00;00;10;21 - 00;00;32;23
Dr Jasmine B. MacDonald
Hi there. And welcome to this episode of Psych Attack. I'm Doctor Dr Jasmine B. MacDonald. Today I'm catching up with Doctor Dr Taylor A. Braund to hear about his research in digital phenotyping for mental health. In particular, we're going to talk about Taylor's research into the link between mental health symptoms and keystroke metadata from smartphones. Welcome, Taylor.

00;00;32;25 - 00;00;34;05
Dr Taylor A. Braund
Thanks for having me, Johns.

00;00;34;07 - 00;00;42;12
Dr Jasmine B. MacDonald
Yeah, it's great to have you here. I always love to kick off episodes with some background. Tell us your origin story. Why psych.

00;00;42;14 - 00;01;13;26
Dr Taylor A. Braund
Last? Zack? I mean, it's pretty long and, twisted journey, but I guess, like, I started in, undergrad doing, psych and then, did on his research in evolutionary psychology. So kind of looking at facial expressions and the difference between Duchenne and non Duchenne smiles. So or genuine non genuine smiles. And then from there I kind of didn't really know what I was going to do once I finished, because I knew I didn't want to be a clinician or get registered or anything like that.

00;01;13;26 - 00;01;48;26
Dr Taylor A. Braund
So I ended up, going to industry and working on, being clinical trials. So two large international clinical trials for a number of years. And then after a while I was like, I think I want to go back and do more research. So, went back inside of the PhD at the West, made chief medical research, and that was looking at treatment predictive biomarkers, for antidepressant treatment outcome in people with depression, specifically people with co-morbid anxiety and depression.

00;01;48;29 - 00;02;24;08
Dr Taylor A. Braund
So finished my PhD again did not plan on what I was going to do after so I ended up, getting a role, with the Black Dog Institute in their workplace mental health group. So they were running a large, cluster city. They, needed, management, operational staff. So I took on that role. And then after about six months in that role, I transitioned into another postdoc looking at digital phenotyping and kind of been working in that area ever since.

00;02;24;08 - 00;02;39;13
Dr Taylor A. Braund
So kind of jumped around a little bit that, it does kind of fit together a little bit in terms of looking at treatment predictive markets, just different data types. And yeah, you find a way to make it fit together.

00;02;39;16 - 00;02;48;18
Dr Jasmine B. MacDonald
nice. a couple of questions of pop to mind for me that I'm wondering, did you enjoy your time working on asked.

00;02;48;20 - 00;03;17;21
Dr Taylor A. Braund
And did I mean, look, they were like international large clinical trials. So you got to travel internationally for it. Got to nice. And when you're younger, like younger, it's like very appealing to, you know, do travel and see the world and all that kind of stuff. As I've gotten older and, during world travel for conferences, I'm a little bit less enthusiastic about it, you know, but, yeah, it was a it was a exciting time.

00;03;17;21 - 00;03;39;07
Dr Taylor A. Braund
And I mean, the research was like super interesting. And all the people working on it were really, like, amazing and smart and, you know, a great kind of networking and gateway into the research that I do now. And I guess at that kind of time, I didn't realize the overlap between the researchers who work in these fields. And then it's it's very diverse.

00;03;39;07 - 00;03;46;25
Dr Taylor A. Braund
So it's you see, a lot of the same researchers I see a lot of the same researchers now. So yeah, just a great experience.

00;03;46;27 - 00;04;02;27
Dr Jasmine B. MacDonald
I've always been interested in, in researching methods quite broadly and being really open minded. But you meet some folks who are on one side of the fence or the other with, citizen and what they can tell us. And I think often the point is missed that it's like, well, actually it's not what method? It's like, what's the question?

00;04;02;27 - 00;04;19;09
Dr Jasmine B. MacDonald
And that leads us to, to which method to use. So I thought I just yeah, that's something you shifted out of. But it sounds like it was a like, transferable skills for what you're doing now as well. It's not it's not completely different.

00;04;19;11 - 00;04;47;20
Dr Taylor A. Braund
Yeah. For sure. And I mean, you know, learning about the operations of arts, it's like kind of, what's happening behind the scenes in terms of like IAB and ethics and data quality and patient participants safety and all these kind of, you know, diverse roles that are applied within kind of large scale assets. And then, you know, seeing that applied in new assets, you know, with different kind of goals.

00;04;47;20 - 00;04;53;06
Dr Taylor A. Braund
But yeah, very like this set of skills that's very inverse transferable.

00;04;53;08 - 00;04;55;25
Dr Jasmine B. MacDonald
Awesome project management skills for sure.

00;04;55;27 - 00;04;57;09
Dr Taylor A. Braund
Yeah for sure. Yeah.

00;04;57;11 - 00;05;16;26
Dr Jasmine B. MacDonald
I want to admit naivety around what Black Dog Institute looks like on the inside. Like I obviously I know about the Institute and I've seen a lot of products that have come out of the institute. But you mentioned you're working in a team that's focusing on occupational context. Right.

00;05;16;29 - 00;05;26;22
Dr Taylor A. Braund
So there's a workplace mental health group with you. Yeah, the Black Dog Institute, and they focus on yeah, pretty much anything to do with workplace mental health.

00;05;26;24 - 00;05;31;16
Dr Jasmine B. MacDonald
Not one of the kind of other major teams that you have a black dog.

00;05;31;18 - 00;06;00;10
Dr Taylor A. Braund
So there's suicide prevention. that's quite a large team. Is the, population, mental health teams. There's the digital scene and typing streams of research so you can get that diverse. there's also a lot of cross project work as well. So you kind of get the experience of working across these diverse research fields, which is amazing having, you know, worked across diverse projects before us.

00;06;00;12 - 00;06;08;28
Dr Jasmine B. MacDonald
And nice. Okay, so we've mentioned the term a couple of times now and digital phenotyping. Tell me about it. What does it mean?

00;06;09;00 - 00;06;35;05
Dr Taylor A. Braund
I guess if you Google digital phenotyping or you read any of like the big papers that have come out, they'll all use the same definition that came out, in 2016. And I'll read for you verbatim. It says digital phenotyping is the moment by moment quantification of the individual human phenotype in situ using data from personal digital devices.

00;06;35;07 - 00;06;37;17
Dr Taylor A. Braund
Okay. So that means that's a challenge.

00;06;37;17 - 00;06;39;18
Dr Jasmine B. MacDonald
On the working memory. But yeah.

00;06;39;20 - 00;07;11;02
Dr Taylor A. Braund
Yeah yeah. So I think like as these fields evolve and also they become a bit more fit for purpose, the way that we operationally define these things. So how I would define digital payments shopping within our projects or within the projects I work on would be, you know, leveraging any digital data that you can collect through a smartphone to make inferences about mood or behavior, to try and improve aspects of clinical care.

00;07;11;07 - 00;07;18;26
Dr Taylor A. Braund
So diagnosis, symptom monitoring, predicting treatment outcome, predicting relapse, things like that.

00;07;18;28 - 00;07;23;20
Dr Jasmine B. MacDonald
I'm wondering about what kinds of data can you use for, phenotyping.

00;07;23;22 - 00;07;50;06
Dr Taylor A. Braund
So the two broad categories of data, things we call active data and passive data. So passive data are things that are collected in the background on your phone, without any kind of input. So think about the global positioning system, the GPS, the way that you move your phone and the angles, the accelerometer and gyroscope. so these things are just kind of being collected in the background.

00;07;50;09 - 00;08;23;24
Dr Taylor A. Braund
the other types of data are active data. So things that require manual input. So you can think if you ask someone to complete a typing task, ask them to give a voice sample. Maybe if you work in, psych research and you use ecological momentary assessment or EMR data where you ask someone to complete, you know, miniature version of questionnaires throughout the day of regular silence, measurement of mental health, all these things would would be characterized as active data.

00;08;23;27 - 00;08;59;15
Dr Taylor A. Braund
So collectively digital phenotyping use the active data and the passive data. And then from these data we do a process called feature engineering, which is where you try and develop features that you can map on to specific mental health symptoms. So to give a concrete example, you have raw DPS data comes in. You say, I want to try and develop a feature that's related to maybe something to do with social behavior or social withdrawal.

00;08;59;18 - 00;09;24;27
Dr Taylor A. Braund
So we'll say how often does the participant move around? Do they leave the house? How far do they go? Are they going to significant locations? Are they going to shopping centers? Are they going to pubs. And then ideally those would be able to map on to the symptoms and give you some information, that you can then apply.

00;09;24;29 - 00;09;49;26
Dr Jasmine B. MacDonald
I think in the work that we're going to talk about, one of your papers, it seems like it is mostly active data, but I'm wondering around like because I, I have the sense that folks who might be slightly skeptical or concerned about research in general, but then as soon as there's like any kind of meta tracking, what kind of permissions and stuff do you need, to access this kind of stuff on people's phones?

00;09;49;28 - 00;10;17;13
Dr Taylor A. Braund
There's a range of proprietary and open source digital and typing platforms that exist out there in the world. some of the popular ones, like the bakery platform, Alliance lamp, a black dog. There's an in-house platform called migration and these platforms, you know, all the kind of checks and balances in place to kind of guarantee participant.

00;10;17;15 - 00;10;43;28
Dr Taylor A. Braund
privacy and security and things like that. But they also offer quite diverse in the amounts of data that they collect. So, some differences between the platforms, but you can think of all the passive and actively collected data that are being, you know, continuously, collected, through the fines. And there's also some novel kind of digital phenotypes being explored now as well.

00;10;43;28 - 00;11;04;18
Dr Taylor A. Braund
So I saw a paper recently where they were looking at selfies within camera rolls and the angle and the lighting of the selfie, and what that might map onto specific to other mental health systems at times. So, I think people are getting really creative in terms of like, what they're using these phenotypes for.

00;11;04;20 - 00;11;12;05
Dr Jasmine B. MacDonald
That's interesting. It's not what we we're going to dive into today, but can you remember any like high level takeaways from that selfie study?

00;11;12;07 - 00;11;27;20
Dr Taylor A. Braund
It was something to do with the lighting, the lighting in that angle. So in terms of the features I extract from the selfies, I think those were some of the important features. But yeah, I need to read it again.

00;11;27;22 - 00;11;49;16
Dr Jasmine B. MacDonald
Yeah. That's fun. I, unreasonable for me to expect you to be able to speak to other people's research on the CMOs. so I think that, like, any time you're doing this kind of study, right, you're working on large data sets. So it's, making inferences across a large number of people and not lacking a massive amount of data.

00;11;49;16 - 00;12;03;05
Dr Jasmine B. MacDonald
I imagine them because stuff is being into, you know, accessed all the time. So your paper talks about the role of machine learning in this data. is a data analysis that we're using machine learning.

00;12;03;07 - 00;12;35;26
Dr Taylor A. Braund
Yeah. For the data analysis, I mean, you can use machine learning for like different tasks. So I guess our application is what would be called supervised machine learning. So we have the features and we know what the data labels or the outcomes are. And then we kind of use the models to try and find the patterns within the data as opposed to something, say, unsupervised machine learning, where we are asking it to explore the data to find hidden patterns within the data.

00;12;35;28 - 00;12;42;28
Dr Taylor A. Braund
So yes. Yeah, as a family, I guess we'd be using supervised machine learning models.

00;12;43;00 - 00;12;49;16
Dr Jasmine B. MacDonald
Okay, cool. So on a practical level, how do you do that type?

00;12;49;19 - 00;12;56;28
Dr Taylor A. Braund
Well, I think in now we've come a long way and it's a lot easier to apply machine learning models.

00;12;57;00 - 00;13;04;15
Dr Jasmine B. MacDonald
Because you're not chucking it into ChatGPT and going like, know what does this look like?

00;13;04;18 - 00;13;21;11
Dr Taylor A. Braund
But it is funny that you mention that because they are developing large language models that you can give written text from participants data and then you can, you know, infer, mental health status as well. That's like a whole nother kettle of fish.

00;13;21;11 - 00;13;23;05
Dr Jasmine B. MacDonald
But interesting.

00;13;23;07 - 00;13;57;03
Dr Taylor A. Braund
On a practical level, how it works is we get the data we want to split the data into, say, our training and test, samples, usually 80, 20 or 73%. Then within the development, samples. So that 80% we apply some full machine learning model. We can say maybe a shrinkage model like elastic net lasso. And that involves tuning a data one or a set of hyperparameters.

00;13;57;06 - 00;14;33;28
Dr Taylor A. Braund
So you can do that a number of ways as well, either by saying choose a random block numbers, or providing a grid of numbers to search across to find the optimal, combination of hyperparameters. And then it basically optimizes the model to find the, you know, the best prediction for whatever it is you're looking for, whether that's, you know, maybe you have mental health symptoms and you have your features and you want it to try and find some combination of these features that are going to predict, mental health symptoms.

00;14;34;00 - 00;15;04;26
Dr Taylor A. Braund
I think the advantages of using the machine learning models, as opposed to say, some kind of classical statistical models like a regression or, you know, something that might be more universally applied is I can, handle these large, complex datasets. So it has the ability to find these complicated, potentially non-linear patterns within the data, to enhance that predictive accuracy.

00;15;04;28 - 00;15;36;18
Dr Taylor A. Braund
and it also can include some of, methodology like feature selection that traditionally you would have to do separately can be inbuilt within the machine learning model. So you don't have to say, you know, I would use these variables or those ones based on maybe some tree conceived idea of what their predictive ability might be. You can just leave that up to the machine learning model to decide.

00;15;36;20 - 00;15;59;04
Dr Jasmine B. MacDonald
Nice. I am wondering, is this like a kind of a high tech specific version of essentially what I think my the first lab manager where I worked would talk about data mining, like having really big data sets and trying to like, I clearly don't know enough about this to be able to extend that question any further. But are you like, no, just that's a completely different thing.

00;15;59;06 - 00;16;32;28
Dr Taylor A. Braund
Well, you can definitely mine the data. I think the one of the important things for data mining is that you need a dataset large enough in order to be able to externally validate the predictions, because you can mine any kind of data set. But I think unless you're able to have a realistic, view of what those accuracies and predictions are going to be, then it really does limit the kind of translate ability of those findings.

00;16;33;01 - 00;16;54;18
Dr Jasmine B. MacDonald
For sure. Yeah. For you being in this, Tyler, this is probably going to sound really dumb. But from the outside and really kind of grasping the practical aspect of this, the assumption I'm making is you have some suite software that you're working with, right, that is set up to do this like it's a souped up version of a basic like, almost like what I would do in say, SPSS.

00;16;54;18 - 00;17;06;01
Dr Jasmine B. MacDonald
I'm saying here's my data and the variables I want to create. And then I'm looking for relationships across them and running tests. You have like a souped up machine learning software package that you use to do this. Correct?

00;17;06;03 - 00;17;28;11
Dr Taylor A. Braund
I mean, I use R, so nice programing. So and there's a lot of packages within R that are very well developed. the carrot package is probably the most popular. It's a very big library that you can apply a lot of machine learning models to. There's also the TensorFlow and the Keras packages if you want to get into deep learning.

00;17;28;13 - 00;17;51;01
Dr Taylor A. Braund
I mean, they're really computationally expensive. You need to start leveraging high performance computing, which can also get expensive. But for run of the mill machine learning models that you can apply, these can be run on local PCs with, yeah, sufficient programing skills and latest version of the heck yeah.

00;17;51;01 - 00;18;08;04
Dr Jasmine B. MacDonald
I'm glad I asked the dumb what I thought was a dumb question, because I think that's that's great for listeners to know that that makes it much more accessible. if you're interested or if they're like, oh, you know, I might want to reach out and do some work with Taylor or similar folks. So cool. Thanks for that.

00;18;08;07 - 00;18;26;09
Dr Jasmine B. MacDonald
I'm wondering if you could, you know, talk to me about, digital phenotyping and some kind of broad, examples to help the audience understand the kinds of questions or the the kinds of real world problems we're trying to solve with this.

00;18;26;11 - 00;18;56;16
Dr Taylor A. Braund
Yeah, sure. So one of the, kind of immediate translatable applications would be using the digital phenotypes to identify these critical windows where we can use these interventions or just in time interventions to increase the efficacy of the treatment. So is there a specific critical window that can be identified using the digital phenotypes? A translatable example from medicine would be the glucose monitoring systems.

00;18;56;16 - 00;19;15;25
Dr Taylor A. Braund
So they provide real time feedback on the blood sugar levels. And they can give the patient information on whether they should be taking insulin or consuming carbs. That would be similar to digital phenotypes. Identifying when the intervention should be applied at the optimal time to increase efficacy.

00;19;15;28 - 00;19;42;06
Dr Jasmine B. MacDonald
That helps make things quite concrete, actually. I guess the way I've been thinking about this when I was reading your work is the information that this can potentially give clinicians, but from what you describing, it also sounds like actually it can just be a prompt to the person for their own insight and awareness of like, yeah, increasing insight essentially, and like a little checking of like, hey, could be a good chance.

00;19;42;06 - 00;19;48;23
Dr Jasmine B. MacDonald
Like, this could be a good time to reach out to your GP or to reach out to your psychologist or whatever else.

00;19;48;25 - 00;20;22;11
Dr Taylor A. Braund
Yeah. And I think that is definitely going to kind of make the clinicians in terms of their appetite for, digital phenotyping as well, because there is such like a broad spectrum of applications that know people are investigating from diagnosis, symptom monitoring, treatment, prediction, but then also ways to, integrate, the clinician into their decision. let's say how to improve these, diagnostics and treatments.

00;20;22;14 - 00;20;38;17
Dr Jasmine B. MacDonald
That's really interesting. Are there some disorders compared to others when we come back to psychology or mental health now and there's some disorders or conditions that phenotyping works better for predicting.

00;20;38;19 - 00;21;09;00
Dr Taylor A. Braund
So I think some features are probably more relevant to some disorders. So if you're if you have specific symptoms within the disorder, maybe as kind of circling back to social engagement or withdrawal and GPS style, with mood disorders. so, so schizophrenia. Yeah. voice and speech is, critical kind of symptom that can be addressed with digital phenotyping.

00;21;09;00 - 00;21;46;06
Dr Taylor A. Braund
So there's quite a broad research area looking at the structure of language within, patients. And they get quite good predictive results using those features. So different I think different disorders will be better suited to different, digital phenotypes. And also I guess you need some kind of symptom fluctuation. So if you have kind of uniform, symptom levels and it's very hard for digital phenotypes to track or pick them up.

00;21;46;09 - 00;22;03;12
Dr Taylor A. Braund
And then I guess also disorders that are not externally visible. So if it's yeah, if you're taking movement and measurements and it's not something that can really be tracked with the phenotypes and that can make it difficult as well.

00;22;03;14 - 00;22;33;03
Dr Jasmine B. MacDonald
so you we'll talk more about this in a seq you used, for instance, in at least one of the studies I've written task. But perhaps like it's sounding like what you're saying is that for some other disorders, like a psychosis in schizophrenia, then potentially you could have built in some kind of speech task like, I don't know what that would look like, like maybe having a prompt to talk about some aspect of the day or their week or something.

00;22;33;05 - 00;22;41;13
Dr Taylor A. Braund
Yeah. So in the future, proofing study, the one that I'll talk about in a minute, there is a voice task within that, the max time.

00;22;41;17 - 00;22;49;27
Dr Jasmine B. MacDonald
So let's do this. Let's talk about the future proofing tasks. Tell me like what's what was the deal with it. What what did it aim to do?

00;22;50;00 - 00;23;24;14
Dr Taylor A. Braund
So the future proofing study was, this, kind of massive study. It's collected data from over 6000 Australian adolescents across Australia. And it was a cluster randomized, controlled trial looking at prevention for depression was the primary outcome. and yeah, as part of that study, they have the future proofing app, which was a custom built app, digital phenotyping app that collected, a range of these digital phenotypes that we've been discussing.

00;23;24;14 - 00;23;56;13
Dr Taylor A. Braund
So in addition to the passively collected data, so the GPS and the accelerometer and the gyroscopes, they also completed a number of task based activities to collect data. These included cognition tasks, typing tasks, voice tasks. And so all that, massive amounts of data is kind of being investigated at the moment. And the one that I have been diving into is the typing data.

00;23;56;16 - 00;24;32;09
Dr Taylor A. Braund
So this captures the detailed information in time in between key presses. So this kind of research area has come from the field of biometrics. So authenticating people's identity. So the way that you type on a keyboard is kind of, you know, like a password in a way. And we can I guess mental health has been leveraging this application to see if typing behavior is related to mental health symptoms or diagnosis.

00;24;32;11 - 00;25;07;03
Dr Taylor A. Braund
And I guess the main symptom that they believe that it should be mapped onto is psychomotor impairment or agitation. So that's a core symptom within depression. And this should kind of theoretically manifest itself. with this symptom. Spoiler alert I just finished the paper looking at the individual symptoms, and it doesn't seem like that's ageist. but leaving that to the side, the idea is you can kind of use these total scales of mental health.

00;25;07;03 - 00;25;21;22
Dr Taylor A. Braund
So depression, anxiety, distress or insomnia, which is what we looked at in the current paper, as opposed to looking at specific individual individualized, symptoms.

00;25;21;24 - 00;25;54;12
Dr Jasmine B. MacDonald
so what's really interesting, I noticed through recording and editing, audio that people have this pattern to the way that they speak. And I noticed that in myself that I do these pauses before, like partway through a sentence. So it sounds like what you're saying is that there's probably, baseline intonation, for lack of a better word, of basically like the, the rhythm or the time that people typically spend typing or, or the pattern of typing.

00;25;54;18 - 00;26;08;21
Dr Jasmine B. MacDonald
And then, like you were saying before, what we need is like a fluctuation in that patterns that, we're trying to look for a correlation or association with mood in this context. Is that like a good basic summary?

00;26;08;23 - 00;26;35;13
Dr Taylor A. Braund
Absolutely. Yeah. And I think the I guess the way that you would expect these relationships to manifest depends on how you think depression is associated with typing data. So let's say you make an assumption that if your if one person is depressed, they're going to have psychomotor impairment, they're going to be taught. So the hypothesis they're going to type slower right.

00;26;35;15 - 00;26;54;25
Dr Taylor A. Braund
On average. But as we know depression symptoms are very diverse. and some people with depression have psychomotor agitation. So they might potentially be talking faster. So it's also very individualized in terms of how these relationships are going to work.

00;26;54;27 - 00;27;13;23
Dr Jasmine B. MacDonald
I'm chuckling to myself, why are you speaking about that? Because I'm thinking of being in early high school, having a sweet Nokia 3310. And it was this. You were pretty awesome. If you could type super fast and actually that, like I'm in my head, I'm like, I'd be thinking this is associated with some kind of, you know, anxious, elevated state.

00;27;13;23 - 00;27;35;25
Dr Jasmine B. MacDonald
But again, that comes back to looking for the fluctuate motions and, and changes and, and having large sets of data and probably in a clinical context, being able to have other kinds of input that's brought, you know, data collection is brought in alongside this information that helps inform for this person or broadly, for people who experience depressive symptoms or anxiety symptoms.

00;27;35;25 - 00;27;39;10
Dr Jasmine B. MacDonald
This is the kind of changes that we say.

00;27;39;12 - 00;28;04;19
Dr Taylor A. Braund
Yeah. I mean it really depends on what the research question is as well. So for this paper we are looking cross-sectional. So we're kind of asking population level questions like what generally are the associations between say talking behavior and mental health symptoms to people you know to fast or when they're more depressed or do they talk slower or a lot faster when they're anxious?

00;28;04;21 - 00;28;39;10
Dr Taylor A. Braund
once we have this kind of rich data, this longitudinal data, we can ask, I guess, more detailed questions like, what are the individual typing profiles of someone? The participant? And how do these, you know, maybe predict future depressive symptoms? Do they predict whether you're going to respond to treatment or do they predict potentially new onset of mental health, issues or future mental health episodes?

00;28;39;13 - 00;28;47;01
Dr Taylor A. Braund
So it really opens door to like quite wide right. wide applications.

00;28;47;04 - 00;29;14;02
Dr Jasmine B. MacDonald
Totally. I'm wondering about in the process of establishing different phenotypes and thinking about how predictive they are, what is validation look like. Is this like would be working with a large like you had 6000 participants. Do you have information about whether there is a current diagnosis or they're currently accessing support for a certain disorder, or what does that look like?

00;29;14;05 - 00;29;41;27
Dr Taylor A. Braund
So of the 6000 or so who completed the future briefing study, about a thousand had typing data. So they completed the typing task. And then of those participants, I guess the the population cohort of future proofing is just high school students. Like there's no like inclusion criteria that requires them to have a mental health disorder at a population level.

00;29;41;27 - 00;30;06;19
Dr Taylor A. Braund
Some are, some aren't. Yeah. this is just a population scan of associations generally. So we do some subgroup analysis like throughout the paper. So we look at talking differences between males and females and people who make clinical cutoffs for, for certain disorders. But yeah that that's kind of the scope at least this paper covers.

00;30;06;22 - 00;30;14;23
Dr Jasmine B. MacDonald
Yeah. Nice. What are the tasks look like. What kind of stuff are you asking these adolescents to write about?

00;30;14;26 - 00;30;39;21
Dr Taylor A. Braund
Well there's two as you mentioned, there's two typing tasks. So we call them the composition typing task. And the prose typing task. Names aren't that sexy, but it doesn't matter what they actually do. So the composition typing task is the participants are given one of a prompts. So a time I was happy, the time I was sad, my hopes for the future.

00;30;39;24 - 00;31;01;29
Dr Taylor A. Braund
And then they're asked to write a certain amount of text on that topic and the other one is the prose task. So in the prose task, participants are given one of a scripts, and these scripts, kind of like short stories. And the participants asked what as much text from that short story as they can, within 30s.

00;31;02;02 - 00;31;35;01
Dr Taylor A. Braund
So we got the prose typing data. So the time between the key presses on the prose task and the composition task, and we kind of do some supplementary analysis looking at the timing of the typing between the tasks. And we generally find the same patterns with the mental symptoms. So for this paper we combined the typing behavior across the across the tasks so that if you if you're really interested you can find the supplementary material online.

00;31;35;04 - 00;31;50;27
Dr Jasmine B. MacDonald
Excellent. Presumably you thinking that there was going to be a difference between the two tasks. Right. And like we love as researchers to be proven wrong. And that's fine. You adopt, you can put those two together. But what did you what was the team's thinking on having the two tasks and how they'd be different?

00;32;03;29 - 00;32;13;09
Dr Taylor A. Braund
The idea basically is you want to have kind of two different, goal directed tasks, one where you have kind of a time limit.

00;32;13;16 - 00;32;38;26
Dr Taylor A. Braund
So, you know, 30s you want them to be typing as fast as I can. The other one, the kind of human free range to have some expression to write out the answers and the composition task where they're given the prompts and they can type, you know, freely, you can then take that text, and then you can apply linguistics analysis to it and develop linguistic digital phenotypes as well.

00;32;38;26 - 00;32;44;24
Dr Taylor A. Braund
So that's something we've been doing in other studies and in the future proving study interesting.

00;32;44;24 - 00;32;52;23
Dr Jasmine B. MacDonald
So that's less about, say, things like time and patterns and typing and more about the words used in the sentence structure and stuff.

00;32;52;25 - 00;33;26;05
Dr Taylor A. Braund
Exactly. so you can maybe say, what is the proportion of words in their body of text that have a positive affect connotation. How many have a negative affective connotation? Kind of, very frequently cited for this stuff. People with depression, it's a high frequency, a first person, second pronouns. So a lot of, in the sentence structure and this is kind of mapped on to symptoms, with rumination.

00;33;26;08 - 00;33;52;01
Dr Taylor A. Braund
So when you're ruminating there's a lot of self thoughts. So and that manifests in writing styles. So and then you know ideally somewhere down the road you could potentially map typing features onto linguistic markers and how they relate to each other. Maybe you have a composite marker, which is a combination of their typing behavior and what they're actually writing.

00;33;52;04 - 00;33;56;15
Dr Taylor A. Braund
So there's a lot it's kind of ways that these can all be brought together.

00;33;56;18 - 00;34;12;26
Dr Jasmine B. MacDonald
Yeah. Yeah. Cool. All right. And you had the two writing tasks, but they were brought together as the one, as like a composite task. And the major outcome was, depression. So what did you find?

00;34;12;29 - 00;34;51;06
Dr Taylor A. Braund
So basically what we found is that overall higher mental health symptoms were very weakly associated with faster typing speed. So more depressed faster typing speed but less typing frequency. So they're typing less. This is kind of counter intuitive to what you would expect, considering most people have found previously, and would expect that more depression or higher levels of depression would lead to psychomotor impairment, which would lead to slow typing speed.

00;34;51;09 - 00;35;13;21
Dr Taylor A. Braund
So I think this is a really interesting finding in terms of what is it that's unique about the data collected that we have and how that differs from other studies that are out there. I mean, a few off the top of my head is an adolescent sample. Most of the other studies, in adults and in clinical samples.

00;35;13;23 - 00;35;47;01
Dr Taylor A. Braund
Right. also the type of typing that they're doing. So these are task based tasks, where there's very, popular digital phenotyping app called the BAE effect app. Alex Liao developed. And this captures typing behavior in, I guess, natural use of your smartphone. So it's not a goal directed toss. It's just collecting the data passively throughout people's day.

00;35;47;01 - 00;35;53;14
Dr Taylor A. Braund
So maybe there's something that's happening between asking someone to do a task and just collecting it.

00;35;53;16 - 00;35;54;08
Dr Jasmine B. MacDonald

00;35;54;11 - 00;35;55;28
Dr Taylor A. Braund
In the wild.

00;35;56;00 - 00;36;17;02
Dr Jasmine B. MacDonald
Well, this is fascinating, Tyler, because I'm thinking about stuff like, not at all trying to pathologize individual stresses or anything, but, you know, you can. I think everyone listening has a sense of when you're feeling more or less anxious or stressed about something, how there is info on your phone that would easily help show that or remind you of that.

00;36;17;02 - 00;36;36;25
Dr Jasmine B. MacDonald
Like for instance, even picking up your phone to check, you know, unlocking your phone or going into a certain app, checking messages or. Yeah, I think the range of data and the practical applications is really interesting, but I'm sorry I cut you off because I was musing of how. Sure, that's nice.

00;36;36;28 - 00;37;01;09
Dr Taylor A. Braund
Yeah, absolutely. I mean, the end I mean, you know, picking up the phone and unlocking it, that's activating like the gyroscope. So that's like sending stuff and a lot of those, kind of metadata that are collected, through the phone. So call logs, text logs, you know, how active you are. And then, you know, you can even kind of insert a sleep patterns based on, like sewing movement.

00;37;01;11 - 00;37;28;14
Dr Taylor A. Braund
Or if instead of a smartphone using a smartwatch that your feet and you can kind of have more detailed sleep, and movement patterns on there. But yeah, it can be very diverse in terms of the things that it's capturing. The other results in the paper, I guess, is we be interested in looking at any differences between males and females, and females show this similar pattern of faster typing.

00;37;28;18 - 00;37;58;03
Dr Taylor A. Braund
It's higher mental health symptoms, but for males it seems to flip. They do actually show slower typing with higher mental health symptoms. So I think like kind of investigating these, why these findings are occurring is something that's like going to be of high interest, moving forward, because a lot of the times people even just want to partially are these covariates or it's, not addressed.

00;37;58;03 - 00;38;09;00
Dr Taylor A. Braund
So saying that it has, you know, quite a different impact between male and female. This. Yeah. Interesting.

00;38;09;03 - 00;38;34;06
Dr Jasmine B. MacDonald
I'm wondering about what you would like to come from this work like that applied aspect. You we've talked briefly already about phenotyping broadly and the applications. So I have no questions about that. But if people, reading your work or looking to apply the work or you yourself like what would you like to see come out of this, the kind of practice applications of the key learnings?

00;38;34;09 - 00;39;02;24
Dr Taylor A. Braund
Yeah. Well, I mean, I would characterize the findings as descriptive as kind of these are like some very high level population like associations. Obviously, the the associations are very weak. Like there's no practical application. Like we run a whole range of machine learning models to see if there was some non-linear kind of association between the features that could predict mental health symptoms as well.

00;39;02;27 - 00;39;42;04
Dr Taylor A. Braund
The best way I can kind of described is like trying to get blood out of the stone. There's just really not not much there. So I think one of the implications from it is to move the focus towards applications that might be more fruitful. So, you know, exploring whether they can be used to, identify these critical windows to, implement these just-in-time interventions, whether the data captured longitudinally can identify, you know, these person, specific signatures on the type of behavior that might be more fruitful.

00;39;42;06 - 00;39;54;14
Dr Taylor A. Braund
So as I think as important as knowing where to go is at least tells us where not to go, but resilient locations are going to be.

00;39;54;17 - 00;40;08;00
Dr Jasmine B. MacDonald
Yeah. Yeah, absolutely. That's really important foundational work that future work is going to build upon. Cool. I think that's a very reasonable, valid response for sure.

00;40;08;02 - 00;40;38;13
Dr Taylor A. Braund
Yeah. I think that just another thing to add is I think there's so much we don't know at the moment about digital phenotyping as well. Like it's such a young field. And I mean, just the differences between the tasks, differences between the digital phenotyping platforms, how they've collected the differences between the samples, like there's just so many factors that we don't know how they're impacting the results.

00;40;38;15 - 00;41;05;11
Dr Taylor A. Braund
So I think like methodologically, there's like quite a way we can go to improve this. But yeah, it is. And I mean, when I first started that paper, there was I did a systematic review and there was eight other papers that have looked at key strike features associated with mental health symptoms, like that's in a paper. So there's not many people working on it or looking at it at the time.

00;41;05;11 - 00;41;23;06
Dr Taylor A. Braund
So kind of if you want to develop like more comprehensive digital phenotyping, you kind of have to explore these novel methodologies like like talking behavior. It's not something that's, you know, included in many digital phenotyping platforms, if at all.

00;41;23;08 - 00;41;44;04
Dr Jasmine B. MacDonald
That's really interesting that there's so much available data out there. But this really sounds like cloud in emerging space as a methods. So that's pretty exciting I think. So what's next for you like hip hit me up with some shameless self-promotion. Like what. What have you got going on. What are you excited about.

00;41;44;06 - 00;42;21;18
Dr Taylor A. Braund
What am I excited about? So recently I'm diving into accelerometer data. So sound movement data and specifically mapping it on to individual symptoms of depression. So, Antonio, two of the things I'm excited about is looking at different, data modalities. The second is we've recently, got funding for use of high performance computing. So we can really explore, some more rigorous methodologies and pipelines.

00;42;21;18 - 00;42;54;10
Dr Taylor A. Braund
And even if we have the bandwidth go into deep learning. it's quite a learning curve. But I think, you know, there's there's some application there. And also having the bandwidth to kind of leverage multiple data sets with digital phenotyping data now that they've completed. So yeah, merging the data sets, looking at specific applications of these just in time interventions, new feature development and specific symptoms.

00;42;54;13 - 00;42;56;05
Dr Taylor A. Braund
Yeah it's a there's a lot going on there.

00;42;56;05 - 00;43;07;17
Dr Jasmine B. MacDonald
But yeah yeah sounds like it. That's really exciting. And if folks who are listening want to follow along and see what you're up to, what's the best way for them to keep track or reach out?

00;43;07;19 - 00;43;19;26
Dr Taylor A. Braund
Yeah. So I put everything on Twitter. I think my handles are tailored for one so you can find me on there tweeting about, you know, everything science and sometimes known science.

00;43;19;28 - 00;43;20;15
Dr Jasmine B. MacDonald

00;43;21;13 - 00;43;38;29
Dr Taylor A. Braund
I'm on LinkedIn and yeah, I have profiles on the, you NSW website and the Black Toad website. So my email should be on most of those. So if you're interested in learning more or wanting to collaborate, feel free to reach out.

00;43;39;01 - 00;44;09;18
Dr Jasmine B. MacDonald
Awesome. I mean, this is I think how we came into contact was through Twitter. So, it's a good place. Like most of the folks that I talk to on this podcast I've met on Twitter, I'm like, this is how we do our water cooler conversations and find like minded folks. So yeah, definitely encourage people to especially, I think if you're a researcher, how to find your people on Twitter or LinkedIn, all of that stuff that we just talked about, links to the paper for, that we've discussed today and Twitter and LinkedIn and stuff.

00;44;09;18 - 00;44;25;10
Dr Jasmine B. MacDonald
I'm going to pop that in the show notes. If, folks are interested to go and check that out. I'm going to segue out of something that you said, Tyler, which was, mostly tweeting about science stuff, but there's other stuff. What are you doing when you're not nerding out on research?

00;44;25;12 - 00;44;38;13
Dr Taylor A. Braund
I play, discussing amount of chess, like I play with a lot of chess, so I play online. It's really an addiction, and it's it's bad. So I play one chess.

00;44;38;15 - 00;44;45;07
Dr Jasmine B. MacDonald
Do you need help? You asking for help right now?

00;44;45;09 - 00;44;47;07
Dr Taylor A. Braund
that.

00;44;47;09 - 00;44;49;19
Dr Jasmine B. MacDonald
Is that. Have you always been into chess?

00;44;49;21 - 00;45;19;01
Dr Taylor A. Braund
Well, I played it in high school, and, you know, I liked it. And then I didn't play it for a while, like a decade. And then I actually started playing in the game of Little People. So I started. Yeah. And I guess I'll competitive like science just flat and yeah, from there I've just been training and playing a lot of the time, so just making sure he's levels not higher than mine.

00;45;19;03 - 00;45;41;01
Dr Jasmine B. MacDonald
Nothing wrong with some healthy competition. for folks listening who, kind of long term, sock attack listeners. Gabe is Doctor Gabriel Tillman, and Gabe has been on two episodes of Sock Attack. So it's one of the first episodes where we talked about, mathematical psychology, which there's a huge amount of overlap with what we're talking about today and Gabe's work.

00;45;41;01 - 00;45;58;15
Dr Jasmine B. MacDonald
And, turns out that, you know, good people now will find each other, and Tyler and Gabe, are actually mates, which is cool. And, yeah, the other episode, Gabe interviewed me about my research in trauma, and, Newsweek is. Okay. So what do you like about chess?

00;45;58;17 - 00;46;05;11
Dr Taylor A. Braund
Oh, that is a really great question. What do I like about chess? Because if I'm not winning, I don't really like it at all.

00;46;05;14 - 00;46;07;29
Dr Jasmine B. MacDonald
this is the competition aspect of the challenge.

00;46;08;02 - 00;46;32;14
Dr Taylor A. Braund
I think it's the competition aspect also like learning techniques and then being able to apply them effectively is very rewarding. So, you know, learning like it's a lot of pattern recognition and, you know, puzzles and learning openings and things like that. So, you know, applying them once you've won at them and beating, people is yeah, very rewarding, I think.

00;46;32;14 - 00;46;32;19
Dr Taylor A. Braund
Yeah.

00;46;32;19 - 00;46;40;25
Dr Jasmine B. MacDonald
Cool. Do you go back and watch like, YouTube clips or research winning moves or like these kind of stuff?

00;46;40;27 - 00;47;02;29
Dr Taylor A. Braund
I think it's a lot of analyzing games afterwards. So it's, you know, every game after you play, you can go through analyze, you can find out where you made a blunder. I always made a mistake what the optimal move should have been, what the continuations are, what the lines are, what the right lines should have been. So you can probably spend more time analyzing the playing field.

00;47;03;02 - 00;47;12;15
Dr Taylor A. Braund
Yeah. Getting sensitive. But yeah, it's a lot of, time commitment if you're doing it regularly, I bet.

00;47;12;18 - 00;47;22;08
Dr Jasmine B. MacDonald
Yeah, I feel like, I wish I knew this before because I could have worked in so many chess puns.

00;47;22;11 - 00;47;43;07
Dr Jasmine B. MacDonald
Taylor, this has been a really great chat. I yeah, I really like speaking to folks like yourself who are doing work that we have shared language, but what we do is quite different. And I really appreciate learning and thinking about the world and data and and research and psychology in different ways. So I appreciate your time.

00;47;43;15 - 00;47;48;23
Dr Jasmine B. MacDonald
I think what you're doing is awesome. Thanks for coming along and sharing your expertise with us.

00;47;48;26 - 00;47;51;12
Dr Taylor A. Braund
Thanks. I mean, it's been a great time. Great speaking with you.