As part of our AI in your community series, I had the opportunity to talk to Katherine Heller, an Assistant Professor at Duke University, who studies machine learning and Bayesian statistics and applies that work to problems in the brain, cognitive sciences, and social interactions. We discussed her many projects, how they connect, and what she finds exciting about Artificial Intelligence as an entire field.
Katherine Heller
Tara Chklovski: I’d love to learn a little bit about your research. What area, topic, and problem are you working on?
Katherine Heller: In an applied sense, I do research on a couple of different areas. In a methodological sense, it’s really one thing. I’ve done a bunch of work related to social science and social media networking and how people influence one other on Twitter or, in discussions or other contexts — looking at controversial Supreme Court cases or Federal Reserve board meetings and things like that.
On the other end of the spectrum, I have done a lot of work in application to medicine — everything from sepsis prevention to neurology. I have a secondary appointment in the neurology department.
TC: Sepsis prevention? How?
KH: So basically, just, like, early detection of, you know, whether somebody is decompensating in a hospital setting. Most people get sepsis in the emergency department or the ICU, where there are signs that they picked up an infection, maybe off of one of their lines or something like that.
TC: So how does all of your work connect?
KH: Underlying all of these projects is machine-learning methodology that I work on. So, it’s like Bayesian statistical methods. It’s always time-series Bayesian statistical methods with the social stuff; only with some of the medical stuff is it networking. Taking the sepsis work for example, you look at a patient’s vitals and labs and other things like that over time as they’re staying in the hospital, and basically you learn that when you see certain patterns in their labs and their vitals jointly, that they’re at risk for having sepsis. And what you want to do, because they’re taken so infrequently, is be able to estimate in between the times when they’re taken.
TC: Interesting.
KH: So also, there’s this app that I developed for multiple sclerosis. We use electronic health records and at Duke we’re working on pulling data out of the electronic health records in real-time for patients that are in the emergency department. But one of the things that exists less frequently is data on people who have chronic diseases. There are a lot of people who have chronic diseases, and the problem is that they’re mostly not in the hospital setting. They’re mostly at home, living their life. And so, the question then becomes, how do we collect data on these people? That’s why we’ve started with iPhone apps and other wearable sensors, trying to get data on a more continuous basis about people who have particular chronic diseases. We can use that data to make predictions both for them and for the population as a whole.
TC: That’s fascinating. So how many projects are you working on now?
KH: The neurology one and the sepsis one are both current. I’m also working on something that’s Twitter-feed related. It started as an identification of online bullying in the Twitterverse, and in the end, it’s more like identifying who started a thread and whether there are latent events going on.
TC: So online bullying in adults?
KH: In adults or kids. And you’re right, the behavior might not be quite the same, but online bullying overall was what we wanted to get at, but it’s very hard because it’s hard to get that data. I don’t work at Facebook or Google and I don’t have access to all of the Facebook or Google data, but Twitter lets you scrape data. And so, we can do things like scrape data from Twitter, which has bullying-related hashtags… but that’s not necessarily instances of bullying. It’s very conflated. It could be that that bullying movie came out and then you could have some other instance of bullying that’s just not tagged as bullying at all.
TC: That’s so interesting. Our project is basically empowering children and parents to find a problem in their community and to use some of these technology tools to solve that. So, one question to you would be: what advice would you give them on how to find a good problem to solve using technology?
KH: Oh, so you want to choose a problem where you can say, “okay I’m going to collect these particular things, these particular variables or predictors and I’m going to use them to protect some quantity of interest to me.” Say I want to be able to link water-quality measurements to health outcomes. I can do that because I’ve got the health outcomes and because I’ve got the water quality measurements. If you don’t have access to those things, then it becomes really difficult, particularly for young people because it’s hard to show up and say, “hey, give me your water-quality measurements.” So you want to have a problem where you have predictors, where you have an outcome, and where you have access to both.
And then you also want to be dealing with a problem that’s important to you. People work a lot more when they have a personal investment. For me, for the medical stuff, my mom died of a traumatic brain injury, and I lived in a hospital for six months. Everything I learned and saw impacts the research that I carry forward. I was not doing medical research at all before then. I kind of decided at that point that what I really want to do is be able to make the practice of medicine better. I think I’m way more motivated to address that problem than I was almost about any problem before that.
So yes, something where you can sort of logistically get ahold of the data and where you have a problem that’s meaningful to you.
TC: Thanks for sharing that story. What inspires you?
KH: I talked a little bit about medicine and medical applications and why I ended up doing that. When I was [in the hospital] I saw how bad the usage of data was, and I thought “we can do a lot better than this. I know we can do a lot better than this.” So, you see the problem and then you also see the opening to apply what you know in order to make it a lot better.
In terms of machine learning and AI as a whole, I really got into it starting when I in high school, and it was interesting because computing was just kind of coming about at the time. It was something that was present in my parents’ life a little bit with punch cards and things like that, but not to the extent it was in mine. I really liked it because it was something adults didn’t know a lot about.
It was like almost like my own space where I could do my own thing. And that’s me. I have a very independent drive, and so I wanted my own thing and I wanted something where I felt like I was contributing in a way that somebody who was 50 years old couldn’t. That’s how I ended up going into computer science and then AI.
TC: Cool. What do you find difficult in your work and how do you overcome that difficulty?
KH: There’s a lot of difficulty. There’s a lot of difficulty on multiple different levels. So right now, I would say one of the things I find really difficult is working with the medical school at Duke. I’ve got to work with people who are, in a sense, culturally very different from me. This is not an ethnic thing or a gender thing. But it’s a job thing. And the way that they have MDs, have been trained, and even the things that they value are very, very different from the way that I’ve been trained and the things that I value.
TC: What do you value that’s different from theirs?
KH: Data science, for example. I see data and its ability to impact medicine and the way that I think that medicine is going to operate in the future. And I think that I’m right. I think the future of medicine is to have a lot of these predictive algorithms in place. I hope so, because they do a lot better than human judgement. But I think for somebody who’s an MD or somebody who’s running a hospital, they care a lot less about that and a lot more about the day‑to‑day. If they’re a doctor, it’s more about “am I addressing this person’s need right now at the hospital?” And there’s a lot less focus on the future of where everything is going.
TC: Interesting. Last question: what field or aspect of AI excites you?
KH: So kind of all of it. I think its ability to make an impact on all of these areas. I think that these are all things that we can do going forward. I think we can have an impact of medicine so that the level of care that we give people, the amount of lives that we save, are significantly improved over what’s happening now. I think we can eventually get algorithms which do a good job of predicting and flagging whether somebody is being bullied online, and figure out how it is that we want to report that or deal with that.
I think that we’ll be able to identify who’s influential, in what discussion, and what it is that is actually going on under the rug.
TC: Awesome. Thank you so much Katherine for sharing your time and your thoughts.
KH: Thanks. I really appreciate it.
