Good morning. My name is Justin Baker and welcome to this edition of the APA's virtual immersive session. I'll be talking today about integrating data from smartphones and wearables into psychiatric clinical practice. Why do we want to begin thinking about integrating data from smartphones and wearables into a clinical practice in psychiatry? Although there are many reasons to consider doing this, one way that I think really comes from patients is to say that it's really better ultimately to be grounded in reality when working with patients. This is true for patients with all medical conditions, but especially in neuropsychiatric conditions where subjective reality often seems to take precedence. I'll be talking today about how bringing reality anchoring into psychiatric practice can benefit both clinicians and patients. What is digital phenotyping and how is it relevant? In essence, digital phenotyping represents the neocreplinic approach to history taking and longitudinal assessment, but without having to follow someone around with note cards. Those among you who are historical buffs in psychiatry will recognize the face of Emil Kraepelin, who is one of the forefathers of modern psychiatric taxonomies. What he did at the time was really walking around different psychiatric asylums at the time and taking detailed notes of longitudinal trajectories. In some cases, what we're doing today with the advent of technology such as smartphones and wearables is being able to do this, but without the cumbersome methods that were necessary back in Kraepelin's day. I'll take you through some of the opportunities and challenges of digital phenotyping for a busy psychiatrist, as we described in this paper from several years ago, published in Psychiatric Annals. To start with some of the opportunities, we often ask psychiatrists to start patients or maintain them on therapeutic interventions, whether that may be psychological, behavioral interventions, or medications. We often don't know how that person is doing, whether they're experiencing side effects or whether they've stopped taking the therapy entirely. One of the important ways to think about digital phenotyping is it allows psychiatrists to still maintain a sense of how that patient is doing even outside of the hospital, outside of the clinic. There are many opportunities for anticipating and even preventing relapses of behavioral conditions using remote monitoring or digital phenotyping. We'll come back to that as well. Perhaps the one that's most relevant for today's remarks will be how we use digital phenotyping to provide quantitative data pertaining to aspects of mental and behavioral health that can aid the clinician and the patient. Finally, although I won't get into it too much today, one exciting area about digital phenotyping is what's called just-in-time adaptive interventions, which is a way to use remote monitoring to deliver optimal dosing of therapeutics based on real-time information about the person's state or environment. Those are some of the opportunities, and we'll come back to digging into several of these in a moment. But there are some challenges. One is the lack of standardized methodologies. This field is still relatively new, and so the lack of standardized methodologies means that systematic data analysis using sensory mental health, associating sensor data with mental health, has been challenging. But there is this move towards open-source platforms that helps lead toward transparency and reproducibility, and that allows larger and larger studies to be able to replicate findings. Similarly, with lack of standardized methodologies, the statistical tools to handle the amounts of data one can gather through digital phenotyping methods are still in development, and this highlights the need for further innovations in statistics to interpret the complex data you can get using these technologies. But of relevance to today's remarks, and I think that this audience is likely to be mostly practitioners, one key challenge is how do we build trust among patients and providers? And I think what this really requires is educating and guiding patients and other health care providers on what are the appropriate use cases so we can demonstrate how and when digital phenotyping can be genuinely beneficial to the patient's experience and the provider's experience. Believe it or not, efficacy has yet to be really demonstrated in real-world clinic applications, and so finding consistent associations between sensor data and mental health status is going to continue to be a challenge. And finally, there are a host of ethical concerns given that the types of data we collect with digital phenotyping can be very sensitive data, and therefore this raises questions around patient autonomy and privacy, and we believe that it's really essential to include patient perspectives in all the ethical discussions so that we can ensure a good balance between privacy and accessibility. It may be that in some cases providers have a higher tolerance or rather are more sensitive to privacy than in some cases the patient might be in order to get the best experience. So with regard to the ethical concerns, we've developed an ethical checklist for doing digital health research in psychiatry. We felt that this was a practical solution after convening a number of stakeholders from both patients as well as legal and ethical experts, and so what this does is allows someone who's wanting to engage in this kind of work, whether it's for research or clinical purposes, to really reevaluate key things such as informed consent, equity, diversity, and access, privacy and partnerships, regulation in the law, return of results, as well as aspects of duty to warn and report when conducting this kind of remote research. Of particular relevance for today's presentation is this notion of return of results. We feel that it's essential to begin incorporating return of results from whether it's research studies or also in clinical use cases, and that this can potentially benefit patients, but it's still early days and so there's some data that we can collect using digital phenotyping that we may not want to provide to the patients directly. We might want to have that mediated through a clinician, and so this paper goes through a rubric for how to consider when it's appropriate to return individual research results from digital phenotyping to individual patients. I want to emphasize that this article that was published in the American Journal of Bioethics had a number of very insightful comment articles that came from this, so if you're interested in this area, I highly encourage you to check out this series. So we feel that it's essential to approach problems of bringing technology into psychiatry from a design-centered thinking approach. So what do we mean by this? It's to really empathize with the stakeholders, whether that may be the patients, the providers, or other people who work in the system, with particular pain points or problems they may be experiencing, and to really, after empathizing with them, define what it is the problem is, ideate on potential solutions together, prototype those solutions, and then actually test them and determine whether we've hit the mark or whether we need to continue to iterate, and iteration is going to be the norm in this kind of work. So several questions that come to bear when you're beginning to think about integrating data from smartphones and wearables is what results do you even show the stakeholder? When in the treatment process do you show them? Is it between sessions? Is it during a session? Do you have the clinician first digest it and then make sense of it to the patient, or do you show it to them together? Who receives them? Patient? Provider? Both of them? Do they see the same information or different information? How much do you go beyond what's seen in the raw data? In other words, how much inference or interpretation do you provide? You can think of the analogy of the EKG, where we have a diagnosis or a potential diagnosis shown, but also for the expert, that person's able to go back and see all of the granularity that went into making that interpretation. So same with digital phenotyping, we have to consider how much or little granularity is going to be useful for the end user. And then of critical importance, how do you design protocols around when you have shown results to a patient, what happens when that data doesn't match the client's perception of reality? What if it doesn't match the provider's perception of reality? How do you handle that from a therapeutic standpoint to maintain trust, but also in some cases mitigate or get through issues where you have multiple perspectives that you have to reconcile? And so we like to think of this from a sort of multiple modes of data collection, informing on individual latent constructs that are going to be useful to the provider. So you can think of this as there are many types of raw data that one can collect through digital phenotyping and other modes of data collection, whether it's traditional surveys, video, audio recordings, movement data from a wearable or from a phone, interactions with the phone, location or other sensor data from a phone. And then of course there are other types of more comprehensive phenotyping such as blood tests, MRI scans and data from an electronic health record. All of these can be engineered to extract just the important or the relevant features from those raw data sources that the clinician or the patient might find useful. And then it's ultimately mapping those features into the key constructs that are of relevance to the clinician, whether that may be problems with sleep, executive function, language, affect regulation, stress sensitivity. These are the variables that all clinicians are trying to understand about their patients as they provide care. We can also apply more data-driven approaches to synthesize and detect changes in these key latent constructs. And we can ultimately feed those changes and synthesis into user experiences. And that's what we'll talk about more today. And here again, we're going to use a data or design-centered approach to provide experiences back to not just the patient or the person at the center of the care, but people like their provider, potentially their peers, family members. And this design process is meant to result in clinical support tools as well as potentially research tools, and ultimately even potentially tools that can be used by other parts of the health care organization to anticipate capacity requirements or be able to determine optimal dosing of therapeutic regimens. So today I'll start with a relatively simple use case, which is using actigraphy or data from a wearable device to assess sleep and activity on inpatient psychiatric services. And in addition to capturing it here, we're going to combine it with data from the electronic health record to help, in this case, clinicians make sense of changes that may be occurring when different medications are being provided to the patient that are changing their sleep and activity profiles. And so we'll start here with digital phenotyping on inpatient psychiatric practice. And so why have we done this? Well, today, at least at our hospital at McLean, outside of Boston, sleep and activity data are collected still in a very human resource-intensive and somewhat more paper-and-pencil kind of approach. People walk around with pieces of paper, and they will note how long particular patients have been asleep for. And they'll do that by popping their head into the room and observing breaths. And this is all part of the standard Joint Commission requirements for being able to observe patients by eye to assess their sleep. But we often find that patients and providers find this to be cumbersome. It requires someone potentially interrupting sleep simply to observe it. And so we thought it would be a worthwhile exercise to work with clinicians to devise a way to leverage wearable data to capture the same information and, again, be able to provide that in a way that provides additional utility to the clinicians related to dosing of medications. And so over the past several years, we've collected data from over 500 participants. We use this research-grade wearable called the GeneActiveWatch to collect data 24-7 from patients. And we can collect this raw data, which allows us to use open-source platforms to analyze the data and extract periods of weight and periods of sleep based on an individual's own unique activity profile, recognizing that individuals are going to sleep or have different levels of activity. We can then turn those activity profiles into representations over the entire hospitalization. So you see in this example, this particular patient came in and was getting very little amounts of sleep. But then over the course of the hospitalization, their sleep normalized. And then they had certain days where they had high durations of sleep or very low activity. And so we have several places where this can be useful. One is in simply deriving the distributions of sleep features. So when patients go to sleep, when they wake up, how long they were asleep, as well as important ratios such as the sleep-wake ratio and activity during daytime, such as pacing or being somnolent during the day. So in order to derive population-level distributions for these same parameters, we use data from the UK Biobank, which is a large dataset collected in the UK of about 100,000, more than 100,000 individuals who wore a very similar watch to the one we used on the inpatient unit. And using this approach, we were able to generate a discovery dataset and a replication dataset where we screened all these individuals who were a mix of healthy individuals with psychiatric conditions. We were able to screen out unusable data. And then using this approach, we were able to find the normative pattern of sleep-wake activity over an average week. So obviously there's the daytime circadian changes that we all experience, but also most humans have a week dependence of their sleep-wake pattern. And you can see that here that Saturday and Sunday have a slightly different pattern than the weekday. And that's very consistent across both discovery and replication datasets. But then we were able to do it using this approach to see how sleep and activity change not only over the lifespan, but also in relation to important psychiatric variables such as depression. And so what I'll show you next is how these week-level patterns change as a function of age. So here you can see that as individuals get older, their sleep pattern changes such that they're moving less during the daytime period. And they also have some changes in their sleep patterns. You can see that this is similar, although not identical, between the sexes. But what's relevant for this presentation is noticing the differences between individuals in this sample who endorsed having recent depression relative to individuals who did not have depression. And while there are a number of changes in the sleep onset time and waking time, the really key example I want to focus your attention on is this change in daytime waking activity. So what you can see here is that as you age, your daytime activity is coming down in this roughly linear fashion. But of note that individuals who experience or have experienced recent depression, they're actually moving as if they're about five years older than their current self. And so that's it. That entire normative distribution have shifted to the left in individuals who have a recent history of depression. So we can use that to generate actual percentiles for patients, and so here what I'll go into next is how we begin to generate useful reports for actual clinicians that are able to leverage these kinds of data from the UK Biobank. And so what we've generated and have in a clinical quality improvement project is these radiology-style reports for summarizing behavior. And I say radiology-style reports because what we have is there's two pages to the report. There's this more non-expert summary that appears on the first page of the report. And so this would be something that if you're a clinician who's requested a report and you don't want to get into the weeds, you could still go over this report and get a quick impression. As some clinicians might do, they might skip directly down to this impression that this is the summary of key findings. But if they want to get more information, they can see the average of the key features. So what the sleep duration was, what the percentile was for individuals of that age range, as well as changes over time. So again, this was being deployed on an inpatient service. And so you can see that the recording duration was over about a month and six weeks or so. And you can see then how this automated report summarizes the changes in this individual over time. But let's say you're wanting to get more into the weeds. Well, there's an additional page to this which provides a summary of those key features at the top of the page. It has this description that summarizes where we got the percentiles from, as I was mentioning. And then it shows the reference ranges for those values and the recording duration. So again, this is meant to be a quick snapshot for clinicians who might want to get this information but don't have the time to get into the weeds. We also have a summary of how these different key features changed over time. And again, this is shown in the style of a lab report where you can see the result, see the percentile and see the reference range quickly. There's also a graphical representation. So here we show the same features changing over the entire recording period. But then of critical importance for the inpatient clinician, we also represent the medications that this individual was receiving throughout that same period. So we can see the medications that were being provided. We're seeing the daily doses of those medications. And you can see on red and blue here, the periods, just those highlighted periods where we increased a dose or started medicine or decreased a dose or stopped medicine. And finally, if there's questions about exact timing of medications, we have a representation of exactly when individual medications were being administered. And all of this comes automated through the electronic health record, but it's being shown so that the clinician can see where the sleep epoch occurred relative to the individual dosing of those medicines. And then finally, for those who really wanna be able to have a sanity check, the somewhat, the raw but slightly processed data are shown just below all of this so that the clinician can quickly see whether there are other periods of sleep, such as naps that weren't part of the primary sleep epoch, but might still be relevant for that patient's individual trajectory. So again, the idea here is to take the radiology approach, which is we don't expect everyone to want the expert level report, but for those who are trained in interpreting these kinds of graphs, it's provided for the clinician. And otherwise we have a summary that's provided for the non-expert who wants to just get a snapshot of changes in sleep and activity over the period being assessed. And I'll just throw in here a couple of real world examples. This is an example of someone who had sleep and activity within normal limits. These are all real patient data that have been, there's no identifying data on them. There's also, this is an example of someone who had elevated sleep fragmentation as part of starting a course of ECT, which you can see here in terms of the propofol being administered. Here's an example of someone who had a large number of medications in the context of increasing sleep fragmentation and duration while they were being hospitalized. And finally, here's an example of someone with very low waking activity in the context of elevated sleep duration. And as I'll come back in a moment, this may be a hallmark of what we see in some cases of major depression. So do patients and providers use the report? So we were continuing to use these in a quality improvement context. And so during a period of time between August and December, we had 155 patients admitted to the unit. The first step is to approach individual patients to see whether they would wanna be part of that quality improvement pilot. And we saw that just over half of patients were approached by clinical staff. The one issue was just in a quality improvement project. Not everyone was offered the wearable. This was, as a new initiative, this is often something that's gonna be a challenge of getting staff oriented to doing new things that they haven't been part of. As of the patients who did accept the watch or who were approached rather, about 75% or roughly three-fourths of them did accept the watch. We did see about a quarter of them decline to participate. The most common reason that was stated was hesitation around data sharing and being sensitive to potential privacy concerns. We didn't always have a reason recorded. Again, this is issue of just when staff did approach, we didn't always record the reasons why the person didn't agree. But in total, we did get a number of reports generated for the 42 participants who participated in this pilot. We've gone on to do this in a number of patients outside of the quality improvement initiative. The final thing I wanna mention as a use case for collecting these kinds of data on inpatients is that we can begin to compare how different kind of medications affect sleep under different conditions. Here, what we've been able to do is compare movement profiles on days with and without particular medications for the same individual, the same patient, and then use this approach, so same time of administration, but on a day with a drug versus without a drug, to derive movement profiles as well as effects on sleep. And so we were able to use this pragmatic approach to determine, for instance, that we can see that across the benzodiazepines, which are the most common drugs requested as a PRN medicine, that we see the effect of clonazepam affecting movement profiles for a much longer period of time than, for instance, lorazepam or drazepam, consistent with their pharmacodynamic properties. So you can see that while this is still a relatively small number of observations, in this case, it's about 150 observations of clonazepam and about 200 observations of lorazepam, you can imagine that these kinds of data at a larger scale could help us understand pharmacodynamic properties of a range of psychiatric medications. We were also able to test how these same kinds of drugs that are often used for sleep do affect different sleep parameters. So here, using a Bayesian approach, we were able to see that clonazepam, but not lorazepam, lowered sleep fragmentation in hospitalized patients who received it just prior to bed. And so if you're not familiar with Bayesian approaches, what we can see, essentially, is that the difference between the means of the patients who'd received, or the nights where lorazepam was given versus the nights where lorazepam was not given, you see that this difference of means overlap zero, which means that you can't confidently say that it changed the level of sleep fragmentation. Whereas in clonazepam, that difference of means is well below zero, meaning that clonazepam did reliably shift that sleep fragmentation to be less fragmented. So finally, in summary, inpatient services have many opportunities for integrating wearable data into clinical practice. First, this can replace the cumbersome and often inaccurate sleep tracking methods that are often pain points for staff and patients. Also, it allows us to study the effects of medications and other interventions on movement and sleep at multiple levels. We've focused so far on individual patients providing sleep reports to aid in the treatment and sleep reports to aid in dosing and other clinical decision-making for inpatient providers. But as these approaches are adopted, these can be used at the level of a clinical service, for instance, to see whether certain initiatives such as walking groups or sleep psychoeducation is able to improve sleep across at a unit level or a service level. And then these could be adopted at the level of healthcare systems to improve efficiency, operational efficiency across multiple units. And then as I mentioned with UK Biobank data, this can also be collected at the population level to establish norms in different settings and in different demographics. Nonetheless, there are many challenges for widespread adoption. We still need more population level norms. I didn't emphasize it, but the UK Biobank data is only available for individuals 45 and up. We need other datasets to fill in the gaps of different demographics. But also we need to really critically help providers see the value over and above what the current standard of care may be or standard of practices may be. And this is important because it justifies their effort to learn and integrate the new technology. Even if the previous standard of care is gonna be in some ways deficient compared to the new standard of care, there still is some activation energy required to get over this hump that justifies the benefit in the long run to patients and providers. All right, so that's inpatient psychiatric practice. Now I'd like to shift gears to talk about how we can begin thinking about integrating digital phenotyping into outpatient psychiatric practice. So when we think about integrating digital phenotyping in general, I wanna emphasize that we can really focus on triangulation of the client's condition. So, so far I've really emphasized measuring sleep and activity, but of course, even on the inpatient setting, we're gonna be trying to assess those objective markers in relation to how the participant or the patient themselves feels, as well as how they appear to the clinician. And so this is gonna be even more critical when we're evaluating clients on the outpatient setting where we don't always get access to these different signals. And we certainly don't get to see them continuously over time. And so we like to think about the recovery triangle where we have how the patient feels, which are the subjective markers, how they function, which are the measures of daily living and sociability, and then also how they appear. So how do they look or sound, whether that may be in session or between sessions. And critically, we're gonna be looking at both the coherence among these. So do we see a consistent expansion of the patient's subjective and functional appearance, or do we see that there's somehow a coherence across the modalities that requires further discussion between the patient and the client? So obviously how people feel is gonna be pretty intuitive. It's how does the client's mood or subjective experience change over time. In a research setting, we were able to collect over two years of data from almost 100 patients who were receiving care. And this is one example of a patient who we asked to provide self-reports through an app experience over the course of two years. And you can see here from the representation that this is someone who began the study feeling very positive, endorsing these different ecological assessments. But over around 100 days into the study, they started reporting feeling very negative and starting endorsing feeling hostile, irritable, all of these negative adjectives. And then you can see that you see a very granular trajectory for this patient in terms of their episodes of depression. You can see how the frequency of those episodes starts to increase over towards the second year of this person's assessment. So this level of subjective experiences is something that you might think patients wouldn't be willing to do, but this patient was willing to do it even during the period of time when she wasn't motivated to leave her house much at all. And you see this when we compare this person's ecological assessments with the assessments that she was able to do when she came in to visit the clinic and had a more standardized depression assessment or assessment for positive and negative symptoms. And you can see that her depression scores went up consistent when she was feeling very negative, but just as you might expect when she's coming in less frequently, her clinical scores that she sees in the clinic are not as granular as what we get when she's providing those responses to an app. So we can provide different client-facing and provider-facing visualizations to motivate clinicians and providers to provide this information. So for instance, we can provide an app that's more engaging, that allows the client to reflect on their mood and provide responses on this circumflex model of emotion that is a way to provide them with a more enjoyable way to provide their mood rating than just answering a series of questions. We can then visualize these mood ratings for the provider, and that provider can then socialize that back to the patient to be able to start conversations around when and how their mood changed over the course of a period between visits perhaps. We can also measure function. So these would be questions like how is sleep and activity changing over time? Are they getting out of the house? Are they interacting with others? So here's a different example of a person who was assessed over two years of that same research study. In this example, we can see is that the representation of the individual's locations over a two-year period from dawn till dusk. So you can see that where they're staying at night is this gray location, that's their home location. We can also see other key locations popping in. So for instance, this is when they were hospitalized. This blue location is when they were hospitalized for a period of 10 days or so, different times during the study. You can also see that same location coming in during the week. This is when they're going to their therapist's office and also when they're AA meetings. So you can also see that when we break it down by day of week, that they consistently have therapy on Wednesdays and AA in the evenings. We can also see when they're functional in terms of work. So you can see their part-time jobs coming in is in the T location and this black location towards the latter half of the study. We can also get information about social environments, such as when this person starts to have a relationship and they start to be able, they start to sleep at this other person's house. And you can even see things, subtle things like smoking. So these little red dots is when this individual is going outside and smoking outside their home, all based on location data that the person can collect without having to do anything on their phone. We think this is a really powerful way to extract social and environmental context. So these are seven individuals now with distinct patterns of location. And just by going through this with you briefly, you can see that these seven individuals have very different patterns of sociability and interaction with the outside world. So this sort of representation of location is very attractive for being able to individualize. And so we can begin then in the same approach we would take with a client-facing app and a provider-facing report. We can motivate the client to allow collection of these kinds of data by enabling location access on their phone and justifying it with some psychoeducation. We can also then provide a provider, show the provider how the patient's time at home was changing and whether that might have relevance for their functional status. And this of course would be something that the provider would then discuss with the patient during a clinic visit perhaps. We would also then have similar reports of sleep and movement, all derived from wearables or the phone. And again, this would allow conversations around changes in sleep and activity over time that could have clinical relevance. We also want to integrate this with how the person is appearing. So are there any signs of psychomotor changes relative to the person's baseline? What about the movement patterns between sessions? How do they look? What about within sessions? How do they appear when they're within session? Also changes in language that could have clinical applications. So these aren't directly subjective, but they are things that clinicians are going to want to assess to get a holistic or comprehensive picture for how the patient is doing. So here's an example of that same person who I mentioned brief earlier who's having repeated episodes of depression. And so this is their movement traces over that same two-year period using that same watch that I've referred to in the inpatient work. And so what we can see here at the high level is that this person was able to wear the watch fairly consistently over that two-year period. But I want to highlight is we see changes in their sleep, particularly these abrupt changes in their sleep and activity are occurring multiple times during participation in the study where we see that sleep is getting longer. She's waking up much later in the morning and she's much less active during the day. Similar to what I showed with that UK Biobank data where individuals who are depressed are less active while they're awake. And so we can also look at movement patterns within sessions. So we can measure data from a camera and be able to quantify a patient's expressivity, how they move their face, as well as how their pose data changes during a session. And these data can be represented relative to normalized groups, such as larger samples of individuals or relative to themselves. So if they've been with you for multiple times, you can see how their movement patterns, how their affect, how their pose dynamics, whether they're leaning out or leaning in to a session, how all those things change relative to themselves. So again, this allows us to fill in that recovery triangle and get a sense quantitatively for how this person appears in different contexts, both at the level of time over two years, but also over periods of 30 minutes of a particular session with the clinician. And so then how do we synthesize all of this together? Ultimately, the clinician is going to want to bring all these data types together to make inferences. And so here I'm bringing back that individual who I showed you their self-report data and how it compares then with data from the watch. And what we see, or it may be hard to appreciate here, but the timing of the changes in her sleep and activity coincided with the changes in her positive to negative affect. And so again, by being able to visualize this for the clinician, we believe that synthesis where you can combine mood reports with perhaps habit changes, changes in consumption, with changes in sleep and activity and changes in location is how we begin to give the clinician and the patient that 360 sense of how someone's doing that can help the patient and the provider identify potential insights around ways, influences, causal factors, but also help reconcile when the subjective and the objective don't seem to match so that they can understand this better. So in summary, digital phenotyping and outpatient psychiatric practices have many opportunities for integrating smartphone and wearable data that are very relevant for the outpatient psychiatrist. They provide a more ecological way for clients and providers to convey, for clients to convey their symptom patterns over time and in relation to events, as well as stressors and context factors. They also allow a triangulation of client recovery through distinct modalities that can inform in ways that can be either coherent or discrepant, leading to a more comprehensive assessment that can augment and even mimic the natural clinical discovery process. We as providers are often trying to triangulate between collaterals, what the patient says and what we may be hearing from a spouse or a family member, or even their patient's own words. So the patient may say one thing, but then act in a slightly different way in terms of the way that they express something in their tone. And so collecting these data and having that triangulation approach or the latent construct approach allows us to formalize that through data visualization. Again, with the goal here not of usurping the clinician's role, but really to help provide more quantitative resources so that the client and the provider can be on the same page with regard to different sources of data. And finally, I think this is really critical, is that care transitions are often one of the challenges for providers and clients. And so being able to transcend individual client-provider relationships using both subjective and objective data sources is a way to mitigate this if you have someone who is maybe moving and is going to see a new provider, or if you have somebody covering while you're on vacation and you need someone to be able to see your patient without having to start from scratch. But also for the client, being able to have that long-term treatment history. So over even years, they would be able to see how they've responded to different medications and therapies and contexts that could be useful for a future provider. And there's still much to learn about client and provider preferences. And so we really want to emphasize that this discovery process is going to require more providers and clients being willing to try these methods and incorporate them into practice, because I think it's not simply going to be software developers developing these kinds of tools. It really has to come with proactive engagement from clinicians as well as patients who are willing to help iterate on finding the best use cases for digital phenotyping. So again, just in summary, what we're trying to do with these approaches is to triangulate how patients are feeling, functioning, and appearing through different kinds of data visualization. Here today, I've just mentioned a few. We feel that there's many opportunities on inpatient services to do this, where you can directly measure how particular medications and other therapies are affecting movement, as well as other types of, you know, whether it's movement through a wearable or movement through clinical encounters. We feel that synthesized reports, where you can combine how someone's feeling and functioning and appearing, can be of great value to clinicians and patients as well. And finally, being able to visualize simply how someone is doing, or rather where someone is going or how they're sleeping, over time, allows for more conversations to be generated around linking those together in a way that's still privacy preserving for the individual client, but can be very useful when the individual client and provider can collaborate on the locations that have direct relevance for that interaction. So I'll wrap up here. I want to mention that we have a technology and psychiatry conference that will be held December 7th in Phoenix, Arizona. This will be on the day prior to the American College of Neuropsychopharmacology meeting that will happen in the same hotel. So if anyone's going to be at ACMP or just wants to come and discuss more of these issues in person, please don't hesitate to register for the conference at mclean.org slash tips, and I'll be happy to answer any questions. And please don't hesitate to, I see there's a couple questions in the Q&A, but feel free to drop more questions in there if you do have other questions. Okay, so first question from Joe DiBella, what might be the cost of the wearables? Will insurance pay? So that's a great question. So the wearables that we use come in different flavors. An Apple watch is around $300. The GeneActive watch that we've been using in the hospital is also around $300. The Oura ring is around $400. You can get cheaper wearables. This kind of comes back to the question of, do you have individuals wear their own device, or do you have, or do you provide them with the wearable as part of their treatment? On the inpatient service, it might make sense to provide the individuals with the wearable. On the outpatient service, there may be some benefit to having the individual use their own wearable. So in some of the graphs that I was showing previously, we did have individuals using their own wearables, and that was either an Apple watch or an Oura ring. The advantage there is that they're used to doing it. There's no additional cost to doing that. And so it can be good for capitalizing on what someone's already willing to do. It can be a challenge. It's another challenge to try to get someone using or wearing a device that they're not already using as part of their daily life. But that's a great question. The question of whether we'll insurance pay, I think once we can show that there is a value back to patients and providers, that we can get insurances to potentially cover the actual cost of the device. I'm not sure we're there yet, but I think that there's potential as we begin showing how this can help either on inpatient or outpatient services that we could get to that point. So the next question is from Fabio Oresta. So could you elaborate on the kind of analysis performed on the data? You said OS software programs with statistical methods. So I mentioned a number of different approaches. So I can tell you that the, I will also put this into the chat, but if I go, yeah, so there's different analyses that I'm showing here. This visualization that I showed a few times uses open source data from this watch. That watch is that watch is providing raw accelerometry data. And so we've developed in my lab, open source software that allows you to take that raw accelerometry data and turn it into these sleep estimates. And so there's something we call DP sleep, which is a open source pipeline that I'm going to try to drop the link into the chat for anyone who may be interested. We also have a manuscript that's been published that describes this approach. And so I'll also put that into the chat here. Now for the other examples, the data that I'm showing was directly from either the Apple watch or the Aura ring. In those cases, we're reliant on Apple and Aura's algorithms to generate the features. The downside of that is it doesn't really adhere to this thing I mentioned earlier, which is that we really would prefer open source tools and algorithms so that we know exactly how the different features were computed. The challenge with that is that a lot of the device manufacturers don't release their algorithms. And so we have a rough sense about how Aura and Apple compute the different sleep stages, for instance, but there is no manual to derive that. So I think that's going to be a key challenge of how to integrate the data from proprietary algorithms into clinical care. There's another question here from Alan Mastry. I know in Europe, I think Germany, they're using the analysis of patient language to assess if a schizophrenic patient is doing well or not. Is this something you may consider adding? Yeah, so we have done work using language analysis. I didn't emphasize it today because I was focusing on data from smartphones and wearables. Although with smartphones, you can use audio diaries. We've done a lot of that kind of work where you can collect even daily samples of how someone's doing from an audio diary. You can also analyze the session recordings from interviews. And so we published on that as well. I can drop a couple of links into the answer here. But that's absolutely something that we would be very interested in. And there's a whole set of papers that go into this. I'll just type a couple into the chat. One from our group and then an example of linguistic findings that are specifically about schizophrenia. And I would say that's something that language in particular being such a core feature of psychiatric evaluation, is a very fruitful way to assess even computationally how someone may be doing. So linguistic features is something, whether we're talking about the words that someone's using, or things like the coherence of someone's speech, or things like errors. So how many verbal errors is someone making? In that example that I posted in the Q&A. And in the second example, we studied basically how spoken language is different in patients who have a psychotic condition. We've also looked at this in individuals who have a disorganization. So specifically looking at this feature of disorganization in schizophrenia. And we found that there were features of language that correlated quite tightly with disorganization. So I'll put that also in this answer here. And yeah, just following up that this could be used for mild cognitive impairment versus depression versus dementia. I completely agree. I think what I'm showing you today is really just meant to inspire those of you who maybe work in a different area of psychiatry, neuropsychiatry, that this kind of work can and may be useful in a whole range of conditions. And I think that's where, just to go back to my very first slide, if I can do that. Well, that'll be too challenging. But that first paper of being rounded to reality, I think is the key here for all of these conditions. You want the subjective response of the patient to not be the only source of data that the clinician has to go on. I think that generates more confidence in the treatment process by both the provider and the patient. So I hope that resonated. Okay. Well, I think we're about wrapping up. So I'm going to thank you all for your attention and your excellent questions. I hope this was informative and inspiring. And please don't hesitate to contact me either at my email or over the X platform or LinkedIn if you have any further thoughts or questions. And please do consider coming to our TIPS conference in just about a month in Phoenix. And then finally, I do want to say that we're going to have in our next presentation, Kim Bullock is going to be coming back next to talk about the practice and potential of using extended reality in psychiatry. And I think that will be very interesting for everyone. So please join us in our next lecture.

digital phenotyping

psychiatric practice

smartphone data

wearable technology

patient monitoring

mental health

privacy issues

data analysis

treatment efficacy