Bringing People Alive Through Machine Learning

Connect with Ryan:

Connect with Dwight:

• 
• dwight@turetskyconsulting.com 

Resources:

• Google Duplex
• HR Data Labs Season 2 Episode 19
• HR Data Labs Season 2 Episode 17

• “Understand the process that created the data that you’re going to be using. You get a better product, and you also don’t lose the humanity in the numbers.”
• “When we start modeling things like turnover rate and we’re able to use explainable AI to see why it’s coming up with the results, we’re really giving those people a voice. We’re being able to ask an entire population of employees, or patients, or whoever, what do you think about this?”

Announcer: 0:02

Here’s an experiment for you. Take passionate experts in human resource technology. Invite cross industry experts from inside and outside HR. Mix in what’s happening in people analytics today. Give them the technology to connect, hit record, core their discussions into a beaker, mix thoroughly. And voila, you get the HR data labs Podcast, where we explore the impact of data and analytics to your business. We may get passionate and even irreverent, but count on each episode challenging and enhancing your understanding of the way people data can be used to solve real world problems. Now, here’s your host, David Turetsky.

Dwight Brown: 0:46

Hello, and welcome to the HR data labs Podcast. I am Dwight Brown, your host for today. And today I am joined by a very special guest, Ryan Allen, from the UnitedHealth Group. Hey, Ryan.

Ryan Allen: 0:59

Hey, how you doing? Right?

Dwight Brown: 1:01

Good. Great to have you here. appreciate you spending the time with us.

Ryan Allen: 1:04

Yeah, me, I’m happy to be here.

Dwight Brown: 1:06

To give you a little background on Ryan, Ryan and I actually worked together at the Mayo Clinic. And he’s a good friend of mine. And he does a lot with machine learning mostly in the healthcare space. And so there’s a lot of crossover though, on the people side with people analytics, because it all deals with people. And so I thought it’d be great to have Ryan here, just to give you a little background of who he is. He’s the senior principal data scientist in the Office of the Chief Data Scientist at United Health Group. That’s a mouthful right? title. And there, he helps make transformational technologies, foundational creates AI systems that change how healthcare operates. And he’s also an evangelist for machine learning. And he’s part of the team that’s setting the vision for the future of AI at an enterprise and industry scale. He’s worked in machine learning and the healthcare industry for about 10 years. And like I mentioned, at the beginning of his career in machine learning, he worked with yours, truly. So it’s great to have you here, Ryan.

Ryan Allen: 2:12

Yeah, it’s nice to talk to you again.

Dwight Brown: 2:14

You too. So one of the things that we do with all of our podcast is we have a fun fact about every guest. And so Ryan, you used to work in a game shop in Dublin, Ireland.

Ryan Allen: 2:29

Yeah, Yeah, I did. So I went to grad school over in England, in Sheffield, and got a job in Dublin right afterwards. And you know, before that, I’d really enjoyed playing games, Dungeons and Dragons, and then all the board games and things like that. And I was walking back from work one day, and I came across a game shop, there was like a block from my apartment. So I went in, and I eventually started making friends with people that that, you know, play there. And then that worked there. And I got to know the owner really well. And he offered me a weekend job. And I realized that if I worked there, I could get a discount on the things that I was already going to buy. So it was really hard to pass that up.

Dwight Brown: 3:12

And you got pub money in the meantime, too, didn’t you?

Ryan Allen: 3:14

I did. There’s never enough pub money.

Dwight Brown: 3:19

Isn’t that the truth? And that never changes throughout life either.

Ryan Allen: 3:22

I’d trade pub money for sleep at this point. But yeah,

Dwight Brown: 3:27

Exactly. If only we could put a value on that. So the topic for today’s podcast is bringing people alive with machine learning and just a sort of level set for people. Just from a nomenclature perspective, we were going to talk about machine learning, but there’s some interchangeability with the terms of artificial intelligence or AI, as well as a data science. Ryan, can you just expand on that just a little bit so that our listeners kind of have a level set of what we’re talking about today?

Ryan Allen: 3:59

Yeah, yeah. So I tend to use AI and machine learning fairly interchangeably. But in reality, machine learning is really a subset of artificial intelligence in a way that an algorithm learns over time. And it learns from historic data, whereas artificial intelligence could even include things like localization of robots based on geometry and things like that. So well, AI is actually the broader term and ml is a subset. I tend to kind of use them interchangeably on a day to day basis.

Dwight Brown: 4:32

Good. Yeah, that at least helps us get a level set as we’re talking about some of the terms throughout the day. So this brings us to our first topic, which is cool people related trends in machine learning. So, Ryan, you’ve worked with a lot of people related machine learning you work with patients and pay In the context of medical care, what are some of the cool trends you’re seeing with machine learning that deals with people?

Ryan Allen: 5:07

Yeah, so I think my two, I guess, favorite, right, the two front runners, for me, are the advances in natural language processing and how that allows people to interact with technology, as we’re more and more tied into technology, computers, bones, Ai, and ml is everywhere within that. And the second one really comes down to focusing on the, I guess, the ethics and the bias within machine learning and how it affects people, right. So the first one is how we interact. So for instance, if you were to call into a call center, right now, the current process is not say I call and I want to know about vision benefits, and my insurance say, what would happen is, I would say, whatever, I’m wondering what my benefits are, how much do I have left the deductibles, blah, blah, blah. And what would happen is the agent would actually have to sort of put you on hold, and then they go on, they type in and they basically do an internal Google search, right, and it pops everything they need. But there’s a break in that conversation right between you and the agent on the other side. But with a lot of the new advancements in natural language processing and natural language, understanding, the capability is there now to have really like a machine learning agent in between the two. So I’ll be talking to you, right, as a member and an agent. And in between would be a machine learning system that would be listening, and actually able to answer the questions, right, so I’m talking about vision, and benefits, and would pick all this up. And instead of you having to put me on hold and break that dialogue, we actually bring up the information as needed on a screen. And then we can keep having like an organic human to human connection, right, there’s not this brain transaction, things like that. Or even something like Google Duplex, which is way out there. In my mind, it’s fantastic in that, you can say the magic words that makes your phone listen to that I won’t hit.

Dwight Brown: 7:09

Everyone’s phone’s going off right now.

Ryan Allen: 7:11

Exactly. But you can you can say that and then say, I would like to have a reservation at my favorite restaurant for four people on Tuesday night, they will look at your Google Calendar. And know that while you have a meeting until five, so that’s B after five and whatever else you have on there. And they’ll actually call the restaurant and use actual human nomenclature when it’s talking. So it’ll use specific slang, it sounds like an actual human. And it can actually pick up what the person on the other side is saying as well, regardless of dialect accent, how fast they speak, what slang they use. And it’ll actually set up that reservation for you by calling a human being and doing that, which I personally think is fantastic. Because I have like a mini panic attack every time I have to call for a haircut or whatever,

Dwight Brown: 7:59

you know, am I gonna goof it up somehow?

Ryan Allen: 8:02

Exactly. Am I gonna slur my words, and then they’re gonna think I’m crazy, you know? So I mean, that kind of thing, I think is huge.

Dwight Brown: 8:10

And I think of how far we’ve come with that, because you talked about regardless of dialect, and the ability to use slang, I think, early iterations of this kind of thing. There had to be perfection and what you said and how you said it. And the systems weren’t able to process different dialects or accents, or whatever that might be where what I hear you saying is, it takes into account all the imperfections of our speaking in our language, and it’s able to process those and still come up with accurate results. And then very timely results, which kind of creates the flow of conversation. Is that would that be accurate?

Ryan Allen: 8:51

Yeah, yeah. It’s fascinating. I don’t know if you ever tried text on your phone by just speaking to it. But I frankly, gave up on it. Because it took me more time to go back and correct. It misunderstanding a word or saying that it was I just typed it out. And now right point where you can call on somebody can have a really thick accent, right? Let’s say you’re calling an Indian restaurant, right? or whatever it is. Yeah, they have a very thick accent, it’ll still pick it up. It’ll still understand. It’s like agnostic, really,

Dwight Brown: 9:21

Man. And if you think about the application of that in the healthcare arena, to where we can start to have these interactions, we’re not just calling a call center, we can have these sorts of medical interactions where we’re working, it’s what kind of like going to see the doctor, and that’s what you’re doing, but it’s machine based, and then helps not have to set appointments and those sorts of things.

Ryan Allen: 9:44

Yeah, yeah. And I think with this push towards telehealth as well COVID has kind of opened up this idea of telehealth even further than it already was prior to the pandemic. But imagine if you could have some sort of machine learning in between you In the doctor in this telehealth visit, that’s helping a doctor with cues that maybe they couldn’t get otherwise, because it’s not in person. I mean, they can’t necessarily see things as well in different aspects that you would get in an office visit that you might be lacking in telehealth, right, the potential to be able to elevate that and bring things to the doctor’s attention that maybe they would miss. In a normal telehealth situation, I think I’d really push things forward, I think right, you know, rural, rural populations where they might not have the same access, or whatever it happens to be really help out with that.

Dwight Brown: 10:42

So that brings us to our second topic, which is the thought of losing people element in the numbers. Sometimes it feels like we lose people in the numbers we get. So heads down and crunching, crunching numbers going through spreadsheets, running algorithms and those things. Sometimes it feels like the people element gets lost. Would you agree with that? And what are some of the things that you do to prevent that in your work?

Ryan Allen: 11:07

Yeah, so I know for me personally, that was, while it is an issue that I run into, like you said, to get heads down and zeros and ones, you’re really kind of looking at an interesting problem. But you forget what the interesting problem really means. So at Mayo Clinic, when we’re working together, you know, I get that feeling. And what I would do is I would actually just walk the block or sale over to the patient care of buildings, Gonda, Mayo, and all of them downtown there, yeah, we’re lucky that we’re only a block away. So I could walk over there, or the lunch hour and really see the patients, they were no longer numbers, they’re no longer rows in the database. Now, it was people who were most likely having one of the most difficult days of their life, because we’re talking about clinic, it’s not, they’re not there, because they’re feeling great.

Dwight Brown: 11:59

Yeah, you’re not having a great day when you have to go to Mayo Clinic

Ryan Allen: 12:01

No. So so it really helped to bring that back into focus this sort of create that empathy in the data, I guess, no, in my current role, or for people who don’t have that same access, to be able to go and see who they’re really working for. At the end of the day, what I found is to remember that to be a good analyst, a good data scientists, the good machine learning practitioner, whatever it is, you have to understand what the process is, that’s creating the data you’re using. And in a lot of the work that I do in healthcare, any, if you’re doing any sort of people focused work, that process is generated by people. And so it kind of helps to bring it back together. And to be able to say, Yes, I’m looking at rows of data, and I’m creating these complex algorithms or whatever it is. But the process creating this as people and the outcome from this has a direct impact on our life. And to try to keep that in mind through the entire cycle of analysis, development, and eventual production integration as well.

Dwight Brown: 13:04

I think of of every project that we work on, and our process of developing those projects, and whatnot. And one of the key aspects of any project is good requirements gathering at the beginning. And the key piece to that is really sort of filtering things down to what’s the goal of this project. And I would imagine that becomes very important in the people related side, because, ultimately, and this is what I hear you saying is that there are people behind everything that you work on, and you need to continually be able to go back and remember that. So as you’re building your algorithms, and you’re crunching the numbers, you still have in your mind, why am I doing this, it’s as opposed to a project that just gets shoved across your desk, and somebody says, I want you to run this number and that number, but you really don’t know why. And so I would imagine that for you. There’s a lot of importance in the process of coming up with these projects and sort of filtering down to that one common goal and what are the requirements to get there?

Ryan Allen: 14:16

Yeah, yeah, exactly. And I think even beyond trying to keep the the idea of people in the data alive, simply because it’s the right thing to do. I also think you end up with better outcomes, right? It’s all part of the same thing, if you can, like you said, How really scope it out, scope out the project, here’s what we’re going for, and then understand the process that created the data that you’re going to use. It’s kind of a win win, right? You get a better product, right? Better project and you also don’t lose the humanity in the numbers.

Dwight Brown: 14:47

Right. Yeah, I think that this also translates well over into the HR analytics arena where ultimately we’re dealing with people from the HR side, they’re no matter what we’re doing. Whether It’s looking at turnover rates, or if it’s looking at competencies or performance management or recruiting, they’re always people behind things. But it’s so easy to get lost in in not realizing that as you’re going through things. So

Ryan Allen: 15:15

yeah, yeah. And I think too, for instance, with recruiting, and I’d mentioned earlier the idea of ethics and bias within machine learning. And part of that comes down to remembering that it’s people as well. So some of the bias that comes into things like machine learning, you know, let’s say that you have a algorithm that helps you with recruitment. And it’s just learning from historic data. And if historic data is biased, right, we mostly hire white men that come from an Ivy League school, whatever happens to be now your hiring process, because you’ve implemented this algorithm is also going to be biased, because you forgot, right? Where the data came from, that there’s people in this data, including the people that generated the data that correlated it, and put it together. So I think it’s important, even from that standpoint, from a hiring standpoint, you’re going to hire better people, if you realize where your data came from. And you keep the people in that. And Yep, and the entire population of people, not even just the subset that you might potentially be looking at, you can have a much better outcome in that respect as well.

Dwight Brown: 16:24

Oh, definitely, we’ve talked on the show before about AI in the recruitment process, and you just touched on this one of the one of the examples is a frustration that I think a lot of people have encountered, and that is the actual resume submission process. And the algorithms are built to pick up particular keywords. And if you don’t, if you don’t know that, if you haven’t gotten those keywords into your resume, when you submit it, oftentimes, the algorithm will filter it out. And right away, you get a rejection notice, where you might be a very skilled person, but you just haven’t been able to learn how the AI works, and how you need to tweak your resume to that. And we’ve even talked about the idea of the hacks that people build and being able to hack the system, by one person I talked to, you said that I put, I took all the keywords out of the job description, put it in very small white font on my resume, that was not, you couldn’t see it visually, but the machine picked it up. And so it would put my resume to the top is what this person was telling me. And I think that’s a bad is one of the downsides. Because again, you’ve got people that are behind this, and you’ve got people with good skills that you want to be able to get into the workforce. And so I would imagine it’s a challenge for you. Being able to keep that in mind, especially building whoever build these AI algorithms for the recruiting process, trying to remember that and trying to account for that, take the biases and things out of the system, like you’ve talked about. Yeah,

Ryan Allen: 18:05

yeah. And sometimes the bias too, isn’t even in the data itself, which is kind of the first thing you look at is the data I’m looking at for hiring or whatever it is representative of the population. But the right study that was done in conjunction between a postdoc at MIT, and I believe Stanford, where they’re looking at facial recognition software, you know, using deep neural networks, scenes. And yeah, you know, the first thing they found was that the datasets that were being used to train these commercial systems were highly biased, who is something like 77%? white males, right, which is not representative translation at all. So what they did is they actually went and created their own sort of representative sample right here is sort of probabilistically representative of the population that we’d be looking at. And they ran those through the same architectures and tested the results. And what they found is that even though they created an unbiased dataset, the algorithms themselves and become biased, so you know, mathematically biased anyway. And that, if it was trying to predict the identity of a light skinned male, right, that is right now, okay. It had an error rate of less than 1% is like point 8%, some crazy successful modeling. But the issue was, is that when it when it was trying to identify, and the worst case ended up being darker skinned females, the error rate was anywhere from 24% to 30%. Which, wow, which is crazy. And it’s not because of the bias data, they corrected for that. It was just some way that the modeling had worked out that it was having issues with that. And you know, and it’s now it’s outside anyway, kind of looks like Well, that’s an interesting scientific fact. And let’s write a paper and whatever but then you think about how these are being used, right? criminal justice or whatever happens to be and now We’re misclassifying people based on gender and skin color. Right on an AI model, right? So again, we have to remember that there’s people involved in this. It’s not just an error rate like this is potentially going to affect somebody’s life negatively or positively, depending. But it’s something that we really can’t take out the entire process.

Dwight Brown: 20:22

Yeah, we talked about DE&I quite a bit. And that really helps to underscore what we talk about when we’re talking about that particular topic. Because how do we become more diverse? How, where does it? How do we manage that, and what you just pointed out is sort of that starting point, if we look at it just from the recruiting process, if our if we’ve got a faulty AI algorithm in there, then it’s not going to get us where we need to be. And it’s not going to get us the people that we need. And so a lot of implications that go with that.

Announcer: 20:59

Like what you hear so far, make sure you never miss a show by clicking the subscribe button. Now. This podcast is made possible by Turetsky Consulting and listeners like you. Thank you for your support. Now, back to the show.

Dwight Brown: 21:14

So that brings us to our third topic, which is in our last topic, we talked about just keeping the people element in machine learning. But more than keeping the people element in there. How do we actually bring people alive? It’s not just about keeping them considered, what are some of the ways that you’ve done that or still do that?

Ryan Allen: 21:36

Yeah. So it occurred to me when we were speaking earlier, that in my prior sort of career life when I was working as a paleo economist, and an arco botanist, so I used to look at plant remains from it’s a mouthful.

Dwight Brown: 21:50

Yeah, that’s a mouthful. I don’t think I don’t think I could say those, you must have practiced that

Ryan Allen: 21:55

I had to get a master’s degree just to be able to say it to be fair. But I used to look at plant remains from archaeological sites. And, you know, at that time, we call them statistical methods to be able to identify variations in these plant remains. But now we would call it unsupervised machine learning, we’re using things like DB scan, and K means and all these sorts of distance based metrics. So even way, way back in the day, part of my role was to sort of figuratively bring people alive, using machine learning, trying to figure out how they’re dealing with their crops, and what was their sort of economic spread and things like that more recently, what I think has been really interesting is this idea of explainable AI. So as technology has advanced, we’ve gotten more and more powerful machine learning models. Now we’re using deep neural networks for pretty much everything. And they’re really good at predicting things. But they’re really bad at telling you why they predicted what they did. So like Pinterest, for example, I’ll stick with an HR for trying to predict future churn rate, right, something like that. And we generate this deep neural network to be able to identify in the next two months, we’re gonna have a churn rate of 25%, or whatever happens to be, we don’t know why we just know what’s going to happen. And now there’s, there’s no people in that it’s just numbers coming in one side and numbers going out the other side. And what do you do about that? With a lot of the advancements in explainable AI, what it’s allowing us to do is to look inside and say, Well, why are people leaving, and it’s really giving people a voice, right? That wouldn’t necessarily have a voice, I mean, I I write tend to not be the kind of person who would go knock in my boss’s door and say, this job stinks. Because of X, Y, and Z. Right? So right, but when we start modeling, things like overturn rate or whatever it is, and we’re able to use explainable AI to see why it’s coming up the results are really giving those people a voice, right, we’re being able to ask an entire population of employees or patients or whoever, what do you think about this? What are you doing? And what are where do you want to go?

Dwight Brown: 24:14

I think of that, from the perspective, almost every analytics process that I’ve ever worked on that element is, number one, the biggest time consumer number two, very prone to error, because, as you said, you come up with these numbers. So let’s say turnover rate, we’ve had a, we’ve had a spike in our turnover rate in the last year. And so you go to the board or whoever you’re presenting the statistics to and they say that’s nice. Now, why is it happening? And so with explainable AI, then you’re basically taking that full process where you identify, and you can explain at the same time, I mean, I Think of just the time savings that goes with that. Because every time we report a statistic, we always try to dig down and figure out why whatever is happening is actually happening. And so what I hear you saying is, a lot of that is taken care of in the algorithms and kind of short circuits, a lot of this a lot of these manual processes that we have. Yeah, that accurate assessment? Yeah,

Ryan Allen: 25:24

yeah. And it really helps in a business context. What levers do I need to pull? What buttons do I push to have any effect on any of this? even beyond giving people a voice? Like you said it also you go to the board? Now they can say, Well, here’s what we need to do. Or it’s because of PTO number. So can we have more PTO, or people want to work remote, or they want to have more connection at work and have it be less transactional? And more relational, that sort of idea. And they can do something with that vs. People are going to leave? Okay, now what?

Dwight Brown: 25:57

Right, exactly? Yeah, it gives us that context that we need to truly understand what’s coming out in the report. So we’ve talked about a lot today. And we’ve talked about bringing the people through and machine learning, we’ve talked about some of the cool trends and in machine learning, including things like explainable AI, and other human centered algorithmic elements that are out there, as well as some of the pitfalls of some of those, we always need to keep the people in the numbers because we’re dealing with people and we are a people business. And so we talked about the fact that AI will never replace that you can’t lose that. But the goal is to really enhance the our understanding and our learning as we go along. And truly bring the people out and that and truly bring the people alive. So Ryan, I appreciate your being here. A lot of fascinating stuff that we’ve been through any parting thoughts that you have for us?

Ryan Allen: 27:10

No, I just really appreciate getting down to and being able to talk to you again.

Dwight Brown: 27:16

It’s good to be able to talk again, I always love talking with you. I always loved working with you and when we work together, so thank you for listening. And if you liked this episode, please hit the subscribe button. And if you know somebody who might find useful this episode, please do forward it to them. And if you have any comments or questions, please leave it to us on your favorite social media platform or go to Turetsky Consulting comm slash podcast and let us know your thoughts. Thank you very much. Take care and stay safe.

Announcer: 27:47

That was HR data labs. Please visit Turetsky consulting.com forward slash podcast to review the show. add comments about this episode, or add new ideas about upcoming shows you’d like to hear. Feel free to be creative. But please be nice. Thank you for joining us this week on the HR data labs podcast and stay tuned for our next episode. Stay safe