Teen mental health crises are rising — but what if providers could intervene before symptoms fully take hold? In this conversation, Duke University School of Medicine's Jonathan Posner, M.D., professor of psychiatry and behavioral sciences, and Matthew Engelhard, M.D., Ph.D., assistant professor of the Department of Biostatistics and Bioinformatics, break down the "Duke PMA" — an AI-powered predictive model designed to identify adolescents at high risk for psychiatric illness. They explore how sleep, device use, and a myriad of other factors shape mental health risk, and how this technology could transform prevention, especially in underserved communities.
View Transcript
00;00;00;01 - 00;00;19;04
Tom Haederle
Welcome to Advancing Health. Many a parent has wondered what's really going on inside the head of their beloved teenager. As we hear in this podcast, a new predictive tool helps answer that question and can flag teens at elevated risk for developing a psychiatric illness in the near future.
00;00;19;06 - 00;00;41;25
Jordan Steiger
Hi everyone, and welcome to the Advancing Health Podcast. My name is Jordan Steiger, and I'm the director of Behavioral Health and Violence Prevention at the American Hospital Association. I'm joined today by Dr. Jonathan Posner, who is the J.P. Gibbons Distinguished Professor of Psychiatry, and Dr. Matthew Englehard, the assistant professor of biostatistics and bioinformatics, both at Duke University School of Medicine.
00;00;41;26 - 00;00;46;11
Jordan Steiger
So, Dr. Posner and Dr. Englehard, thank you so much for being here today.
00;00;46;14 - 00;00;47;18
Jonathan Posner, M.D.
Thanks for having us.
00;00;47;20 - 00;00;48;29
Matthew Engelhard, M.D., Ph.D.
Yeah, thanks for having us.
00;00;49;05 - 00;01;07;02
Jordan Steiger
We're really excited to talk about the work you're doing. This is a little bit different of a topic that I think we usually have on our Advancing Health podcast, more kind of focused on the research and what could be coming down the pike in terms of behavioral health. So I'm really, really excited to share everything with our membership today that you've been working on.
00;01;07;02 - 00;01;25;23
Jordan Steiger
I would love for you to just tell the audience a little bit about what you do, the Duke predictive model of Adolescent Mental Health, or we'll call it the Duke PMA. To shorten that a little bit on this podcast and just tell us what this tool is, who it's for and what it can do. Dr. Posner, why don't you get us started with this one?
00;01;25;29 - 00;01;54;27
Jonathan Posner, M.D.
So we designed the Duke PMA to be able to automate the assessment of psychiatric risk in teenagers. So essentially what the tool does is it can identify which teenagers are most likely to develop a psychiatric illness within the next 12 months. And it does that by using pretty standard questionnaires that can easily be collected in a primary care setting.
00;01;55;02 - 00;02;15;13
Jonathan Posner, M.D.
And then not only does it tell you who's at risk, but it also tells you what's contributing to that risk. So as an example, let's say I am a pediatrician and I'm about to evaluate a 14 year old. I might only have 5 or 10 minutes with that kid, and it's not enough time to do a thorough psychiatric assessment.
00;02;15;15 - 00;02;42;19
Jonathan Posner, M.D.
But the Duke PMA will automate that for me, and it will tell me that the kid that I'm evaluating who seems stable today, actually has an 80% chance, or a 90% chance of developing a psychiatric illness in the next year. And a key contributor to that might be poor sleep. So as the pediatrician, if I can get that kid sleeping better today, I might be able to prevent him from ever developing a psychiatric condition.
00;02;42;21 - 00;02;58;01
Jordan Steiger
Wow. So that could be transformative for adolescent behavioral health. That's really incredible. You mentioned sleep is maybe one of the predictors. And I know in our conversations you've said that's a pretty significant predictor. What are some of the others?
00;02;58;03 - 00;03;25;22
Jonathan Posner, M.D.
So we looked at a bunch of different factors that relate to psychiatric risk. And so some of the other contributors are things like conflict within the family, the school setting, some demographic features, a history of childhood adversity and family history of psychiatric illness. So some of them are things that we could - that are modifiable, that a physician could say, okay, I'm going to try to intervene on that.
00;03;25;25 - 00;03;30;29
Jonathan Posner, M.D.
Other things, like family history of psychiatric illness are more challenging to intervene on.
00;03;31;02 - 00;03;40;00
Jordan Steiger
Absolutely. But that does give the clinician a great opportunity, at least to get ahead of it and start intervening early. And it sounds like that really is the goal of this work.
00;03;40;05 - 00;03;48;02
Jonathan Posner, M.D.
That's right. Yeah. The goal is really to move psychiatry into a model that's much more proactive and preventative.
00;03;48;05 - 00;03;58;14
Jordan Steiger
Dr. Englehard, tell us a little bit more, just why AI tools? What is the most transformative thing about using AI tools like this to predict mental health before symptoms appear?
00;03;58;16 - 00;04;22;04
Matthew Engelhard, M.D., Ph.D
Yeah. So I think I think AI tools are really good at figuring out what in a specific type of information ends up being predictive. So compared to more traditional models, I think of them as being able to, you know, extract all of that signal. And when you couple these tools with a really large data set like the one we've been working with, which is coming out of the Adolescent Brain Cognitive Development study, it really enhances our ability to do that.
00;04;22;04 - 00;04;51;19
Matthew Engelhard, M.D., Ph.D
So first and foremost, it's about just sort of being able to pull as much predictive power as we can. At the same time, AI tools are really good at being able to combine different types of information together to say sort of what is what is shared between this questionnaire and this questionnaire and this questionnaire, and what is distinct between them and being able to draw on those individually and together again, to enhance our ability to make the best predictions that we can.
00;04;51;26 - 00;05;05;12
Jordan Steiger
That makes total sense, and I think it would be great - I probably should have started with this. Can we tell the audience a little bit about where you're at in your research and just this whole work? It sounds like you've done quite a bit already to show that this model is really impactful.
00;05;05;15 - 00;05;29;16
Matthew Engelhard, M.D., Ph.D
So we've already developed our model, and it's a matter of making sure that this model, that we have everything in place so that this model can reach as many people as possible. So we're really careful about making sure that our model is what we call generalizable, that we've built it in this this study, which is intended to be which is nationally representative, but it doesn't include everyone.
00;05;29;18 - 00;05;52;01
Matthew Engelhard, M.D., Ph.D
Right. We want to make sure that we've built it in this particular study, and now we have a system that we're confident is going to be able to work for everyone across the United States, regardless of what region they're part of, regardless of their demographic background, and even regardless of some of the specific data collection practices that might be in place where they receive their care.
00;05;52;01 - 00;06;08;18
Matthew Engelhard, M.D., Ph.D
So right now, we're in a phase where we're getting ready to take our model that we've built and to validate it in a bunch of different rural communities and in several states. But before we do that, we're making absolutely sure that it is going to work effectively for everybody.
00;06;08;26 - 00;06;27;25
Jordan Steiger
That makes total sense. And I think going into rural communities, you know, presents such an opportunity to expand access to care, make sure that we're focusing in on that prevention aspect a little bit. But tell us a little bit more about that, Dr. Posner. Maybe just why you're starting to test in these rural areas and what's next in terms of that?
00;06;28;01 - 00;07;07;05
Jonathan Posner, M.D.
Yeah, I mean, one of the key motivators in developing the PMA was to try to make psychiatric assessment much more efficient by automating the process. The availability of mental health providers across the boards is extremely limited. But then if you go into rural areas, the problem is even more severe. So the reason why we wanted to bring the Duke PMA to rural communities is for exactly that reason, that that the availability of providers is so limited in those communities that we thought the Duke PMA could be particularly helpful in that setting.
00;07;07;13 - 00;07;26;12
Jordan Steiger
I was thinking about the workforce aspect as you both were talking about this, and just how that can, I'm sure, really, really help clinicians in these areas that maybe don't have the time, you know, to spend with the kids that really need the help and the mental health support, you know, or they might not catch something that's happening because they don't have the time.
00;07;26;15 - 00;07;42;24
Jonathan Posner, M.D.
Yeah, I think that that is really a huge factor in that we're not making the claim that the Duke PMA could do an assessment better than a trained professional, it's just that it can do it much more efficiently. And those trained professionals aren't available in lots of settings.
00;07;42;26 - 00;08;02;14
Jordan Steiger
Absolutely. It's an additional tool for trained professionals to use. Thinking about how this is going to play out in a real world clinic setting in some of these rural communities. What do we think? I know we just talked about workforce and expanding access, but some of those other big opportunities and maybe some of the challenges that might arise in practice.
00;08;02;17 - 00;08;32;22
Jonathan Posner, M.D.
One of the big opportunities that I would see is for this to become a standard part of a pediatric assessment. So the comparison that I like to use is with cardiac risk. So essentially all adults in the US get screened for cardiac risk, for cardiovascular disease. And then the risk gets tracked over time. And the goal is to intervene well in advance of anyone ever having a heart attack for example.
00;08;32;24 - 00;08;51;01
Jonathan Posner, M.D.
And so we would we would love to see something similar to that with mental health risk, so that kids are identified as being high risk before the conditions have really taken root. And the earlier we can intervene, our chances of being able to help these kids is substantially higher.
00;08;51;07 - 00;09;06;22
Jordan Steiger
Absolutely. And, you know, walk me through before we go to maybe some of the challenges, say a child is identified as having a risk for sleep. Let's use that as the example because we brought it up earlier. What happens after that? Just tell me kind of walk me through the clinical process after that.
00;09;06;25 - 00;09;29;04
Jonathan Posner, M.D.
So fortunately with sleep, we have really good evidence based interventions that can help kids sleep better. So depending on the setting and the community and the availability, being able to refer that child for a psychotherapy to address poor sleep would be one natural route to go.
00;09;29;06 - 00;09;41;14
Jordan Steiger
The reason I'm asking that, I guess, is I'm thinking about these rural communities that might not always have the resources available to kind of refer to those additional services. So I'm, I guess thinking about challenges and maybe some of that.
00;09;41;22 - 00;10;08;08
Jonathan Posner, M.D.
Yeah, I think no, it's a great point. One of the things that that Matt and I have talked a lot about is we would ideally love to be able to bring the Duke PMA to schools, because that's where kids are. But one of the challenges of that is that, you know, the last thing we want to do is identify a child at being high risk, but then not having any resources for them, not being able to refer them to a provider.
00;10;08;09 - 00;10;30;03
Jonathan Posner, M.D.
So by working within primary care settings, they'll at least be partnered immediately with a medical team. And some of the interventions can be pretty straightforward. So for example, with sleep, talking to the child and talking to the family about better sleep hygiene, better sleep habits that could actually go a long way and doesn't require a separate referral.
00;10;30;06 - 00;10;35;18
Jordan Steiger
No, absolutely. That's a great point. Dr. Englehart, what do you think about challenges and opportunities?
00;10;35;20 - 00;10;57;05
Matthew Engelhard, M.D., Ph.D
Gosh, we have so many opportunities to make an impact here and so many opportunities to get more and more sophisticated in our ability to perform this work. So I think being able to connect with the schools is a tremendous opportunity. We also see there's an opportunity in being able to understand the device use component of the picture here.
00;10;57;07 - 00;11;23;24
Matthew Engelhard, M.D., Ph.D
How do you how do adolescents use of their digital devices? How does that contribute to some of the development of mental health distress that we're seeing? And then how can we in turn interpret information coming from those devices to help us intervene again? Another opportunity that we're really excited about is that as part of the work that we're working on now, we're going to be collecting information about home environments as well.
00;11;24;00 - 00;11;58;20
Matthew Engelhard, M.D., Ph.D
So this is drawing on research that we've been engaged in for a while now to understand relationships between different personal environments over the course of the day, and different health risks that individuals might be exposed to. And it's known that there are relationships between aspects of home environments and adolescent mental health. But we have the opportunity to work with adolescents and their parents to document different components of home environments and understand again how that relates in turn to the mental health picture, and maybe to make recommendations along those lines as well.
00;11;58;21 - 00;12;18;28
Matthew Engelhard, M.D., Ph.D
I think I would also add one more, which is that we've mentioned the importance of sleep, which has already come up, and I think we're very excited about the opportunity to put that knowledge to use and understand whether we can, in fact, use our model to identify children that are at high risk of worsening mental health because of their sleep, at least in part.
00;12;19;04 - 00;12;29;13
Matthew Engelhard, M.D., Ph.D
And to see if intervening on sleep, giving them some actionable recommendations to change sleep patterns, as I think Jonathan mentioned, whether that can indeed move the needle for those kids.
00;12;29;15 - 00;12;50;26
Jordan Steiger
That makes total sense. And, you know, thinking about teenagers to in the home environment, I feel like that's really, really important, thinking about mental health. And, you know, you bring up your personal device used to and I think that's another fascinating thing. I'll be curious to follow your work and see what comes of that, because we know that's a huge component in the way mental health kind of manifests for kids these days.
00;12;50;28 - 00;13;06;24
Jordan Steiger
Tell me a little bit more about what the Duke PMA can be used for. Like, is this a universal tool to predict mental illness kind of across the spectrum of different disorders that children can experience? Or is this really focused in on just a few different things?
00;13;06;26 - 00;13;35;04
Jonathan Posner, M.D.
Yeah, it's a great it's a great question. And what it actually predicts is something called a P factor, which is a global measure of mental illness. So it essentially encompasses risk for the full spectrum of psychiatric disorders. And one of the reasons why we chose the P factor is because if it was specific to a predicting risk for one disorder, its utility would be much more limited.
00;13;35;05 - 00;14;00;12
Jonathan Posner, M.D.
So, for example, if it predicted development of bipolar disorder, that would still be useful. But the number of teenagers that you could use it with would be also very limited. By using a global measure, it can be applied essentially to all kids. We're working to make it as user friendly as possible, so the entire battery of questions are being put onto an iPad.
00;14;00;20 - 00;14;21;11
Jonathan Posner, M.D.
So in theory, a family comes to their primary care clinic, and while they're in the waiting room, they could be handed the iPad, answer all the questions, and then the risk prediction and the profile would then be fed forward to the clinician who's about to see them. In the current version of the tool to complete all the questions,
00;14;21;12 - 00;14;41;15
Jonathan Posner, M.D.
it probably takes about half an hour. And that's one thing that we're really focused on, is to try to get the assessment down to as short as possible so that families could complete it in say, ten minutes, rather than 30 minutes. And so what we're trying to do is to limit the number of questions while still maintaining the predictive power.
00;14;41;15 - 00;14;44;07
Jonathan Posner, M.D.
And that's an active area of work for us.
00;14;44;10 - 00;15;06;26
Matthew Engelhard, M.D., Ph.D
And I would add that from the AI perspective, I think one of the reasons we've been so successful in doing something that is, is new here is the fact that in focusing broadly on all these different pathologies, we've really enhanced our ability to learn about relationships between general psychopathology and all of these different types of predictive features that we've talked about.
00;15;06;28 - 00;15;19;05
Matthew Engelhard, M.D., Ph.D
So, you know, it puts us in a position to give people the help they need. It's also enhanced our ability to understand what risk factors contribute.
00;15;19;06 - 00;15;43;02
Jordan Steiger
So I know we all are in the world right now where everybody is talking about AI constantly. And I think AI and mental health is a really hot topic right now. I know a lot of our members at the AHA are really kind of trying to understand and consider how they can incorporate more AI, you know, predictive models and all sorts of different things that you all know much more about than I do.
00;15;43;02 - 00;16;00;03
Jordan Steiger
But a lot of the audience on this podcast is hospital leaders. So if they're listening and thinking, how can I do this at my organization? How can I get something like the Duke PMA in, you know, for my primary care clinics or psychiatric clinics for kids? What would you tell them?
00;16;00;06 - 00;16;23;11
Matthew Engelhard, M.D., Ph.D
Well, first, first I would tell them, we're always happy to engage with folks that are interested in getting connected on this. So happy to talk to specific individuals who might like to adopt some of this technology. Thinking about interest in AI broadly, AI is so many different things. I mean, I think we're in an era where people think of ChatGPT and other chatbots as being synonymous with AI in some ways.
00;16;23;11 - 00;16;51;06
Matthew Engelhard, M.D., Ph.D
And I would say that this is a little bit different. I mean, when we talk about AI, we're thinking more broadly about systems that are used to make sense of data to understand patterns and data, in this case, patterns that help us understand, again, who's at greatest likelihood of being in mental health distress. So we're thinking of sort of AI in a broad versus a narrow sense here, and how to get this kind of technology up and running at a particular institution.
00;16;51;09 - 00;17;15;11
Matthew Engelhard, M.D., Ph.D
There is institutional know how that is, that is required. But we are doing everything we can to make this tool as broadly accessible as possible. Our model is in fact available in a public repository right now, and we are happy to work with folks to think about how they might take the publicly facing resources that we've made available and put it to work toward a specific use case.
00;17;15;14 - 00;17;34;18
Jordan Steiger
Thank you both so much for being here today. I think this is such a great opportunity just to share some of the really innovative work that's coming out of Duke right now and, you know, spreading across the country. I think this is just an incredible example of how we can kind of get ahead of the mental health crisis that we are all hearing about and experiencing every day.
00;17;34;18 - 00;17;40;03
Jordan Steiger
So thank you both for the work that you're doing to make mental health care better for people all across the country.
00;17;40;11 - 00;17;43;28
Jonathan Posner, M.D.
Absolutely. Thank you for having us and for your great questions.
00;17;44;02 - 00;17;46;04
Matthew Engelhard, M.D., Ph.D
It's been a pleasure. Thanks for having us.
00;17;46;07 - 00;17;54;29
Tom Haederle
Thanks for listening to Advancing Health. Please subscribe and rate us five stars on Apple Podcasts, Spotify, or wherever you get your podcasts.



