 So, thank you for the great introduction. And I'll be talking about digital phenotyping and I want to first give credit to the two PhD students who did almost all the work that I'm going to be talking about that are I'm co supervising. I'm George Albers here at Tolberg University, who is also funded on a Tolberg University grant the impact program, and Karen Vandenberg, who's a PhD student at Maastricht University, and a clinical psychologist. So, what I want to talk about today is predicting and modeling momentary well being and why at the outside should we care about momentary well being, and the critical thing is that we're getting more and more importance that all of the facets that make up well being are important not only for our physical health and our mental health in the short and the long term. So they play a critical role in supporting all the other aspects of our life our social life, our educational life, our work life, and everything else, and it's becoming more and more clear than understanding how our well being is changing on a moment by moment basis is has direct consequences for both our sort of immediate well being but also the long term well being and I think that it's increasingly a focus of research and this is really but in the last couple years the an explosion in research under the heading of digital phenotyping of well being. And there's a couple of criteria that the research kind of needs meet in order to satisfy the criteria of digital phenotyping the first is the quantification of individual level human states, and it is happening on a moment by moment basis, and hopefully using passively collected personal digital information. And I'll kind of flesh that out as I go along here and give you a number of examples, and this is most popular for clinical populations where there's real impact with the mental well being and mental health of the patients, but it's increasingly being used for non clinical assessment, particularly in adolescence and I'll show a couple examples from our lab where it's being used today. To give you an overview, what I mean here is that the target is to infer what I've labeled here as the active sensing. This is where you, you're, you are asked, you know, how anxious are you feeling how stressed are you feeling how lonely are you feeling how depressed are you feeling, and you give an answer usually not a Likert scale I'm feeling for. It's relatively invasive data to collect, and it's relatively in expensive in time you have to ask this question over and over again to someone to build up a moment by moment model of what what they're doing. The goal is to be able to predict what state people are in the digital phenotype that they are experiencing in a moment based on passive sensing so that's the left column here. Things like this is coming from Melcher here these are all things that can be measured on somebody's phone, you can also imagine wearables, as well as ambient sensors could all be used to build models that link passive sensing to inferring the active sensing activity. And I've kind of labeled in the middle here that a lot of researchers are either aggregating or building representations of what I've sort of labeled as theoretical concepts. Things like social activity is really, there's a number of signals that you might be able to collect that inform you about the social activity that a person's engaging in. And the social activity may actually be linked to multiple different wellbeing states something that I'll kind of come back to over and over again. So how does this end up being implemented well what happens is if you look at the phone on the left here you see done in yellow, I don't know if it's big enough to read, but this is the phone collecting the passive data, sort of logging what's going on at whatever is part of the project. At the top you're actually collecting the active data, the active sensing this is asking people with surveys or questionnaires how they're feeling at the particular time, usually with a particular focus on particular phenotypes that you want to study. And then the, once you've collected some data, usually from multiple people, you build a model that is connecting these two data sources with the idea that then that model can then be generalized and put onto other phones and use way other people to infer their state, given that the passive sensing that's being collected on the other phone. So we're going to walk through four digital phenotype being studies that coming out of our lab. First, correlating procrastination in adolescence with using dynamical structural equation modeling. Then two projects looking at predicting sleep and stress in adolescence using machine learning techniques. And then finally looking at the temporal dynamics of mood and imagery in clinical patients with multi level factor, auto regression modeling. That's where we're going the rest of the way. So first the correlates of procrastination in adolescence. What we're looking at here is building models that are aiming to predict procrastination or to understand procrastination on the far right. Based on noted the notifications they're showing up on a phone, the screen time the amount of time that a person is spending on the phone, as well as what apps they're using and for how long. And sort of getting a sampling of different passive sensing coming from the phone. And the technique that George used in this was a Bayesian dynamic structural equation model. And what we're looking for here is looking for correlations within the moment between procrastination which is the central diamond here, and the various measures and what we see is that there's a pretty strong relationship with total phone usage, and relatively strong relationship with social media, and less so with other other categories of app usage, and a very weak relationship with the number of notifications that people are getting, or the fragmentation of how they're using their phone. But this is a model constructed across over 200, almost 250 participants. What we see if you build networks of individual people which this technique allows, is that there's a huge amount of variation across people. The connection with total usage is pretty constant at least across these three, but if you look participant a there's a very strong relationship when that that user is using games on their phone. They're very likely to report that they were procrastinating, but you see that there's no connection between game usage for participants B and C. They're much more higher loading on videos, for instance, when they indicate that they're procrastinating. We can then look at the connection weights that we infer from every single person and plot them to see what the distribution looks like and so this is the correlation between total usage and procrastination for every person. And we see the dotted line here is zero that most people have a positive and many people have a very positive relationship between total usage and procrastination so when they're using their phone war, they're reporting that they're procrastinating, which gives us a pretty clear signal that if we want to have an app on somebody's phone that intervenes, that might be something that that you would end up wanting to track and use as a as a predictor. So if you look at the other aspects of it, that for some people there for for at every top here there's some people who have a stronger relationship with all the other categories, but it's really if you look at the majority of people, there is not a strong relationship for all of the other measures and procrastination, and this suggests that there's one feature that seems to be pretty predictive and then you might want to learn an individual model for individual people on their phone, for instance, is indicating to them Hey, not only are you using your phone where you're using video apps and so that might be pretty indicative that you're procrastinating. Maybe you want to have a notification that interrupts that behavior gets you back on task. Okay, going on to predicting sleep duration in adolescence. What we're looking at here is predicting an interim theoretical concept here sleep based on screen time and sleep is really interesting because it's correlated with lots of other outcomes including academic performance. This is a model that was built machinery model that was built trained on around to around 150 people and then generalize to around 50 people. And you can see that we're doing really really accurate prediction for the people in the training set, which is maybe not surprising but correlation of almost one. And then for people whose data were not trained on at all we're still seeing correlations of point six, which is really pretty high given that with very intrusive sensors that we're reading use in our study the state of the art is currently getting about point seven. And so this is really pretty exciting. And we see again, large individual differences in how well the model model predicts both duration. And if we look at then predicting the self reported quality of the sleep, we're getting very accurate predictions for many people but then very inaccurate for other people suggesting that there may be, again, large individual differences across people. When we look similarly using machine learning to predict stress. So in this paper we're now using additional information like the time of day and the day it is, as well as the apps people are using and the sleep features from the previous study to predict stress, we can look at the same kinds of models so now we're looking at the prediction error of how stressed people are reporting at a particular moment. And we see again that there's a widespread across people, we're doing pretty well for for a number of people, but doing poorly for some and doing really well for others. But if you look at the, if you build a baseline model that doesn't that just looks at an individual's overall performance. We see a different story where there's some people where the these features are just not very predictive they're actually leading us to be worse than baseline just predicting whatever the stable state is, but for other people, it's giving us a lot of traction improving performance, suggesting that there's again, large individual differences in how well we can do in the staff. And then finally looking instead of instead of at adolescents looking at patients with bipolar disorder, we looked at the temporal dynamics of loot and sort of how it evolves across time. And so instead of looking at passive sensing here we're now looking at the constellation of different phenotypes, particularly now we've added intrusive imagery which was theoretically motivated to be a potential causal mechanism for patients with bipolar disorder to either enter manic or depressive states. And what these patients did is they self reported their intrusive imagery their mania their anxiety and depression across many, many weeks. And what we did is we estimated a network of connections between these different states and how they evolve over time. If you look at contemporaneous connections so what values are high together which values are low together, we found that anxiety and depression, we're very strongly related, but the measure of intrusive imagery seem to move along with them suggesting that there might be some some validity to the theory that these go together. When we then looked at how these correlate across time, how, how do they project forward more leaning towards a causal mechanism, we see a different story, where there it seems like there's more influence going towards the intrusive imagery than it is the intrusive imagery projecting out to these other concepts suggesting that there may be a more complicated story underlying this. And if we regress away all of that and we just look at between participants we see some consistency across people, but it's not capturing very much suggesting that there's again, a lot of individual differences here. So to conclude, what we see is that there's relatively reliable relationships and accurate prediction of digital phenotypes across domains and populations. Consistently strong individual differences demonstrated in each of the domains I talked about sleep is maybe the best, and the more moved more variable it is. And I think that I'm excited about this because there's an opportunity for individually customized well being phenotype without data aggregation or centralization that would be more federated learning happening on one person's phone that doesn't need to be shared with others. So thank you for your attention for chatting more about this and you can check out pre prints on this on one page at the bottom.