques that people worked on in the 80's and 90's and 2000's

and see does it work on five million patients

instead of the few hundred you could sort of cobble together and hand label.

Now, we have much larger datasets, much more powerful computing powers,

that we can bring it all together and do really great health care for machine learning research.

[music]