Can AI Help Prevent PTSD?

What if a simple conversation in the emergency room could reveal who’s most at risk for PTSD before symptoms even begin? Katharina Schultebraucks, PhD, shares her innovative work on using machine learning to forecast mental health outcomes and explains how AI could revolutionize how we detect, prevent, and treat psychiatric disorders. Dr. Schultebraucks is Co-Director of the Computational Psychiatry Program and Associate Professor in the Department of Psychiatry and Population Health at NYU Grossman School of Medicine.

🔍 Topics Covered
00:00 Introduction
00:23 Current Work and Research Focus
01:29 Objective Measures in Psychiatry
02:50 Predicting PTSD Risk
04:28 Early Preventive Interventions
05:47 Machine Learning in Mental Health
09:49 Challenges and Surprises in Research
22:46 Burnout in Emergency Department Providers
27:17 Precision Psychiatry and Future Directions
29:35 Conclusion

📚 Related Resources
-Katharina Schultebraucks, PhD
-Computational Psychiatry Program in NYU Langone’s Department of Psychiatry
-“Dissecting racial bias in an algorithm used to manage the health of populations” by Ziad Obermeyer, et al
-PTSD Treatment at NYU Langone Health 

🔔 Subscribe for more conversations at the intersection of mental health and innovation.
📢 Like, comment, and share if this episode sparked your thinking.

Visit our website for more Insights on Psychiatry.
Watch this episode on YouTube
Executive Producer: Jon Earle

Transcript

DR. KATHARINA SCHULTEBRAUCKS: Burnout is really associated also with like a lot of cardiovascular risk factors. Such a higher mortality in physicians than the general public. So when I heard all of this, I was like, okay, we need to do something to better identify those at risk. 

DR. THEA GALLAGHER: Welcome to the Insights on Psychiatry podcast. I'm Dr. Thea Gallagher, and today I have the pleasure of interviewing Dr. Katharina Schultebraucks. Dr. Schultebraucks is an Associate Professor of Psychiatry and Population Health at the NYU Grossman School of Medicine and NYU Langone Health. She's also the co-director of the Computational Psychiatry Program. Thank you so much for being with us today. 

DR. KATHARINA SCHULTEBRAUCKS: Thank you so much for having me. 

DR. THEA GALLAGHER: So can you tell us a little bit about the work that you're doing currently? 

DR. KATHARINA SCHULTEBRAUCKS: Yeah. So I have several studies ongoing, but they all have one overall aim in common. The overall aim is like there was a big frustration of me when I conducted research that there's like a big translational gap between what we aim to do and then also what we really transfer to the patient and what is really beneficial for the patient. So overall, all my studies try to develop tools that are really implementable in real-world settings and help us better to identify patients at risk and also better to select the treatment that is really beneficial for the patient. And why I became interested in it is because psychiatry is still like I think the only medical field where it completely depends on the patient, what kind of symptoms they present to us, but also on the clinician, how they interpret the symptomatology and then select individualized treatment or select treatment and also predict the progression of the symptomatology and we know that only like one-third of patients really respond to those treatments. And it's really difficult for us to predict so far to know who will respond and who will not respond and that's how I became more and more interested in that area I'm working in. 

DR. THEA GALLAGHER: So it seems like going from less self-report to more objective measures and maybe objective measures that don't rely on the patient perspective or experience as much?

DR. KATHARINA SCHULTEBRAUCKS: We want to identify objective markers of this subjective experience of the patient, but also still consider the subjective experience of the patient because what we also found in previous studies is that the subjective response is still super important in identifying individuals at risk. For example, we had a study in collaboration with several colleagues here at NYU and also at Emory University where we collected or enrolled patients after trauma in the emergency department. And we developed an algorithm using, for example, objective data from the electronic health records to predict who is at risk for long-term PTSD through one year after trauma. But what we found is that really also the subjective response, how patients experience that is highly important. However, it is really important also to see if we find good objective proxies of this subjective response. And this is what we try currently in several studies.

DR. THEA GALLAGHER: Can you tell us a little bit about how good we are at predicting that? And then what are some of these predictors or how you're doing that work? 

DR. KATHARINA SCHULTEBRAUCKS: A lot of my work is really to see early after trauma, can we already identify relevant factors of who is at risk of later on developing PTSD. Why I'm very interested in that field and not already when patients already have PTSD is because PTSD is very difficult to treat, especially when it's very chronic and we know that there is a high percentage of patients who are not responding anymore. So I'm very interested in can we already implement more preventive interventions in a subgroup of individuals who really need that? So currently, what we are doing is building a digital tool that can be also associated with the electronic health records and then flag patients at risk for later PTSD risk. 

And what we do is each patient who is coming to the ED is asked like a standard question by each provider, which is what happened. And we video and audio record this narrative and extract objective markers of stress response from those narratives using large language models and digital phenotyping and we are able to use this information and accurately identify up to one-year PTSD risk in those individuals using this type of data. 

DR. THEA GALLAGHER: And then when you have that data, what are you doing? Are you kind of implementing something? Because we know critical stress debriefing has been found to not be as effective. So there's like a timeline probably nuanced there. 

DR. KATHARINA SCHULTEBRAUCKS: You mean how we implement then early preventive interventions? 

DR. THEA GALLAGHER: Yes. 

DR. KATHARINA SCHULTEBRAUCKS: What has been found with early preventive interventions, especially in PTSD, is that there is a huge heterogeneity in if it works or not. And we saw also one of my close collaborators did some analysis and found that really there's a subgroup of individuals who respond to those early preventive interventions such as psychopharmacology but also psychological treatments. And that's actually what we aim to see who is at long-term risk because only those respond to it. And there are colleagues from Emory who did early exposure therapy in the emergency department or early after. And also, again, they found it is helpful for a subgroup of individuals. And this is the next step. We now are able to identify those individuals who are at long-term risk. And now in the next step, we are working on to see for whom, what kind of treatment would be most beneficial. 

DR. THEA GALLAGHER: And with the people when you're identifying it, what are some of the factors that are most relevant or that are the predictors of who will develop PTSD?

DR. KATHARINA SCHULTEBRAUCKS: What we found in previous studies and also now, it's really a combination. So that's why I became fascinated in the field of machine learning because it will be never one single factor. And also, it is more what we see is like the immediate, as I said, biological stress reaction, what we can see in vitals, we can see it in heart rate we can see it also in the immune and inflammatory response after trauma, that this is highly informative, but all in combination also with then psychological risk factors such as what you experienced during trauma, which we also know is relevant and we know that early PTSD symptomatology is relevant, but both separately are not sufficient to accurately identify them. It's more a combination out of biological risk factors and psychological risk factors together that give us a more holistic picture of the risk of an individual.

DR. THEA GALLAGHER: And are the psychological risk factors based on maybe prior conditions? Like I don't know if it's anxiety or depression or even just cognitions. I know sometimes the cognitions about the self and I'm incompetent or I kind of contribute…the guilt even with moral injury and things like that. Are those some of the psychological factors that you're referring to? 

DR. KATHARINA SCHULTEBRAUCKS: Those are also relevant. But what I was now referring to in emergency department patients, what we found, what we examined, was particularly early post-traumatic stress symptoms early after trauma within like hours after trauma. But also, we did another study together with colleagues in Israel where we also looked at early neurocognitive functioning because there's always the question, is it more a vulnerability factor or is it already altered after the traumatic experience and then increasing your risk? And we found that early after trauma, really the neurocognitive performance, especially flexibility, is highly relevant in predicting long-term risk in individuals. So again, I think all of those information give us relevant information that can increase their accuracy in how we identify individuals at risk and also help us what kind of treatment might be most beneficial for the patient. But I think we need to move away from, on average, this is relevant because we see there are different clinical phenotypes with different symptomatology. Like some have cognitive issues early before or after trauma, some have more pronounced physiological stress reaction also long-term and so there's different treatments that might be relevant for those patients.

DR. THEA GALLAGHER: And then to know which to target or which to target first or what might have the most impact in success of healing. And I think one of the things that's so interesting with machine learning is that it's kind of taking the heterogeneity, taking all the complexity, and actually making sense of it to make something more simple, relevant, accessible. And is that kind of the hope of the work that you're doing? 

DR. KATHARINA SCHULTEBRAUCKS: Yeah. So this is definitely to more and more disentangle the complexity and it's not only in the field of PTSD, it's across many mental health disorders that it is not only complex with heterogeneity with regard to the risk factors that are contributing to it and protective and vulnerability factors, but also with regard to how the disorder manifests itself and then also how it develops over time. So I think for all of this, machine learning can help us because we are here able to combine different predictors. We are able to identify also more dynamic patterns over time and see really what kind of symptomatology maybe is persistent over time and which are responding to the treatment. So then over time, what kind of treatment is the better approach for a subgroup of patients than other?

DR. THEA GALLAGHER: Has there been anything that's been surprising or any kind of predictors that you actually found were irrelevant? 

DR. KATHARINA SCHULTEBRAUCKS: Yeah. We found definitely a lot of what was for me in the beginning surprising was also that the relevance, for example, again in the emergency department setting is like factors about the emergency department are really relevant from at what time the patient was admitted, the waiting time. It all makes sense, but it is like really relevant.

DR. THEA GALLAGHER: In that short time frame?

DR. KATHARINA SCHULTEBRAUCKS: In that short timeframe. And I think that is because I'm very interested in besides developing easily deployable tools that are understandable from clinicians and patients and can help us to be more accurate, I'm also very interested in can we identify modifiable factors? For example, I know we cannot change the system and change our emergency departments, but maybe when we would be able to identify the subgroup of patients for whom the environment is very, very stressing and more increases their vulnerability, can we maybe, in that moment, already intervene and give them a more calmer environment or give them more information on the way that they have more communication with the providers to then help them guide through the emergency department stay and then decrease significantly the long-term risk for mental health disorders.

DR. THEA GALLAGHER: This seems like a real integrated model that you're trying to bring into the ER with modifiable factors, not just bringing in embedded psychiatry or something, but really almost kind of working with emergency departments to create environments and, again, modify those factors that you find to be relevant in the persistence of PTSD or the development. 

DR. KATHARINA SCHULTEBRAUCKS: Absolutely. And for example, our team is very interdisciplinary, and I think that's necessary. It is like, first of all, on the team we have emergency department providers who really tell us, okay, where is really the problem? Where machine learning approaches would be really helpful? And I think that's like key. We need to work very close with the providers together. We need to work close with the patient also in what they have for concerns because there's a lot of concerns also from patients with regard to, is the machine learning model doing now the decision and they want to understand why they are receiving what kind of treatment or this prognosis. We need to work closely with experts in machine learning and computer science and biomedical engineering and psychologists. We all need to sit together because we all bring different perspectives together and then I hope that we better target what is really in the clinical decision-making process. Where are the little gaps and difficulties that we can then target?

DR. THEA GALLAGHER: It's interesting because from doing prolonged exposure therapy for PTSD, a lot of the work you end up doing is kind of disentangling some of the cognitions or beliefs about those very critical moments. And it sounds like if you can intervene at that level, it might be able to, again, help with more positive outcomes.

DR. KATHARINA SCHULTEBRAUCKS: Right. There's like one moment where people usually are in contact with the healthcare system in the emergency department. Then they are discharged and then for a long, long time, a lot of them don't seek any help. And it's like something different, it's like I'm also interested in for whom, for example, prolonged exposure therapy works or at what time we need maybe to start doing it. But I'm also interested in providing help on a more scalable level and we know that's only small subset who need help and different subsets need different help. But the problem is we are kind of living in a bubble. We know how to seek help. We know how to see when we are not feeling well and what could be additional steps. But there's a huge group who don't know that and when they are still in contact with the healthcare system, could better guide them, could better see what kind of preventive interventions are meaningful, then we could hopefully reduce the prevalence of PTSD or also other mental health disorders.

DR. THEA GALLAGHER: I think that seems even specifically relevant for people who develop PTSD, maybe who don't have a prior mental health history, and they might not, again, know how to access the tools or it's not part of their life to be connected with mental health interventions but this trauma in and of itself might create a mental health disorder. 

DR. KATHARINA SCHULTEBRAUCKS: Or even when they have prior mental health disorders, still not everyone has access to help.

DR. THEA GALLAGHER: And so using the emergency room and this experience as an opportunity to get people connected to the resources they need and it sounds like you're saying, we want to know exactly or specifically when, where, and what they might need that might be most effective for this individual person and that comes back to that precision medicine. 

DR. KATHARINA SCHULTEBRAUCKS: Yeah. Correct. 

DR. THEA GALLAGHER: It's clear that these machine learning tools can kind of address this gap with treatment and research, and it sounds like you're trying to do that at this rapid pace. What would you say like the main challenge is that exists now and that will continue to persist? 

DR. KATHARINA SCHULTEBRAUCKS: I think there are still a lot of challenges, and I think we are still at the beginning. But I think the most challenges that we are confronted every day is it's about not only to use machine learning models, and they are then the solution for everything, but I think we need to already start, as I said, we have multiple steps before they are useful and one is, again, we need to have interdisciplinary teams to sit together where really the problem is. We need to be better in seeing what is a good proxy of what I really want to measure. And also, when we think about biases, you all know the science publication of Ziad Obermeyer, who talked about racial biases which are a really big problem in machine learning model. But it's not about that our algorithms are biased, but it's about how we are designing the studies, what kind of data we are collecting, but also what kind of proxies we are defining from what we really want to measure. And I think that's still a big barrier. Also when we're interpreting the findings that are there, we need to—the same as in other factors—but we need to consider who we include for whom those algorithms work and also to not further increasing health disparities with those algorithms. There's still a lot of privacy concerns. What happens with my data? How can they be projected? And this is a really big important discussion. But also when we think about those algorithms and we would, for example, use them in electronic health record system and they are, for example, in our electronic health record system, there are already algorithms that are routinely used. And the question is always like how valid are those? And also how this information is understandable for the clinicians and the patient when we think about shared clinical decision-making and a lot of algorithms are like black boxes, very difficult to understand. So we need to also consider to open the black box, to also see if the decision the algorithm makes makes sense and also evaluate if what the algorithm does is really more beneficial than treatment as usual and is really cost-effective but also safe. So there's a long road that we need to still go, especially in psychiatry but I think what our design already shows is there are really a lot of chances to be better in identifying individuals at risk, but also we still need to test a lot of things. 

DR. THEA GALLAGHER: And are you also hoping that this would translate or generalize to maybe other traumatized populations. I'm thinking even on a study about healthcare workers but like secondary trauma or I'm thinking you know what's currently relevant is all the wildfires in California. In different settings, maybe if they aren't going to the ER, are you hoping that there's a way to kind of build this out? 

DR. KATHARINA SCHULTEBRAUCKS: I only talked about one pillar of the work that we are doing in emergency department, patients that are admitted to the emergency department. But we have also studies in, for example, military personnel where we also see already prior to deployment, can we identify risk factors of who are developing mental health problems after deployment? We have studies in emergency department providers because they are confronted with chronic stressors to better see if we can help them also to decrease the burden of burnout and also associated physical and somatic disorders that are associated with those stress pathologies. And so it it's not about only traumas. We have studies here where we enroll patients after severe suicidal ideation and suicide attempts to better identify individuals at risk for later on developing mental health issues also after they're discharged from the psychiatric unit here at NYU. So overall, we look at different mental health groups and to see us in different contexts how those digital technologies and computational approaches could help us and better identify those at risk.

DR. THEA GALLAGHER: Again, we kind of talked about surprises, but has there been a situation where the computational methods actually kind of challenge the theoretical understanding that we may have previously held? 

DR. KATHARINA SCHULTEBRAUCKS: So we just published a review where we examined that question because we were always confronted with that concern. This is a data-driven approach and you just throw data in an algorithm and this has nothing to do with theory or what is always like these two schools like theory versus machine learning. And I think what we also found is, first of all, that there's a huge overlap. We also identify new factors, but it's like all are in line with what we also have for ideological models of PTSD risk. But also, I think it's how we are using machine learning is for this very important. I would like not to see that as two different approaches. I'm a fan of using more theory-driven approaches in selecting what kind of predictors I'm incorporating in my machine learning model because I think that would have been like colleagues examined over decades in psychology of identifying relevant factors is important. But I think what machine learning can help us is using that knowledge and when you look at, for example, PTSD, there are so many different ideological models of PTSD risk. And now we can combine the knowledge of those different models. As you said, cognitive functioning. We know about biomarkers that are relevant. We know about early PTSD responses. And we know many different markers and areas that are relevant, and we have a chance here now to combine it and to have a more what is also aligned with the theory, that we have more psychosocial model of mental health, but we are now able to do this and not just look at single factors and how they are associated with it.

DR. THEA GALLAGHER: It sounds like trying to, instead of having them maybe be at war with each other, kind of integrate them together, and then also bring this data or this information to the clinician who, you can say, if they are working with somebody who's recently experienced a trauma or has a history of PTSD, or again, maybe is a recent experience of trauma maybe utilizing this research to inform their care as well? 

DR. KATHARINA SCHULTEBRAUCKS: Correct. And I think also because what machine learning model can add to what has been found so far is when you think about all the other factors that have been found in more traditional statistical way, looks at average. Like what is an average difference, for example, like someone with higher age is less at risk than someone with younger age or the other way around depending on the disorder or females compared to males. But what I think machine learning models can be very helpful is that there is a huge spectrum and also consider that spectrum and also have a more complex interaction of multiple factors across the spectrum, and then this is relevant for therapy to have then more individualized approaches. But I think that's why theory is important. It guides and it should guide computational approaches. But we can maybe have a more holistic picture but also more individualized picture using those approaches. 

DR. THEA GALLAGHER: We didn't talk about this too much, but you mentioned it, working on identifying objective measures of burnout. Why is that so important? And what are those kind of objective pieces that you've identified?

DR. KATHARINA SCHULTEBRAUCKS: So this is another NIH-funded study where over three years have at multiple time points connect with emergency department providers and see, first of all, how their burnout symptomatology develop but also, as you said, we try to have objective proxies of burnout. And why I became interested in that, so this study already runs a couple of years and was planned before COVID and then COVID, of course, accelerated the relevance of it. But the problem is that self-report is very biased also and there's a lot of social stigma, especially in the work context. And especially when we think about emergency department providers who have a high demanding job, but it's also very often also associated with their identity. And so I was very interested when I worked as an emergency department with my colleagues there when I saw that and I also talked with them to see, can we find better proxies of burnout that not necessarily require them to you know admit I have issues, but then also implement wellness programs. It's not about pathologizing. I think this is now happening more and more that there are more awareness opportunities also that are implemented across different hospital systems here in New York City and we also evaluate which ones are more relevant and also for whom. And that's how we initially became interested in it. And also because burnout is really associated also with a lot of cardiovascular risk factors. There was a study published years ago which had reported such a higher mortality in physicians than the general public. So when I heard all of this, I was like, okay we need to do something to better identify those at risk.

DR. THEA GALLAGHER: I think it sounds like with identifying some of these objective factors with burnout, you might be able to pull apart what are the internal factors, and then what are some of the systemic factors because we know both can play a part in burnout. 

DR. KATHARINA SCHULTEBRAUCKS: Yeah. So we look at both. So in that study, we look at subjective factors. We look at objective proxies of these subjective factors. But we also look at occupational factors. So we examine, for example, how many shifts that like how many patients died, how many patients were on ventilators, which is kind of like a proxy also on severity of the cases, on how crowded the ED was during their shifts. So we have these objective factors. We have this subjective response. And we hope we can better understand really what factors are contributing to more burnout risk. And then we also look at, as I said, it's over three years so every six months, we check in on our providers. And we also look at biological factors and how they change depending, for example, immune inflammatory markers, indicators of metabolic syndrome, etc., and markers of chronic stress. And we want to see also what kind of objective markers of the ED environment they worked in also impact those, but also their mental health and have more understanding about that. And also the neurocognitive performance because there's always the discussion that burnout impacts that which leads to more errors, which then also increases or has the chance to, so there’s like a risk for patients. But it's very difficult. There's no really longitudinal study to really evaluate that and so that's why we look at all these different factors over time to see if they really impact each other. And then what kind of wellness programs would be good to decrease the risk.

DR. THEA GALLAGHER: As we move toward precision psychiatry and really trying to find scalable assessment tools or scalable interventions, what are you excited about with what's happening there and what do you hope to see continue to happen there?

DR. KATHARINA SCHULTEBRAUCKS: I think what I was the most excited about is again, I'm sorry, I'm going back to the emergency department, but I was really excited that just how someone is talking about what happened to them early after trauma is super relevant in forecasting risk. And we used large language models and developed a tool that we hope we can implement at one point, but really it’s super relevant how they are talking about it. And what we also saw, and this was also very exciting for me, is some people could ask, they may not feel comfortable talking about what happened, but the good thing is they don't need to talk about it. Some of them really, they say just a few and talk about other things, and that's absolutely fine. But what we saw is that patients actually enjoy…for them, they feel very seen and validated having the opportunity to talk about what happened. And it's an open-ended question, so no one intervenes. And they are in either way ask those questions. So it is like a standard procedure. But having the opportunity to talk about it is really a relief for a lot of patients. And this was, I think, for me the most exciting that something that is already happening. Every emergency department does that, they ask the question, but we could use that in really identifying those at risk. And it's not an additional burden for providers. It's not an additional burden for patients. And this was, for me, a very exciting moment when we realized that. 

DR. THEA GALLAGHER: It sounds like it almost comes back to that human connection, social support, relationship-building that's so important and might actually be really important to our health, especially in these crucial moments.

DR. KATHARINA SCHULTEBRAUCKS: Exactly.

DR. THEA GALLAGHER: Well, wonderful. Thank you so much for being on the podcast and sharing all this information with us. 

DR. KATHARINA SCHULTEBRAUCKS: Thank you so much for having me. 

DR. THEA GALLAGHER: And thank you to all of our viewers and listeners at home. If you liked this podcast, please remember to rate and subscribe wherever you watch or listen to your podcasts. I'm Dr. Thea Gallagher and from all of us here at NYU Langone Health, thank you so much for watching and listening, our Insights on Psychiatry podcast today.