Motivating Teams and Combating Bias in Healthcare Data Science

DAVE COLE

Hello, and welcome to another episode of the Data Science Leaders podcast. I’m your host, Dave Cole, and today we have Vikram Bandugula. Vikram, how are you doing today?

VIKRAM BANDUGULA

I'm good. Thanks, Dave. How are you?

DAVE COLE

Good. I'm great.

Vikram, you are the Senior Director of Data Science at Anthem. Today we are going to be talking a little bit about overall data science, healthcare, and the types of problems that you solve there.

Then we're going to segue into bias in healthcare. We talk a lot about bias on the Data Science Leaders podcast, something that all data science leaders need to be thinking about, but how specifically does it pertain to healthcare?

Last but not least, I know you have a fantastic team with lots of PhDs. You're going to give us a few tips on keeping a high-performing team motivated. Does that sound good, Vikram?

VIKRAM BANDUGULA

Perfect.

DAVE COLE

Awesome. Let's go ahead and get started.

First of all, you've been at Anthem for a little while. What intrigued you about getting into the world of healthcare?

VIKRAM BANDUGULA

I would say it was an accident, Dave. When I finished grad school, I started working for Discover Financial company. I was there for eight years. Then I worked for GE for a few years, but I was in finance. Then I did some startups and then consulting.

One thing I realized what I was good at is actually solving business problems. It didn't matter what industry they belonged to. I was technically good, but it helped me to solve different things. I realized that I never did anything in healthcare. Then the opportunity presented itself. I jumped right into it. It has been two years in healthcare, in fact more than that, but I feel like I was meant to do this. I enjoy the process.

DAVE COLE

What is it about healthcare? Walk us through a day in your life. What are the problems that you focus on?

VIKRAM BANDUGULA

When I look at my calendar, it's all packed, but the best thing about it is that I'm trying to solve problems which will impact our members' lives. We're not recommending people buy coffee mugs or to read a certain article and then do a workout. Our stuff that we do actually impacts people's lives because we are trying to improve—that's Anthem's actual vision, by the way, improving lives in communities.

When I feel like we are doing something meaningful, even if it's a small thing like recommending the right things for a member to get their screening done. We did something for a member where a nudge got them to do their A1c test and thus manage their condition better. Even those small things, to me, make me feel that we're doing something. To me, that is pretty rewarding.

DAVE COLE

Right. I would imagine that there are not many data science leaders out there who are trying to optimize people’s lives, make people's lives better, and potentially even save lives.

Is there any specific project that you can share with us, at least at a high level? If you're a data science leader in healthcare today, what are some things that you think are current and topical that should be focused on?

VIKRAM BANDUGULA

I think that there are all kinds of problems with healthcare. It’s why I’m excited about healthcare and the problems we're trying to solve. Obviously the data is different because we have claims data. You're talking about people's conditions, how they engage with healthcare. If you look at the calendar, what I see is that today we are at a meeting on predicting women who could be pregnant and how do we provide the right care for them. There was another meeting today to review results on our breast cancer screening awareness program.

DAVE COLE

Right.

VIKRAM BANDUGULA

How do we ensure that especially women above 65, get a screening done every 10 years. We have a bunch of programs we optimize for. We have our meeting on how the model is performing and how the program is doing.

Like with any business, even we solve typical business problems, like increasing revenue, increasing engagement, etc. but there are some unique problems in healthcare that we solve. For example, care management is a big part of the problems we solve. How do we provide the right care to the right member at the right time?

Thinking about data science, if you remove all the top things, it ends up being a recommendation engine. How do you build a recommendation engine? Those impact lives.

Another problem we try to look for is fraud, waste and abuse in healthcare systems. Somebody does those things and sometimes doctors over-diagnose certain things. How do you ensure those things don’t happen? Condition prediction is a huge thing. Can we predict somebody's going to be diabetic in the next six months or one year from now?

DAVE COLE

Wow.

VIKRAM BANDUGULA

Diabetes and hypertension are two big problems people have, even though they’re both manageable, that can cause a lot of other conditions down the line. Can we identify members who are going to be diabetic, members who will have hypertension or even members who are going to have a fall or risk of falling?

DAVE COLE

Wow.

VIKRAM BANDUGULA

These are like condition-predicting models and then the other ones are like management prediction. I look at the calendar and I’ll see a meeting to brainstorm.. How do we build? What kind of model do we need to build? Supervise, optimization, deep learning etc..

In other meetings we review the performance of the models, in others it’s about model bias. It's a big thing nowadays. I have an engineering background, so I like solving problems. We are trying to solve problems with AI, which impacts lives and has a consequence.

DAVE COLE

Right.

VIKRAM BANDUGULA

It impacts members, their lives and financial outcomes. Every day I get up in the morning, I'm excited to do my job and solve problems like this.

DAVE COLE

That's awesome. When you were describing it, it sounded a lot like marketing use cases: right message, right time. The difference is that instead of encouraging people to go buy a product, you're encouraging people to go get screened or to come in for a checkup or what have you. That's incredibly interesting.

Are there any specific challenges that you run into? I know that getting the ground truth can be difficult in all sorts of different verticals. I have to imagine that in healthcare there's such a lag. Are there any challenges there that you struggle with when working in the healthcare industry?

VIKRAM BANDUGULA

Yeah, absolutely. One of the biggest struggles I see is around what businesses actually want to do and what is good for a member. Most of the time they should be aligned.

DAVE COLE

Yeah.

VIKRAM BANDUGULA

Both should be aligned because a happy customer means a happy business. For a health insurance company, you want members to pay the premiums and not go to the hospital.

DAVE COLE

Right.

VIKRAM BANDUGULA

That's the best case scenario but we are trying to build something where the model thinks that someone shouldn’t be doing something specific. There's business bias there. Maybe it's not the right thing to do for our members. That's where I think you would see there's a struggle. A lot of times we’re educating business leaders about what we should or should not be doing.

The other struggle that I see is that the business team looks strictly from a business point of view, at problems. Data scientists solve problems from a data science point of view. They're trying to solve problems from very different angles. The questions being asked from either side are very different. That’s where I see myself playing a bigger role. I'm trying to convert information in a way that the business and the data science team will understand. Maybe the business needs to rethink its problem or the data science team needs to ask different questions. That's where I think you bridge the gap where you're trying to educate the leaders about what a data science team can do and what we can offer. It's not like a magic wand that you just wave around and suddenly it comes out alright. I think when people try to do that, bad things happen in terms of recommendations. There's enough evidence in history where we have recommended bad stuff. That's what we try to avoid.

DAVE COLE

Yeah. That's a recurring theme here on the DSL podcast: how to communicate with your business counterparts. There's certainly educating them, demystifying the dark art which is data science, and explaining how models work—being more transparent in general. Are there any tips there that you have for fellow data science leaders communicating business?

VIKRAM BANDUGULA

Try to avoid Greek letters in your PowerPoints. They love showing how the model works, but I think the key thing we should focus on is what the model is supposed to do. I think we are too fancy with showing these curves and metrics like RMSC, precision recall, and all those things. I think that leaders don't have a lot of time. They're sitting there for something like five minutes and they just think as passers by. They're also probably doing other things while they're talking to you.

DAVE COLE

Right.

VIKRAM BANDUGULA

Let's also make sure that we remove all the basic assumptions that come with the basic tools when you want to explain what the model is supposed to do. Summarize, for example, in two minutes that the model’s goal is to predict the volume of women who will be pregnant in the next six months.

Then ask the business team about what they plan to do, because you want to make sure that they use this model in the right way.

DAVE COLE

Right.

VIKRAM BANDUGULA

Don't use the model the wrong way. You don’t have it and play with those things. A lot of times we are only giving outputs of the model, but we're not asking enough questions on what do you plan to do with that. I think as data science leaders, that's a very important thing because you don't want your models to be misconstrued or used in a way that they should not be used. That's when you'll have un-optimized results.

DAVE COLE

Yeah. If I recall correctly, a famous retailer a few years ago also tried to predict women who would be pregnant, then sent pregnancy-related coupons to a young girl who was at home. Her mother got home and wondered what the heck was going on there. You have to be very careful about how you use the outputs of the models at all times.

Part of the job of a data scientist is educating and talking through the pros and cons. Getting back to your larger point, there’s a time and a place to walk through the how, explain what features and algorithms you're using, how you got from A to B and how you got to that final model.

I think that's best left for your fellow data scientists when you explain how you did your work on a brown bag session or something along those lines, but when you're talking to business, they just want to know what this thing does. Is it better than what we had before? How can we go ahead and deploy this into an existing business process? I think that's very good advice.

You mentioned bias in healthcare. We've talked a lot about bias in previous episodes, but not specifically with regards to healthcare. I have to imagine that there might be some unique twists to what you have to deal with, with regards to bias. We should all be aware of bias in our models, but when it comes to healthcare, I imagine even the degree to which you need to pay attention is even greater. So I'll let you take it away here, Vikram. Talk to us a little bit about bias in healthcare.

VIKRAM BANDUGULA

Nowadays, bias has become a buzzword. Everybody uses the word for everything. It's actually very difficult to even identify bias, forget about fixing it.

Let's not even worry about fixing it because our goal has to be, as a data science leader, to see if we can identify bias in our algorithm. Is the algorithm providing bias? I'll give you an example. Right now we are building a model where we're saying that we want to manage members’ diabetic conditions. We know these members have diabetes and then we want to manage the conditions. The true outcome, what we want them to do, is to get that A1c test done once every quarter. That's one of the best ways to monitor the condition. If it's high, you go to the provider, they do something about it.

Let's say the business comes and says, hey, we have 100 dollars. Optimize this. You have this fixed amount. You have calls, text and an incentive. Put all these members into those programs and then let's make sure that we increase our, let's say A1c test results, whatever. The thing is, you take this problem, you convert this into a simple X and Y supervised learning thing. We have PhDs in our team who can build the best model out there, the best optimization model. You could give them all the features. They will build the best solution. They'll even remove all those demographic ones, things, biases, look for income and all those things and then some, but does it really mean that we don't have bias? Have we checked for everything?

How can we ensure that we are not recommending only to people who are rich, who have access to phones and text messages; not just recommending to them but calling them and reminding them. People who don't have access to those kinds of things, we are not even picking them up because they don't fit in the top decile. How do we know that? Who would need incentive versus who would not need incentive? How can we know? These are very difficult questions. What we do is, for us, take responsibility as a data science team. It's not the responsibility of the business team that bias exists. I think it's the responsibility of the data science team actually to ask if we check for bias.

We have this checklist and we use IBM, the University of Chicago—they have a bunch of checklists that are put together. From there, we have this good checklist of things that we go over and we check for target bias. Is there actually bias in the target itself? We check for program bias. What I mean there is that we see whether the raw data which we have shows any bias. For example, do we see that a lot of programs that we ran were run in richer neighborhoods? We ran it on people who have access to healthcare and things like that. So we ignore certain groups of things. That’s a program bias in that data itself. Is the ground truth biased?

All our models are going to behave as supervised learning. They're going to be very good at exploiting those things so that’s what we look for. The big thing that we would look for is algorithmic bias. Is the algorithm exploiting that bias? Once we build the models up, we actually run the algorithm bias against age, gender, income, ethnicity etc. but you can argue, Dave, that age and gender might actually be part of the features. Should we even check for it? Because there might be a certain inherent bias in those things. How do we know? How do we ask? We need SMEs, people who understand the system, clinicians who understand this, doctors who can apply in all those things etc.. We work with them and ask if it makes sense. We share what we are seeing in the data. Sometimes they call out a program bias and that we shouldn't have done that. They're just calling bias.

After that, we do calibration curves or a stratified sampling of programs to ensure that for calls, text and then incentive, will have the same number of people from income. It might cost us in the performance, but that's okay. That's when we seek agreement between business and data science leaders.

DAVE COLE

Right.

VIKRAM BANDUGULA

I think it's the right thing to do. I'd start with data science because we are seeing the data from the ground view. Then you include your clinicians, SMEs, and then you have approval or awareness and alignment from your business leader about this being the right thing to do and that we should do it. That's where, together as a team, we can actually even understand what bias is before even going to the place of fixing it. It starts with the process. Every team, every company nowadays, should have a good process in check. Testing it is an integrative process. You're not going to get it right for the first time.

DAVE COLE

Right.

VIKRAM BANDUGULA

You need people to question it. You'll need people to check and evaluate it. You need to bring some lawyers too, if you have to. Once you have a good process, it should be part of the job.

DAVE COLE

Right.

VIKRAM BANDUGULA

Your job is not to build a model. It's to actually check to see that it's a non-biased model.

DAVE COLE

Just to go back to something we just discussed Vikram, where the ‘how’ is important. If there is a time and a place, obviously, discuss what sort of approaches, checklists and processes that you have in place. It will help to hopefully lower bias in your models. You talked about a few ways here. Is there a governing body within Anthem, that you work with, that sort of reviews your overall process? Or is it on a case-by-case basis with your business counterparts depending on the model?

VIKRAM BANDUGULA

In my team, when data science builds models, there's a principal or director who oversees it and they check for all the checklist items and it goes through. Then within my peers, we have a governing body. My boss is with whom all the data centers sit. Then we have our own checklist where we check and ensure we can even do peer audits. I want to see what’s required if this is a big visibility product. Then our chief AI officer has a data science team which checks for this bias. It's a governing body which checks, and then Anthem has its own bias analysis. They do audits, regular audits, with models and they go through those things, review them and color-code the findings in green, yellow, or red.

DAVE COLE

Right.

VIKRAM BANDUGULA

Sometimes we even may bring in an external team, like consultants, to come and audit our stuff. A few years ago, a big healthcare company’s models were recommending something unwanted. There's a huge consequence but you don't want to have algorithmic bias in your systems. As I said, our recommendations impact people's lives. We want to at least sleep at night and know that we’re not ignoring a group of people altogether.

DAVE COLE

The data that you get from the higher income members is richer and better than the data from some of your lower income members.

VIKRAM BANDUGULA

Yeah. We have high quality levels at Anthem at every stage: the team within; outside the team; within the chief AI officer; outside of Anthem—we have these teams which actually have regular checks. We go into details: you'll be looking at curves; decile charges; accuracy at each stage. For example, it would indicate that a model is pretty good at predicting women groups, but for men it's not that great. So what do we do? So high-income models are pretty good because there's a lot of information about them.

DAVE COLE

Yeah.

VIKRAM BANDUGULA

Low income, not so much. So what do we do? What kind of questions do we ask? There are a lot of details and those are not extracted in something like 10 minutes. You are getting into details. You are asking people to ask questions.

DAVE COLE

Right. You mentioned a couple times and I know just from my past experience that this is something you have got to be aware of. There's this governing body that will occasionally perform audits, which is a good thing. It's meant to ensure that you, as a data science leader, are consistently trying to remove bias to the best of your ability. Is there anything that has helped you in those audits? Is there any documentation that your team puts together? We talked a little bit about reproducibility; having the story with how the model was built. What would you recommend that you have in place just to be prepared for those audits?

VIKRAM BANDUGULA

One of the first things is you ask the key question that other people don't directly ask: how is the model consumed today? That's the key thing. For example, most of the models are, give it top 10 percentile members and we are going to do an outreach for them because you want to optimize for that. But sometimes it's not that simple. The question you need to ask is how the model is being used. Once you know that pretty well, and its downstream systems impact, we are pretty good. Be very clear on how it's been consumed. Then, for example, if you are targeting all the people, some people get texts, some get ads, some get phone calls. That's a different problem versus targeting the top one percentile, to whom we're going to give an incentive. Now we are going to be very careful about what it is.

DAVE COLE

Right.

VIKRAM BANDUGULA

That's why the first thing is to know how your model is being consumed. Ask. If you think that you already asked enough questions, ask one more. Make sure that you know it. The second thing is you do your checks for the minimum. You do check for your target bias, your program ground truth. Is your ground truth actually biased? First check to see either and then increase, looking for the key things: age, ethnicity, income levels. Those are the minimum things that you should have. Let's make sure that you first check to see your target bias.

After that comes your algorithmic bias. Look for precisions within those groups. Look for recall within those groups. Are you able to do it? If there's no data, that's a different problem. If you have data in enough cases, see what it is. If you do the one and a two, and if you just check for target bias and algorithmic bias, I think you are in a pretty good place because now you know exactly what it is. Let's say you're only using the top decile, then you just focus on the top decile. Check to see within those individual subgroups: what are your metrics? How are you doing?

DAVE COLE

Right.

VIKRAM BANDUGULA

Once you have identified those things, you'll be at a pretty confident place that your model's not biased, or if you need to check something here or there. Those are two things, I think at a minimum, we should be able to do.

The second thing is that your data scientist should know what's happening in your models. Don't create black box models. Nowadays, everything can be explained.

DAVE COLE

Right.

VIKRAM BANDUGULA

You should be able to explain your models. I sit in meetings and I'm going to ask them for the top features. Not even top features, like give me a random forest of office topics. No. Tell me what a certain term means, or whatever. Tell me why you scored this member high. I think you should expect that. Nowadays they have too many auto mail tools and people are becoming lazy data scientists. What they do at college is take the data, have the ground truth, throw a bunch of peaches at the wall and then see what sticks.

DAVE COLE

Yeah.

VIKRAM BANDUGULA

I think that's not the right way.

DAVE COLE

No, especially in your world for sure. You can't get away with that. Well, cool. Let's say we are into our last topic. You have a team, it sounds, that has advanced degrees. My guess is that they came from a world of research, potentially working on cutting-edge problems. These days you can even get a PhD in data science, believe it or not. Most folks who have PhDs, however, usually acquire them in an area that is not statistics or data science. Sometimes it's physics, chemistry etc. How do you keep your team highly motivated, engaged and excited, other than saving lives? That, in and of itself, would certainly keep me motivated but what other things do you do to attract and retain top talent?

VIKRAM BANDUGULA

As I said, solving problems keeps me excited. Solving problems with people with whom you enjoy solving the problems—that's important.

DAVE COLE

Yeah.

VIKRAM BANDUGULA

You don't want to be a lone person solving problems. It's going to be pretty tiring there. It's going to be a pretty empty hill up there or it will involve being alone. I think I enjoy the process because I'm doing it with my coworkers. My team has the same passion as I do, like solving problems and members' health. The only thing is, I was joking around this one, the majority of my team has PhDs and I don't.

DAVE COLE

Yeah.

VIKRAM BANDUGULA

Don't get me wrong. I don't think it's important or necessary to have a PhD. No, it's not. You don't even have to go to college, as long as you know the statistics, understand math and have done enough projects. I think you're good to go as long as you're able to solve problems. I don't think that you need a PhD by the way. I think it helps because you dedicated your life, those four years of your life at least, to solving a complex problem.

DAVE COLE

Right.

VIKRAM BANDUGULA

As a leader, one thing that I keep looking for is how to keep these guys engaged. How do I keep them solving? They should not feel that they are wasting their time or something like that, so one thing I always help them understand is the ‘so what’. Think about it. The recommendations they build are going to impact 40 million lives. The model is going to go to 40 million emails and then they're going to see what my team member wants them to see. Just imagine the thing. In the care management, fraud, waste spaces, or whatever it is, there’s a focus on a dollar amount. Affecting 300 million dollars. Yeah, that's okay. But having the knowledge that they, without being a doctor, saved somebody from a surgery with a simple and well-constructed suggestion, is powerful.

The second thing I'll try to do is try to understand where they each are and want to be. Most of these guys, they're highly passionate about certain things. You would be surprised, Dave, where they want to be. Very few times actually it's a promotion or more money. It's about something else. They want their job to be much more fulfilling or to lead a team.

They want to get better at solving business problems. They want to get better at leading teams, like me. I'm trying to understand what's the language I need, to speak to them, motivate them, inspire them to be good at their jobs. Once I know the language, who they are, what and where they want to be, it's just creating opportunities. How do I create opportunities? How do I put people in those positions to be successful, to be good at what they are? Then I continue to mentor and encourage them because I am who I am because of the work and the experience that I've gotten. Some good, some bad. Working for good bosses, bad bosses and learning lessons that I've learned.

And then for them, I want to provide the same thing. I want to provide the goods I've learned so that I can inspire them to work. So two things. How do we keep people engaged? One: I think if they can associate themselves, recognizing that the job is bigger than just writing code, like it means something and it has a purpose. I'm trying to understand the ‘so what’ of them. The second thing is that if I know where they are right now and if I understand where they want to be, how can I put them into the place for them to be successful? So create opportunities for them. Those two things have helped me engage my team members. How does a person who has a PhD, and probably from a very good school, listen to a guy who doesn't have the same credentials, but then believe that what he says makes sense?

DAVE COLE

You touched on a bunch of great stuff there, Vikram, but I think there's a fallacy that you don't need your boss to be able to do your job. They don't have to be smarter than you or a better data scientist than you. What they do have to do, I think, is listen to you, understand you and help unblock things or help you go down that path to meeting your career goals. I think the nature of your business, working in healthcare, indicates that there's a larger purpose that a lot of different industries don't necessarily have. It's just not the same.

It’s great that you have a team and you're motivated. Everyone should be motivated to improve the lives of others. Then there is the more general point of just listening and understanding that each team member has different goals. Some want to lead teams and be managers. Others just want to be better data scientists, experts in deep learning or a particular type of model. Helping them get there will help keep them fulfilled and energized. So all really good pieces of advice there.

Vikram, I have thoroughly enjoyed learning about healthcare and how you all are helping to improve the lives of your members at Anthem. I'm just also hearing some of your tips around bias and trying to eliminate and reduce it in the process. I've thoroughly enjoyed it. Thank you so much for being on the Data Science Leaders podcast.

VIKRAM BANDUGULA

Thank you very much, Dave. Appreciate this. Appreciate your time.

DAVE COLE

Thank you.