← Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
How Data Science Competitions Are Shaping Mental Health Research
00:35:10
Gregory Kiar and Arianna Zuanazzi talk data science, sleep, and how a Kaggle competition is reshaping mental health research at the Child Mind Institute.
Loading summary
Transcript36 lines
- [00:05]
Logan
Welcome to Reshaping Workflows with dell Pro Max PCs and Nvidia, where innovation meets real world impact in high performance computing.
- [00:20]
Greg
Welcome back to another episode of Expanding Workflows with Dell Pro Max and Nvidia. I'm Logan, your host. This episode's gonna be a great one because up until this point we've talked a lot about Nvidia, we've talked a lot about delpro Max and as you know, within kind of the Dell Pro Max, you know, the old workstation, you know, traditional industries, there was a lot of markets that were served that are ran by different ISVs, those being, you know, for example, like healthcare, media, entertainment, engineering, aco. The one we're talking about today is data science, which I don't think gets a lot of. It's probably just due because that's kind of the building block of everything, you know, AI and predictive wise it's been going on kind of in the world in the last 50 years. Now with that, we're going to take a different spin today on that. We're not going to talk necessarily pure data science and the nuts and bolts of, you know, how do you build a data science workflow? What we're going to talk about is we recently partnered and wrapped up a Kaggle competition with our friends at Childmind going over problematic Internet data. So we're going to kind of recap that conversation, but also really focus on once you've done the analysis, once you have the results back and you've kind of, you've done the things you need to do to get to an answer, how do you implement them? What did you observe? How did you arrive at that answer? So we're going to get in kind of more of the analysis phase today of data science. But before I get ahead of myself and out in front of my skis like I normally do, we've got two great people here to join me. So Greg and Ariana, take a second, we'll start with Greg. Just 10, 15 seconds, introduce yourself, give the listeners a little bit of background.
- [01:57]
Ariana
Sure. Hey, thanks for having us, Logan. It's great to be here. Yeah. I'm Greg Kerr. I'm a research scientist at the Child Mind Institute and director of the center for Data Analytics, Innovation and Rigor. And so our group, it's a lot of words. What we effectively do is we try, you know, build data science tools and principles and philosophies for how we handle messy brain imaging data, mental health data, quantitative psychiatry data, these sorts of things. And we're trying to bring it into a place where we can actually generate meaningful insights from these sorts of data and apply it to, you know, enhance the way we actually provide care.
- [02:30]
Logan
Yeah. Hi. Hi everyone. So my name is Ariana Zwonazzi and I am the project collaboration specialist in the research department at Chinman Institute there. I am really focused on leading open science and open data initiatives as well as facilitating collaborations among researchers internally and externally. Internally and among this. There is a very exciting thing that I do which is organizing data science competitions which I'm really excited to talk about today.
- [02:57]
Greg
That's amazing. So let's jump right into it. We've got about 30, 35 minutes. Let's jump right into it. So I'll kind of tee it up, kind of a simple one. Why does Childmind Institute run data science competitions? You know, why go out to kaggles of the world and run a competition? Why do you all do that?
- [03:16]
Logan
I'm going to answer that, Logan. And before answering this question, I really want to give some context about the role that the Child Mind Institute plays in advancing open neuroscience. So bear with me. It might be a long introduction.
- [03:30]
Greg
Please. No, I mean, we've had Greg do it, but let's do it again. Go, go, go.
- [03:36]
Ariana
I'm sure Ariana will do it better.
- [03:38]
Logan
In 2010, Dr. Mai Milam, who is the Chief Science Officer at CMI, launched the Functional Connectome Project, or FCP, which is an initiative which goal is to improve our understanding of brain connectivity by doing something very cool which is sharing, openly sharing resting state FMRI data. So making this data openly available to everyone. By sharing data, researchers were really able to work, and are still able to work with much larger data sets, which means that they can really, they are really able to address questions they would not be able to address with their smaller research data set. This really helps speed up the pro the progress in neuroscience and in, in particular in, in looking at brain connectivity. And just to give you an idea, since 2010 the Tianman Institute has shared resting state data from over 30 sites across the world. So this is really a worldwide initiative and a worldwide effort. We are now in 2025. So in these 15 years there was another initiative called the International Neuroimaging Data Sharing Initiative, or indi, which aimed at expanding beyond rest in state FMRI and it really aimed to share a wider range of neuroscience data globally. And nowadays this initiative really includes more than 50 openly available data sets that cover a variety of data types, cover a variety of participant demographics, as well as sharing both human and non Human primate data. All of these data sets were really thought with neuroscience and generally researchers in mind from academia. But if you think about that, this is a lot of data and it's quite exciting because it's a rich amount of data as well. It's not one type, it's like diverse data. And so it's really interesting for other people that outside academia, outside research to be able to use this data. So to go back to your question about why we want to run data science competition, we were really curious to, and we're really excited to share this data beyond academia and we wanted to get people excited about data science in the mental health space. At the same time, we were really interested in hearing and thinking about our own data in a different way. So we were interested in involving people from different industry in different backgrounds to work with this data and to be able to see what they would do with this data, how they would approach the problem, how would they solve the challenge that we posed to them. So that's, I think the main drive for us to really get, get people interested in mental health research as well as getting new ideas and new ways to look at the data itself.
- [06:15]
Greg
Yeah, I think that's, you know, it's an interesting answer and it's kind of where I kind of thought you would go with it. I do have a question and I, you know, I run into this in my day to day life as well is, you know, you're working on a project. You know, I work on kind of AI, go to market solutions, alliances, Dells for Dell Pro Max. So I've got kind of a very narrow window like that. I kind of see the world through. How much do you think it helps having, you know, obviously people have to be data scientists, duh, you know, check the box there. But then how much does the outside industry perspective you think help someone? And what are some other industries specifically in you know, mental health data or you know, kind of behavioral data, other industry perspectives that someone may have that would, you know, lend itself well to them being able to understand the data and bring something different to it.
- [07:05]
Logan
Yeah, that's a great point. I think that, you know, having a different background and way to look at the same thing, I think it really helps like coming with new perspective and new ideas. So in general, I think this applies to many other things when there are people from different backgrounds and different perspective perspectives that work on the same thing and can also talk to each other. This really is interesting because different questions come up, different way to approach the same question, come up and this was really evident in the discussion. Discussion board was amazing and I was really interested and excited to read what people were coming up with. And you know, it's a way, it's a very unique space, I think the Kaggle competition for people that usually don't work together to be able also to talk to each other and exchange ideas. So I think that, you know, what you were asking is whether, you know, there is anything that other industries can bring to mental health research. And the answer is definitely. So, you know, like thinking of a problem that can be a real world problem is like mental health is really a real world problem that a lot of people are working on needs different kind of perspectives and the industry perspective can be interested, for instance, for developing new, you know, new project and developing new products as well for that address this specific question that we were answering that we were, sorry, asking in this competition as well as, you know, if you have a clinical background, you can look at the problem from that perspective and bring information and solve the challenge from a perspective that can be helpful to other clinicians. So I don't know if this answers your questions, but maybe Greg, want to add to that?
- [08:41]
Ariana
Yeah, I was going to say one thing that I would add even more than that and like Ariana's hit the nail on the head, I think. But one thing like in terms of specific industries, you know, one of the nice things with this data is it is multimodal. There are different types. You know, we have risk torn accelerometry data used in this competition. We have a bunch of survey data used in this competition. We have a bunch of biometric data used in this competition. And so, you know, these are formatted in different sorts of ways. So for instance, tabular data is extremely common in some domains of research. You know, similarly time series data. So you have people coming from economics, from natural language processing, from all other domains, from physics, who, who are able to kind of apply the techniques of their niche to our data. And I think that's one of the biggest things that we've noticed. And this is a problem that academia has suffered from for ever. As far as I can tell is the number of times the same or effectively the same approach has been re, quote, unquote invented in a new domain. Just because the two disciplines of science weren't communicating effectively is too many to count. And so one of the really nice things here is it gives us an opportunity to try and shortchange that process a little bit. We don't need to go on reinventing wheels. We just need to see what wheels on economics side of the table work over here.
- [09:53]
Greg
I love that answer. I mean, that's, it's very interesting. I mean, different industries kind of run at different paces and you know, over the course of time they're going to catch up. It's very interesting. So the reason, one of the reasons I asked that question is that, you know, when it comes to kind of data science is that people get very spun up in having a, a vertical focus. Like, you know, like, oh, I work in healthcare, I work in banking, I work in finance. But with data science in general, I asked the question because I want people to know is they can bounce around to different domains. Long as you have the underlying kind of base skill set of data science, you can pop anywhere in and out with that. But speaking of data, you know, we've talked a little bit and we've teased out, you know, the problematic Internet use and how the multimodal data that we used in the, in the Kaggle competition. But tell us a little bit more about the data. What was, and I guess big question is what was the question or the big questions that you were asking in the data that you were asking everyone who participated in the competition to go answer?
- [10:48]
Logan
I'm going to take this question just because I'm always very excited to talk about the Healthy Brain Network. So I was mentioning before that there was an initiative that is called indie so International Neuroimaging Data Sharing Initiative. As part of this initiative, the Childman Institute in 2015 launched a what is called the Healthy Brain Network, which is a initiative itself. So it's an initiative within an initiative, but it's like a community referral program basically where families who are concerned about the mental health in their children, they are encouraged to participate and they basically receive no cost diagnostic evaluation. So basically this works very well for research as well for the community itself. So I'm very excited about this because of these two aspects. It's very unique initiative and we are always interested in both how can this data help research? But as well, how can this data help communities and families. All of this data is openly available, is openly released, which is also another aspect of the Healthy Brain Network which makes it unique. And so far we have released over 4,000 data sets from participants, mostly from New York City. And another very interesting aspect of this dataset is that it covers a wide variety of data types. So Greg was mentioning that before and I feel like this is really a staple of the dataset that we have, the Timeline Institute So it's a multimodal dataset. So it includes kinetic restaurants, imaging data, electroencephalography data, as well as phenotypic data. And this is the dataset that we use for the problematic Internet use competition. So more broadly, we know that problematic Internet use among children and adolescents as well is an increasing concern. And understanding it better could really improve the way we address mental health issues such as depression and anxiety, and issues related to the use of Internet as well. So this competition was really driven by this real world problem. I was saying before, we really like are interested in giving competition participants real world problems and real world data as opposed to artificial data or made up questions. We are really interested in getting a solution that can be used then in the future. And the participants of the competition were tasked to predict the level of problematic Internet use in children and adolescents based on physical measures. So I was saying before that one of the reasons why we are interested in running data science competition is because we have a rich data set. And this is the case for this competition as well. So in this specific competition we had a variety of data types. We have demographic information, we have Internet habits, we have general functioning data, as well as data from six different physical AS assessments and questionnaires, such as, for instance, cardiovascular fitness, aerobic capacity, muscular strength, sleep difficulties, and so on. We also incorporated something that we were very excited about, which is the real time physical activity measures that was measurement of physical activity through an accelerometer that participants work for 30 days. And all of these measures altogether were used to predict Internet addiction. Internet addiction was measured through a questionnaire that parents had to complete. And this questionnaire was about the Internet habits or, you know, Internet issues that, or use of Internet use going online that their children were experiencing. And they were asked questions such as, how often does your child choose, for instance, to spend more time online as opposed to with their friends. So one last point that I want to make is that all of these measurements and measures that we shared with the competition participants are easy to administer and are accessible. I think this also goes back to your question Logan before, about why we are really interested in involving more people, given that these measures are accessible and, you know, easy to administer. The solution and takeaways that we get from people from different industries as well are interesting for them too, are interesting for, you know, clinical practitioners are interested for people that work in other sectors. So these are not like niche measures of academia or of a specific type of research, but are really interesting for people that work from in different fields as well.
- [15:02]
Greg
Well, this kind of leads kind of really into my next question is that I loved how, you know, and I'm not going to use data science terms, I'm not a data scientist, but how robust, unique and broad the data set was where, you know, I followed along in the Kaggle competition. A lot of it was like Greek to me obviously. But you know, as I was reading it, like some of the time series data and all of this, it was just very interesting where, you know, the accelerometer. Right. I remember reading a thread about, you know, one of the kids, some of the kids didn't wear them at all. Some of them took it off like for days at a time. Some kept them on the entire time. So it was just interesting to see how the data kind of impacted that. And I think that leads to. My next question is, and this is for Greg, is that, you know, kind of through the competition, you know, we, we stayed connected the entire time. We kind of watched it unfold. It ran for a little bit of two months, but there was obviously a very big shake up overall in the competition that, I mean, I don't know, did it get people like wound up? Probably not, but in from what I had experienced and have seen in my limited window of Kaggle seemed like a pretty big deal. So why don't we tell everyone a little bit about that and why that occurred?
- [16:11]
Ariana
Yeah, totally. No, it's a great question. And so just for context, for anybody who isn't familiar with the term shakeup, a shakeup is basically when you have a big difference between the public leaderboard which competitors are using to kind of guess or gauge how well they're performing throughout the competition, and the eventual private leaderboard which is released only after the competition closes. And so as the competition hosts, we see both throughout the entire time and again, they're just, it's the same test that's being performed, it's just on separate data so that essentially nobody can really, by making more and more and more submissions overfit to the private score because they can't see it. So yeah. And as our competition progressed, as you mentioned, Logan, you know, we, we really saw this developing and in particular the measure we were using for our leaderboard was something called quadratic weighted kappa. This is effectively a measure used for classification tasks when you have multiple labels where the difference between them isn't always the same. So by that what I mean is, you know, as Arianna mentioned, we're trying to predict Internet addiction scores in the case of, you know, the one particular Child, there may be no concern, and in another it may be mild, another moderate, another severe. If there's a child who has, you know, severe Internet addiction, and I rank them as having no Internet addiction issues, that should be penalized more than if they have severe and I rank them as moderate. So it's essentially an evaluation metric that lets us gauge, or rather not penalize more depending on how far, how wrong you are. And so by this score, what we noticed was essentially a peak at around 0.46 performance on the public leaderboard. And as soon as people started doing better than that, we actually noticed their performance on the private leaderboard begin to drop. And so this is essentially the source of our shakeup is on the public leaderboard you were still able to climb, but that performance was not generalizing to the private test set. And so there's a number of reasons why this can happen. And, and to be honest, it's actually a really nice illustration of something that's been going on in, in psychology research for a while. Logan, have you heard of the reproducibility crisis in psychology?
- [18:13]
Greg
I'm gonna guess by the name. Is trying to reproduce the same result that you got.
- [18:17]
Ariana
Yeah, basically. And so marketing major for the win, a well named crisis.
- [18:23]
Greg
Exactly.
- [18:25]
Ariana
So yeah, so I, I, I think that it was started to become widely recognized about 15 years ago as well, this reproducibility crisis. You know, there's a seminal paper that essentially noted that of a hundred studies published in psychology, only about 30 of them were able to be reproduced by other researchers in other contexts. And this, you know, came as simultaneously a huge surprise to everyone in the field and no surprise at all. And, and the reason for this is there's so many variables, variables and so many sources of like heterogeneity in all these sorts of data that it's really, really difficult not to on small samples and narrow focused experiments overfit to, to variation in your sample, that isn't actually what you're trying to the signal you're trying to model. And so essentially what we're talking about here is, you know, in our data set we had, you know, we released something in the neighborhood of 4,000 data from 4,000 kids. And in this sample there's going to be different trends, there's going to be different relationships. And so when you split them up into different chunks. So let's say of those 4,000, I don't remember the exact fractions, but let's say we used approximately 3,000 of them for the public training set. And you know about 750 for the private test set. Something in that ballpark. What you end up having is now you have two different populations that have different structures of noise, different part, you know, attributes of the data that we couldn't actually measure. So for instance, things like what did they eat the day that they took a particular assessment, what was the weather, what was the time of the year? And these sorts of things that we don't really know how they affect a lot of these constructs. And it's really hard to document all of them because, I mean, you could also theoretically say, oh, I'm going to answer a questionnaire differently. If I just had a lot of traffic on my way to the clinic, if my train broke down, you know, you're going to answer things differently. So that's an impossible space to properly capture. And what we're trying to capture are ultimately subjective things. And so you end up having this huge amount of heterogeneity. And it's again, when your sample sizes are small, this is a bigger challenge, which is again, part of the reason why we're so excited about these data sharing initiatives. But what we essentially saw was people were doing a great job at modeling the true relationship between these kind of accessible measures of physical fitness and, and problematic Internet use. But then they kept going. And so it was essentially this exercise where people were trying to get more juice to squeeze out of these, out of these data when there wasn't really any. And what that ended up doing was they were overfitting to noise that wasn't shared in the private test set that they didn't have a chance to look.
- [20:50]
Greg
At from not a data scientist. So I get the point of like diminishing returns, right? You can only put so much into it. You start getting less and less until you're getting nothing than a negative return. How if someone, let's say, who, quote, unquote, overfit the data and you know, or it was beyond kind of that, the point of, you know, no return, we'll call it, how did that affect their results? Meaning, like in the real world, had you picked that as your answer, how would that have affected some of the decisions that you might have made based upon the prediction? Like, how would that have impacted. Because it's very interesting to hear is like doing this is obviously leading to decisions that you're going to make and recommendations you're going to make and papers you're going to write and follow up studies. But let's say you pick the wrong variable, you pick the wrong Answer what happens?
- [21:42]
Ariana
That's a fantastic question. And I mean, the shortest, most brutal answer is a bunch of wasted resources. And so essentially, you know, one of the ways we want to use these results is to, as you said, like, guide the next study. If what we want to do is figure out what accessible measures of physical fitness can help us give a less biased estimate of problematic Internet use, where we don't have to go home and ask the parents and where you can just have an accelerometer on the kid, for instance, we can, you know, take their physical measurements whenever they get their annual physicals done and these sorts of things and whatever other metrics, you know, sleep scores and these things, and use them to provide a guide that can hopefully detect things early and regularly. And if what we're steered down the wrong path of somebody who overfit to the data and they found that, you know, it just so happens that I'm just going to pick an arbitrary signal that definitely has no impact. Handedness, whether they're right handed or left handed was extremely predictive in this particular overfit experiment, hypothetically, then what we'd essentially do is roll out an assessment. We're asking everybody about handedness and that's going to ultimately lead to nothing useful. And so it really, it's a tricky problem. And I think this is one of the beauties of the way Kaggle designs the competitions of having these two separated test sets and we keep extra data back as well so we can kind of do even more testing on our own. But it really is trying to minimize the likelihood that we get into this situation of a model that is finding a relationship, but the relationship is among correlated noise as opposed to the true phenomena as it can be observed in these different populations.
- [23:17]
Greg
Yeah, that makes, makes total sense. So really, I mean if you were to boil it down and to say, hey, what's the biggest takeaway? Obviously learned a lot, you know, through the competition, you know, gained a lot of insights. But I really think it boils down to walking away. What I'm hearing is the way that the competition design is now allowing you to go out and ask more informed questions when you're collecting the data, finding out what matters and what doesn't. So you're not spending. And I'm sure it's not cheap to put an accelerometer on someone for 30 days. I'm sure there's a cost associated with that. So all of this. So really the idea is to have better data collection and that ultimately leads to better science. Is that ultimately breaking down what we're doing here, or is there something beyond it that I'm missing?
- [24:03]
Ariana
I think that that's, that's the crux of it. And then one other piece that I'd, I'd add on the end, so, you know, it's more efficient and accessible data collection, which will lead to better science, which will lead to better care. And one of the critical things, again here with the word accessibility we keep throwing around is, you know, one of the most important things in mental health in particular, but all, all areas of health is early detection and intervention. And there aren't very many good indicators for early detection in most areas of mental health. And so the more we can streamline the ways that we measure the data, we can make them more and more accessible. That means we can measure them more often. And if we can measure them more often, we can start developing these indicators of early onset of various problems, whether it's Internet addiction or other things. And then we can actually intervene before problems become too severe, that they're harder to fix. And so that is ultimately like the big, big picture cycle that we're trying to live in. And so we're right now in that phase of refining the assessments to improve the science, but the eventual goal is to translate that to early intervention and care.
- [25:06]
Greg
Makes perfect sense. Man, it's so interesting. I mean, I remember you and I, Greg, when you were in Austin, us talking about it, and it was kind of blowing my mind in terms of how data science worked, all the questions and all this kind of stuff. I mean, it's fascinating. So if you can, I mean, I don't know if you can share. Hopefully you can. So, you know, kind of final question about the, the competition. What were kind of the big predictive factors that, you know, from a health activity standpoint led to problematic Internet use? What were a couple of the big things? If you can share for those that are watching out there, I'm curious to learn myself because my daughter might, may or may not fall into that bucket. I'm gonna be honest with you. I mean, she plays sports and she's athletic, but I mean, that, that YouTube, it's a, it'll, it'll get you. So I'd love to hear what were kind of a couple of the big things takeaways that were very heavily correlated in the prediction.
- [25:59]
Ariana
So there's a few basic demographic ones where of course, this is correlated with age. And so, you know, it'll trend with different ages differently. You know, a six year old will have a different relationship with an Internet than a 16 year old. But I'd say from my recollection of all of pouring through all these models and Ariana, please correct me if I'm misremembering was sleep was essentially the biggest predictor we saw across the board. Everybody who is, who performed well in the task showed us that what was really important in their models was sleep. And so essentially if you're, if you have an unhealthy relationship with the Internet, it affects your sleep. And that's where we're actually measuring a downstream effect. Ariana, were there any other features that really stuck out to you?
- [26:39]
Logan
Yeah, no, I think that the. We had one questionnaire that is called Sleep Disturbance scale that measures different, you know, aspect of sleep and that was surprisingly, and maybe not surprising, surprisingly a very important feature, a very important variable. There were other, yeah, physical measures, basic measure, like for instance, height, weight. This was also predictive. But yeah, what Greg said is basically what came up. And I think it's very interesting also that people, you know, they took different approaches to challenge this problem and they all came up with a similar first, you know, four similar features. So that was something that I was very curious to see and I was very interested also, you know, in hearing what people did to come up with these four specific features that were selected. And if I can add to that, another aspect that was very interesting, Logan, is not just which features were selected, but also which kind of approaches people took in addressing the challenge. Not just from a let's find the best model, but also how do we perform the feature, how do we solve the problem of missing data? So this is like a big conversation that we had. Also internally, our data set has different sources of noise. Again, we are talking about human participants and children. So we have missing data, we have some measurement errors, we have some inconsistencies. And the participant had to be able to work with this type of data and find solutions to that as well as how did they solve problems such as finding, combining different models that could work together. So ensembling was another aspect that we discussed a lot with the winners. Especially I think that also being able to talk to the winners and seeing their solution was also very interesting to us. And quoting Greg, this week, you and I always talk about kaggle competitions, which is true. We are always thinking about what did we get out of that, what did participants teach us about our own data. So that aspect about getting information from the participants, getting their approach and understanding how they solve the problem is useful to us, not just for the data itself, but also it teaches us how can we communicate about our data with different stakeholders. So I'm deviating a bit from your question, but I wanted to highlight this aspect that was really interesting to me.
- [29:02]
Greg
No, I love the deviation. It's good. And it's also kind of a data science question. Deviation, you know, statistics question. No one likes my jokes. That's fine. All right, so we're. We're kind of, you know, up on it. I guess the. We'll. We'll wrap it up with this. And you know, Greg and Arani, you each kind of had to answer this in your own is, we know the competition was, you know, popular. We know that you got a lot of good learnings out of it. This is always a great question that I like to ask people to see how, you know, introspective and retrospective they are. If you could change one thing, knowing what you know now, if you were to go back and redo the exact competition, what would be the thing you would change?
- [29:43]
Ariana
That's a great question, Arianna. Do you want to go first or would you like me to do?
- [29:47]
Logan
I think we might answer this question differently, so I'll go first, but I'm sure it's not going to overlap with your answer. From my perspective, what I would change is providing participants with more information and background. So what I noticed in the discussion as well is that participants were not just interested in discussing data science and machine learning problems or, you know, topics that they were interested in, but they were also very interested in the background. So why is this question interesting? Have other people already investigated this? Is there any research paper or any industry paper or any. Anything that can find that gives me information about why this question is timely and relevant? So we did that, and we did that throughout the competition. We provided participants more information. We, you know, answered a question they were asking in discussion thread. So it was a great teaching. Sorry, a great learning experience for me to understand that people are really curious about the topic and are curious about the mental health space. So providing them with more information, you know, why this is relevant. Here are some interesting papers that talk about that. Here are some interesting, you know, information about what physical measures are and what questionnaires are and how people usually reply to this. I would, I would probably add that from the beginning.
- [31:00]
Ariana
Yeah, I think that's a great point. Indeed. We do have different answers. I went totally like data went very specifically, knowing that people really got a lot of signal out of sleep. I wish we provided more annotations of sleep from the actigraphy data, which we do have, but we, we didn't share in this pre, pre processed vers. The data was actually we were also using them in our previous Kegel competition that we ran last year and I think really trying to see what other ways we can get at some of these constructs. So, you know, even the Internet addiction scale that we use, there's a parent report version, which is what we used here. We could have also included the child one. I'd be curious to see how that impacted things. Would be curious to see how other measures of sleep and all this that, you know, did things. And so essentially it's my question or my answer to your question, Logan, is a non answer. It's what I want to do next.
- [31:49]
Greg
Okay, well then tell me what you want to do. That's a perfect ending. What do we. What, what's happening next?
- [31:54]
Ariana
Yeah, I mean, it's, it's taking what we learned from this and figuring out what's the next assessment we want to deliver. And what, how do we want, what do we want to measure in. In the Healthy Brain Network study and in other studies we run to get at the construct of problematic technology use more precisely and more different avenues. You know, all of these assessments are imperfect. That's one thing that our competitors were not shy about pointing out. And we know as people who work in this space, you know, even thinking about the entire target of our competition, the problematic Internet use questionnaire. It's a questionnaire filled out by the parent about their children's Internet use. So the parent's own relationship with Internet is biasing their answers. Of course it is. It's unavoidable. And so trying to find different ways to probe these same constructs with an emphasis now on. Okay, well, we need to clearly need more ways where we're probing technology use and more ways where we're probing sleep and see which combination of these things gives us kind of the most, the most signal, the most robust relationships that we can bring across these different populations and groups. That's really where my head's at in terms of how we can take this and spin it up into the next thing.
- [33:04]
Greg
I love it. And that's kind of a good way segue because we are at kind of the end here is that when it comes to data science, what I've learned is it's, you know, an iterative process, right? It is. You are constantly iterating, refining, going back, trying to optimize, et cetera, and fun fact for those that didn't know, if you are on a Calga competition, doesn't matter which one and you're running, you know, preferably a Dell Pro Max. If you're not naughty, shame on you. But if you're running anything with an Nvidia RTX gpu, you can go to just Google Nvidia AI workbench, install Workbench and then go to Git and pull their competition kernel project off of Git and you can bring all of your local compute for your Kaggle competition, you know, for all of your, your data sets. Instead of using it kind of in the cloud, if you have, if you're over your allotted amount of hours, you can actually bring that to your local device. And it's something I wish they had when we ran our competition. So it means that we're going to have to run another one. So that's the organis going to have to do that. So with that, you know, I hope everyone enjoyed the episode. As we keep kind of growing and evolving with this podcast, if you see anything or you want to see a certain topic or you want a certain industry cover, industry topic covered, we're happy to do it. The idea is to really show the power of what Dell Pro Max and Nvidia RTX GPUs are bringing, you know, kind of to transforming workflows from media entertainment to engineering to AI to data science, kind of across it all. So with that, you know, appreciate. Arianna, Greg, you joining me today and we'll see you on the next one.
- [34:59]
Ariana
Do what you want. Do what you want.
- [35:02]
Greg
This podcast was produced in partnership with Amaze Media Labs.