Good afternoon, I see we are at 345 and knowing that Philip and I are keeping you from your evening affairs, let's get started. Very pleased to welcome you to this symposium, a patient-centered research roadmap to inform clinical practice of bipolar disorder. I'm Mark Fry, I'm a psychiatrist at Mayo Clinic and will be leading this symposium with Phil. Our colleague and friend, Dr. Kate Burdick is unable to join us as she is flying back to the United States and so Phil and I will tag team this together on her behalf. So some of you may not be familiar with some of these concepts but the goal of the symposium today is to introduce you to learning health systems, learning health networks and how they really are, I would argue, the best mechanism to really conduct patient-centered, meaningful research, really optimizing large-scale data sets, networks and molecular medicine. You'll be hearing some theoretical introduction to that from Dr. Wong and then some real practical early projects and then I will actually close with thinking of some of that theoretical construct that he's introduced us to and how we will be moving that forward with a very exciting initiative, BD Squared, which we'll present this afternoon. Let me please introduce Dr. Philip Wong, who's the Director of the Learning Health Systems Center at Brigham and Women's Hospital and Professor of Psychiatry at Harvard Medical School. Thank you, Philip. Thank you and good afternoon. Can you hear me okay? Yes. And let me see if I can figure out how to advance. So Mark basically has kicked off what we're trying to do here and what we're trying to do is describe how you would build what's been called a learning health system, actually give you a practical view into how that can be done and it entails doing things like standing up measurement-based care, collecting real-world data as a matter of routine practice, integrating it, using it via things like machine learning and so it will get into some concepts maybe some of you have been exposed to but if not, we'll try to unpack this. So firstly, let me just say why is something like a learning health system even necessary? Part of that has to do with currently we can't really adequately treat the patients that we have and I had the pleasure of being an investigator in the National Comorbidity Survey replication from which these data are drawn and we ended up calling this the rule of halves whereby among the 60 million or so Americans who experience a mental disorder each year, including the over 10 million people who have serious mental illness, the most serious and impairing forms of mental illness, only about half receive any form of care whatsoever and then if you take those that receive any form of care and look and apply some evidence-based guidelines to see if they are receiving care that has any likelihood of helping them, well only about fewer than half are actually receiving care that one would think might be able to help them and then if you look at those who are receiving something that might be capable of helping them, fewer than half actually really fully benefit in terms of achieving recovery. Another reason we need to develop things like learning health systems and transform mental health has to do with the fact that there's slow detection and late intervention for mental disorders and again from the National Comorbidity Survey replication we observe that mental disorders are nearly unique among illnesses in how early in development one is put at risk then initially experiences symptoms, experience full-blown syndromes and ultimately disability and that has to do with the fact that the brain systems that underlie behavior, mood, cognition are often disrupted very early by things that are called social determinants of health. These are environmental exposures and experiences, especially things like trauma which disrupt neural circuits, neural symptoms such that the age of onset, the median age of onset is actually 14 and three quarters of individuals who are going to develop a mental disorder develop them before age 24. Despite that extremely early age of onset, there's unfortunately a 10-year delay between the emergence of a mental disorder and actually being diagnosed and having any form of care instituted and in that time disability sets in such that by the time individuals are actually being treated it becomes largely an effort in rehabilitation as opposed to cure. So let me just begin to unpack what I mean by a learning health system. Well in this case we're talking about what has been called a deep learning health system and the goal of it is to conduct research that can be rapidly forward translated into practice and it starts with creating what's called a discovery cohort. Now these are individuals who are deeply phenotyped, in other words they're very well characterized and if it's possible you want to collect biospecimens from them, things like genomics. There's a number of levels and dimensions of data that are very important for identifying novel disease mechanisms. You know in my entire decade at NIMH as the deputy director, my director and I, Tom Insel, we lamented that despite having spent $15 billion we hadn't discovered a single new disease mechanism, which is really unfortunate because those are also the keys to identifying what you need to target, your treatment targets, and even more importantly for our disorders, prevention targets, because if we don't get there early we obviously are really behind the eight ball. So around this discovery cohort in a learning health system you want to build testing cohorts. These are subsets of what I call, they're motivated clinicians. They're very interested. They're either interested in the specific subject matter or they're interested in research and they also are, I call them champions or astronauts. They hold sway. They're kind of thought leaders in their clinic and if you can recruit these testing cohorts they give you the ability to very rapidly test and implement innovations, discoveries, new interventions and rapidly essentially test them and see whether they indeed help improve patient care and outcomes. Finally in this ever expanding Venn diagram you want to build what are called implementation cohorts. Now these are all patients who might be at the site or the clinic and all of the clinicians who are treating them and the goal is to give them shared visualization tools, things we call dashboards that could allow clinicians, maybe a team that's treating a patient, to all see how they're doing. How are their symptoms and functioning changing longitudinally and so you need shared visualization tools and it's also a way to push out what are called clinical decision support rules or recommendations. So if a patient has a particular risk factor or if there's a predictive algorithm that says they might respond to this particular treatment, personalization, a way to alert the clinician and for those of you who work with EHRs you know they're like BPAs and there are ways to be alerted to the fact that a patient might respond if you prescribe the recommended treatment. So finally the implementation cohort is also very useful for what's called reversal of reverse translation. Now this is turning the paradigm around and saying how can we learn from practice to inform research and you know for those of you who are clinicians you know not every treatment helps every patient and you're literally putting most of your patients through N of 1 experiments, right, trial and error. You see if they respond to an initial treatment, if not you maybe follow guidance from star, step or Katie on how to augment or switch or and you're constantly experimenting and often you reach the end of these what we know from you know large you know practical trials and you're left on your own, right, and you're having to you know literally experiment and what if we could capture all of these N of 1 experiments, integrate all that data into a machine learnable database. Well maybe we could use things and I'll just mention some things, advanced causal inference methods and the goal here is to again learn from your practice, learn from the natural experimentation that you're undertaking and ultimately find treatments, preventive interventions that again lead to better patient care and outcomes. So at the core of a discovery cohort are the research domain criteria and how many of you have heard of RDoC? All right, so not many, maybe, all right, five, five people. Okay, so when I was at NIMH I had the privilege of serving on Tom Insel's initial work group to stand up RDoC. Now RDoC is not meant to replace DSM and often people when they hear about it they say oh this is supposed to replace DSM. What it's supposed to do is help us find what are the underlying dimensions and domains and dimensions, things that could represent disease mechanisms, targets that actually are causally related to developing a mental disorder, both the symptoms and the loss of function and as you can see the goal is to essentially reconceptualize mental disorders based on these underlying neuro, they're fundamentally neurocognitive deficits for those of you who think about cognition. These are deficits in neurocognition that essentially represent deficits in known neural circuits and known neural systems, systems like as you can see here negative valence, positive valence, cognitive systems, systems for social processing and arousal regulatory systems and sensory motor systems. What the overall goal of RDoC is is actually to then find the deficits in underlying levels. Again that could represent treatment and prevention targets. Levels like levels at genes, molecules, cells, circuits, physiology, behavior, even self-reports. There are certain symptoms that may be very good measures of known neural circuits. A good example is for example anhedonia. The other goal of RDoC is to identify interactions between deficits at these levels and the environment. We know environmental exposures, experiences, again determine maybe the majority of the variants in our patients' mental health outcomes. We also are very interested in identifying what those interactions are. We're also interested in identifying what are the interactions with critical periods in neurodevelopment. Again, our patients are put at risk very, very early in development, but we need to know what those critical periods are and what are the critical periods during which you can intervene to change trajectories. That's another goal. This can be accomplished by integrating all those levels of data into what's called a digital knowledge base. This is integrating measures of RDoC domains. You see on the left-hand side with measures, and I'm using shorthand there for I2B2. This is a high-dimensional data set that we work with at Mass General Brigham. Think of it as having, they're both biospecimens. There's also biosensor data, and I'll describe what I mean by that. There's also brain imaging data. These are additional dimensions of data that you want to integrate into the digital knowledge base. There's also, as you can see on the far right, what I'm calling ground truth. I'll unpack that in a little bit. These are basically automated measures of symptoms and functioning, very efficient and accurate. Another thing we also want to systematically collect in a standardized fashion is measures of social determinants of health. I'll describe one particular efficient way that we are exploring. What you can do when you have a digital knowledge base is then subject it to machine learning and artificial intelligence. The goal here is identifying what are called novel biotypes. These are biosignatures that predict a patient's course, the outcomes that they're likely to achieve, and also their response to individual treatments. The goal here is, again, trying to identify ways to get people on treatments that are likely to help them quickly and not have this be subject to experimentation during which patients are at risk for adverse outcomes. What about building testing and implementation cohorts? This requires additional developments. In this regard, we're very grateful to my former employer, NIMH, for giving us two grants to help expand the digital knowledge base with new types of data. I'll mention a few. Also develop the methodology for analysis of the digital knowledge base, and also develop a next generation of investigators who are capable of working with this kind of data. The data that I trained with is in when I went to Epi grad school. My co-investigator, Ben Cook, and I have been very fortunate to receive what's called an NIMH P50 Advanced Laboratory or Alacrity Center grant. I put up there a quote, what our aims are. Our aims are to establish a learning health community, and I'll get to what that is in a little bit, outside the clinic walls in diverse neighborhoods, within the clinic walls, establish a fully integrated learning healthcare system. Within the healthcare system, try to do the best you can. Develop biotypes that can help you predict how a patient's going to do so that you can apply interventions, personalize intervention, but even better, outside the clinic walls, can we learn to identify patients at risk? Can we intervene early? Could we build resilience and avert the development of mental disorders altogether? To expand the digital knowledge base, I've highlighted here, we have been working with collecting ground truth, and I'll say a little bit more about that, but basically, these are patient-reported outcomes collected in routine practice. This is not research data. This is data being measured by clinicians who use it actually to diagnose, to longitudinally follow their patients, and to make treatment adjustments. We hope to expand the available data by using natural language processing of the ubiquitous physician and nurse's notes. In Mass General Brigham McLean, there are millions, literally, millions of notes that every clinician records, and if you're an inpatient your nurses write a note every shift, physicians you know write notes with serious certain periodicity, but it's in free text. How do you standardize it? How do you extract what's called structured data from these? We also hope to merge the claims data with other extant data sets, neighborhood level social determinants of health, you you can get aggregate data on the levels of crime, pollution, you know there's you can learn an awful lot about what what individuals are being exposed to. We have a data use agreement in the center with the criminal justice system because as you can imagine many people end up being routed through criminal justice you know they for example a lot of youth end up in juvenile justice. That's their first sort of brush with treatment often and I say treatment lightly because it's obviously not not ideal, but if we can merge criminal justice system we might have a view into how to intervene earlier in the community right and and before people come for formal mental health care and it and into formal mental health attention. We also want to deploy advanced methodologies. These are advanced causal inference methods, I won't get into any detail here, but they allow you to use observational data, not randomized controlled clinical trial data, but what's called real-world data to generate real-world evidence and there's a huge push on the FDA's part, on a lot of people's part because you know what what the whole field has recognized, well the whole field of medicine is the data from these small non-representative samples that of individuals that end up in randomized controlled clinical trials. They don't represent the populations you're treating, they don't represent the conditions under which care is delivered nor do they generalize to you to you and your patients and so there's a lot of attention being given to use using a learning health systems real-world data to generate what's called real-world evidence. Finally, we are trying to develop a next generation to be able to do this kind of research through a NIMH Harvard School of Public Health T32 research fellowship training grant that my colleagues Miguel Hernan, who's a leading methodologist, who I actually went to grad school with, he's fantastic, and Matt Nock, who's a leading suicide researcher in the country and he's also just a really nice guy. He's a MacArthur genius, too, and then chair of psychology, but incredibly humble and a brilliant suicide researcher. Our goals are to educate a new generation of promising interdisciplinary pre- and postdoctoral trainees in the use of cutting-edge methodologies for comparative effectiveness research. In this case it's specifically around reducing suicides, but obviously the methodology will generalize to all of mental health care. Okay, so I keep using this term ground truth and it's critical for building testing and implementation cohorts and it comes from Tom Insel, my former boss, and he uses it, and maybe you've heard him talk about it, to lament what is missing in mental health care. Right now as few as one in five mental health clinicians actually measure, in a standardized fashion, things like patient symptoms and functioning, and largely they're in integrated care. So if any of you work in integrated care, you probably, you know, your patients probably receive a PHQ-9 or a GAD-7. How many of you actually are in a system where, you know, patient reported outcomes are measured, you know, longitudinally in a systematic fashion? One, two, all right, so fewer than one in five. Right, so that is a major problem, right, if you're trying to develop a learning health system, if you're not collecting, you know, measures of symptom severity, functioning in a systematic fashion longitudinally, right, because then your digital knowledge base is devoid of that data. So Tom and I were very pleased to fund two R01s submitted by Robert Gibbons. He's a leading mental health biostatistician, you see him on the right. I have no conflicts of interest. I don't get paid by them or anything, so I'm simply representing what is useful. We're interested in using them as research tools. They are computerized adaptive tests based on a statistical technique called multi-dimensional item response theory, and I'll just have one slide to show you what that is. They can very validly and accurately and efficiently measure either the screen to formally diagnose or to monitor over time the severity of mental health, substance use disorder, suicide risk severity, as well as social determinants of health. On the left-hand side, you see what CATMH has modules for. CATMH is the adult version. There are modules for depression, anxiety, mania, hypomania, suicidality, PTSD, substance use disorder, psychosis, adult ADHD. We've also been working, we have had a research partnership with the developers of this to develop what we think is the first module that measures not just the presence of individual social determinants of health, but in overall severity, which is important for practical reasons, which I'll maybe answer if anyone has questions about it. On the right-hand side, you see what KCAT or KittyCat for youth, for children and adolescents, can measure. There are modules for depression, anxiety, mania, hypomania, ADHD, conduct disorder, ODD, suicidality, substance use disorders. The thing about KCAT is there's both a parental informant component and a youth component. For example, for children, you would want the parent filling it out. In adolescents, you might want both, and so there are components for the informant as well as the youth, the adolescent. Statistically, you can actually combine the two. You can use either one, but you can combine the two, and actually the findings end up being more robust. So using ground truth by collecting CAT-MH, KCAT, CAT-SDOH, and HUDAS. HUDAS is the World Health Organization Disability Assessment Scale. It's just a brief measure of functioning, and so we also think that that should be collected as a matter of routine practice, if for no other reason, because functioning is often what your patients care most about. When I was practicing, you'd have a patient who has psychosis who is okay if their symptoms aren't fully resolved, as long as they want to work, they want a job, and they have romantic interests, and that sort of thing. That's what they care more about often. The ground truth can be collected for what's called measurement-based care in a number of populations and settings. So we've experimented with collecting it out in the community. These are not patients. These are people out in the community. You can do outreach in an automated fashion. You don't need a bank of people calling. This is using software that's commercially available. For any of you who, like I had my annual well visit, and I get reminders via text, email, and robocalls that tell me I have an appointment, and I have to push one to say, yes, I am going to show up. So there are ways that you can engage people, whether to fill something out or be aware, but you can do automated outreach. You can also use, you could use, these ground truth measures to screen, and there are a number of ways that we have been able to do so. That can then allow you to rapidly recommend for people things like low-cost wellness and prevention programs. This is that resilience building that I was describing to you. Often, if you can catch people early, maybe you can avert either a full-blown episode, if the person already has a disorder, or maybe you can literally change their trajectory so that they never develop one. You can also use it in primary care. Again, like PHQ-9s and GAD-7s are already being used, both to screen and then used to follow people over time, and that can be used to adjust treatments. You can also apply it in mental health specialty care at the central intake level. This is when someone is being referred to mental health specialty care, to the psychiatry department, and if you applied these measures, it could allow you to do what's called rapid data-driven triage. So depending on your CADMH, module scores, maybe you are only moderately symptomatic, so you could remain in an integrated care program. Treated in primary care, have a psychiatric consultant. In this case, we call it comprehensive collaborative care, because collaborative care is, for those of you who are familiar with it, it's integrated care, but it's largely for depression. Depression alone, using a PHQ-9. If you expand your measures, you might be able to provide disease management comprehensively across a broader range of mental and substance use disorders. You can also use it within inpatient and outpatient specialty care. We had experience standing it up in our inpatient units, and again, it's useful for monitoring how a patient is doing, understanding when their symptoms are resolving, and also making treatment adjustments. How am I doing on time? Okay, so let me let me speed up here. You can also use it to rapidly get people to specialized services. So if someone's very suicidal, get them to the emergency room, acute care, or specific programs like first episode psychosis programs, right, if you detect that someone is having their first episode. So let me just give you, close by giving you some examples of how you can deploy these discovery, testing, and implementation cohorts to conduct rapid forward and reverse translation. And just this schematic shows you, you use the digital knowledge base. Ideally, you apply, you know, things like artificial intelligence, and you can discover computer-aided diagnostic and risk assessment tools. You can deploy those in things like clinical decision support rules and intervention selection tools. Here's one example that we've been exploring, and it's essentially to develop and test what are called behavioral biomarkers. You can collect sensor data, right, so these are like your Fitbit or your smartphone, which, you know, can measure your steps, and your activity, and your sleep, and your heart rate, and your heart rate variability, and things like that. You can train the sensor data on, for example, CADMH. You could see when someone is perhaps not sleeping and at risk of, you know, developing a manic episode, right, and you can institute rapid intervention based on that passive monitoring. So, you know, by the bottom you see some examples of that. Maybe you could use that to recommend, you know, ICBT if the patient is having an insomnia exercise. If they're inactive, you could taper drugs that are causing people to be unsteady on their feet. You can also use it, in this case, to validate RDoC, and I won't get into this, but essentially, you could use the digital knowledge base and study treatments as actual experimental perturbations, and in that way, you could actually validate that the deficit in an RDoC domain or dimension is actually causally related to developing symptoms and loss of function. Oops, oh, well, the last, the very last one, I also wanted to give you an example of, if I can go back, well, give an example of reverse translation. That's where, again, you use the data that you collect from routine practice in your implementation cohort. In this case, one novel idea is to use it to identify treatments for other conditions that might be repurposed, and a way you could do that is look at, you know, there's a whole host of pharmacopeia, you know, that's out there that, and some people have very interesting hypotheses, like some of the NSAIDs, you know, these NSAIDs are maybe uniquely beneficial for certain types of depression, which may have an inflammatory basis. Well, you could see that if you had, you know, a digital knowledge base. You could see if someone who started on Tylenol versus Motrin versus, you know, another, do they never get hospitalized again? You know, I mean, it's a very interesting potential reverse translation use, but let me stop there, and I'll turn it back to Mark, who will show you how this could all be done in this wonderful, and Jen Buzuki's here, I just thank you again for your leadership in this terrific BD Square initiative, which will put meat on the bones of what I just talked about. All right, that was a fantastic introduction and visual tour, and I can tell that there's kind of, there's a terminology and a vocabulary that needs to mature as we go forward, so we thank you for your tutorial, and looking forward to learning more further from you and with you. So I wanted to really expand on that and really give you a clinical perspective of what this is, and more importantly, what it could be, and I am at Mayo Clinic, and I wanted to recognize that I've had the honor and pleasure for a good number of months to be on the Scientific Steering Committee and co-chairing the Integrated Network with Kate Burdick, and you'll be hearing more about our breakthrough discoveries for thriving with bipolar disorder or BD Squared. So my disclosures are here for your review. I will emphasize that all the drug treatments that I will be referencing are off-label for bipolar depression. That is actually part of the problem as to why we're having this symposium. I do have funding in a number of, in a number of genetic studies. I will be presenting to you today specific pharmacogenomic data that is not part of those funded investigations. So let's take what Phil introduced and maybe shape that a little bit as to how we might think about that in the context of bipolar disorder. So a ongoing iterative definition of learning health systems, which I generally think of as a single site like Mayo Clinic, and then more broadly a learning health network that takes multiple sites in an integrated way to then continue learning health. I want to talk to you about the clinical perspective of who needs to really be involved if this is going to be successful. So the concept of learning health network or LHN stakeholders and concepts that really need to be everyday, ever-present, and Phil talked about that already, which I think is great. I will underscore that. I think it's always helpful to think about some very simple examples as to how this type of innovative technology and active learning can go forward. And then I'd like to show you a couple different areas that a number of our colleagues at Mayo were talking about as to gaps in clinical practice and might those be low-hanging fruit to start pilot projects within a learning health network. And one of them is bipolar depression and I'll tell you a little bit more about the rationale and then we'll close introducing BD squared, a transformative initiative going forward to identify breakthrough discoveries for individuals to thrive with bipolar disorder. All right, so if we go to a government agency for a definition, the Agency for Healthcare Research and Quality, AHRQ, actually has a pretty good one. A learning health system is a system where internal data and experience are systematically integrated with external evidence and that new knowledge is put into practice. That's a very active intent. Generating new knowledge, one, returning it or reinvesting it back into the practice, two. A recognition that these types of health systems provide patients higher quality care, safer care, more efficient care. So who needs to be part of an LHN? Who are the stakeholders that need to be there from time zero? I would argue there are top line leaders that are needed and front line leaders that are needed. Top line leaders are leaders within healthcare organizations that can set the tone of a culture, can advocate that this is important, can direct resources to really see that this moves in a systematic way. Front line leaders are those that are in the trenches, really busy practices, taking care of patients. That would be me. I actually have ping ponged between top line and front line but I think the key here is you need individuals who understand the practice dilemmas that you're trying to change. That's a busy clinician. That's an individual with shared experience who does not see that they're getting better, can't get back to where they were. Engaging individuals who are in the center of this problem are those who can hopefully solve this problem. So people are part of this but I think the other piece that's critically important and this is where at least the healthcare industry and certainly physicians weren't necessarily originally trained is how do we harness this technology to really move the network forward? This is really introducing clinicians and healthcare policy advisors to information technologists and I think actually large healthcare systems now are embedding these individuals, extremely talented into departments and to divisions. The last bullet in my opinion is the holy grail and it is the hardest one to sort of think about and design but if this learning is going to impact care, there has to be a dynamic learning process that's time sensitive, ideally real time that we then put back into the practice. Phil highlighted 10 years before new research is actually integrated into a practice. I think depending on the type of research it might even be closer to 15 to 20. So a couple of visuals here. I'm not gonna spend a lot of time on this other than to hit a couple high points. Let's go from left to right. Assemble patient data. That can be multiple sources and it's already there in many of our programs. So that's the EHR, pharmacy claims, identifying and having a regular investment in community input ideally from individuals with lived experience. The highlights in that first column are what I think are important add-ons that we're doing in BD squared that aren't necessarily within a conventional electronic health record or for that matter a quality improvement project or a learning health system. Bring in digital monitors, potentially digital therapeutics. Bring in cognition. Think about as it's related to drug therapies, pharmacokinetic or dynamic variation more broadly. Think of genome-wide association studies to guide treatment. Think of multi-omic ways to study and stage illness. Look at brain imaging as a phenotype or as a treatment outcome. Second big circle, with that technology, we start to really have different types of data come together that we never really would have had before. And this is where I think there's an important concept that we identify new vectors of association. With all this aggregate data, we may find that there is a volumetric imaging measure that might correlate to a PHQ-9 score and a serotonin transporter variation that then relates to quality of life. The combinations are endless. The point is that with that aggregation of data, we look for new vectors of association. That would be what I would call a discovery association. We then go back to the clinicians, lived experience individuals. What are the outcome measures that would be important that we can try to standardize going forward? How does that change practice? How does that new knowledge reinvest back in practice? So let me give you a couple examples. And I would say these are really simple, but you should know that hospital administrators think these are really profound, correctly so, because they change practice. So NYU Langone has developed a learning health system, again, for them. And creatively so, they do it in the context of qualitative, randomized projects. They emphasize, again, front liners need to be making these decisions based on where they see practice gaps. And they actually encourage very judicious selection of projects, which I think is very timely and parsimonious. That second bullet makes sense. You don't want an outcome measure that is rare. You want an outcome measure that has high volume events. Ideally, the shortest time possible just means that the quality project can have a start point and a finish point. And to Phil's earlier comment, no new data. This is not about introducing a new research metric to see if it helps improve the practice. That opportunity has passed. This is about analyzing data that clinicians, individuals will be using anyway. When I was chair of the Department of Psychiatry at Mayo Clinic, it was brought to my attention that as a department, we had 22 rating scales for depression. And I thought that was about 20 too many. So we brought all the clinicians together and said, this is stopping. We want to aggregate. We need more opportunities for data harmonization. Please come together and recommend two of these 22 scales. You can imagine how popular I was for at least a week. So, second bullet. Some of the examples of their quality randomization. Here's an idea. For this month, we're gonna do this. And the next month, we're gonna do this. Or this part of our hospital does A and this part of our hospital does treatment as usual. Examples that you see were of benefit. That second one. They recognized that changing the text, these alerts that Phil was talking about that were specifically geared towards clinicians who were encouraging or trying to recommend smoking cessation therapy, changing the text a little bit magnified the intervention and in fact was associated with more prescriptions of smoking cessation pharmacotherapies. They saw a problem. The text of the alert needed to be changed, changing that enhanced medication prescription. The third one. This is so profoundly relevant in any hospital system. They looked at those follow-up telephone calls that would be made by all sorts of different types of people with the idea of can that human contact encourage medication compliance, see how someone's doing in the hopes that we reduce the 30-day hospital readmission rate or improve patient experience ratings. The details of these QI projects I'm not familiar with but what this review referenced is they actually found that those initiatives were of no benefit in any way, shape or form. This is a good example. If your QI shows something is not helpful, stop. So this was a really great example of how those phone calls to try to reduce 30-day readmission stopped and that employee group was redeployed to do something else more meaningful. The conclusion of the NYU experience was that these systems can actually pay for themselves by reducing error and adoption of preventative services that then subsequently reduce more expensive hospital costs going forward. Another example is the Geisinger Precision Health Initiative. This is completely on the opposite scale but it really was based on an Institute of Medicine report from nearly 10 years ago highlighting that healthcare would be improved by genomic-informed precision health. What does that mean? So this is really identifying where there are known genetic variants that people would see as clinically actionable. You have the variant for breast cancer. You have the variant for not being able to metabolize coding. Those sorts of genetic variants, if we can get to providers and individuals, we can enhance care, provide more efficient care and have more cost-effective care. Let me show you an example of that. The figure is from a paper from our group about eight years ago when we saw a benefit of recognizing many patients were getting pharmacogenomic testing to look at pharmacokinetic or dynamic variants and how that might impact selecting an antidepressant and antipsychotic or stimulant. At the same time, we ventured to do a large project called The Right 10K, which was getting the right medicine to the right patient at the right time. And we sequenced some 65 PK and PD, pharmacokinetic pharmacodynamic genes in 10,000 individuals who were in primary care and had a visit with them at least once within the last year. So if that genetic information is in the electronic health record and I'm writing a prescription for a drug that might be variably tolerator or would be important for me to know that that variant exists, I get a pop-up. It says, do you know that your patient is a poor metabolizer at 2D6? You might want to consider something other than fluoxetine, which is a drug that is predominantly metabolized at 2D6. So these pop-ups are examples of this learning health. I would argue this has been, at least for this specific project at Mayo, of some benefit. But the other thing we have to remember is sort of constantly reassessing how valuable that is. And we've heard from clinicians that the number of alerts that pop up just gets to be so exhausting that you develop fatigue and you kind of stop paying attention to those. Or those alerts don't line up very clearly with where the actual data is in the laboratory section of the electronic health record. So a great example of early innovation, but it constantly needs reassessing. And you can see that the platform or the visual interface with the electronic health record can improve tremendously. So what might be some examples of stakeholders, front liners in bipolar clinics who know where the gaps in practice are? If we were to identify a couple sort of pilot projects, what might they be? This is not an exhaustive list. This is just a couple different ideas off the cuff. I'd like to present to you today why a focused effort on the depressive phase of illness, its treatments might be quite useful in a learning health network. I'd also like to present how an awareness of metabolic health and optimizing metabolic health could further enhance more efficient and more individualized care for individuals with bipolar disorder. If we have time, we can go over a couple more. So let's talk about depression, FDA approved treatments and how I would argue this is non-precision medicine. World Health Organization has ranked depression as the second most disabling disease worldwide. A D-A-L-Y is a disability adjusted life year. The estimates for depression worldwide are 120 million people with nearly 850,000 completed suicides annually. While women are more likely to present with major depressive disorder, if we start thinking about the benefits of data and electronic health records and large data sets, it's very clear real men get real depression. I don't know if Phil remembers, but that was an NIH campaign about 25 years ago. And we started to understand that men often didn't endorse depressed mood and anhedonia when depressed, but they very clearly endorsed a pain threshold change, more irritability, more drinking, more emotional lability, and that got better with an antidepressant. So circa 2023, there are about 25 FDA approved treatments in the United States. Antidepressants as a class are the second most commonly prescribed drugs in this country. An estimate, at least from one of our studies, that the peak prevalence rate is about 26% in women age 50 to 64. Here's an eye opener for you. The absolute majority of prescriptions are not written by people probably attending this meeting. These prescriptions are written by primary care providers. They are the initial gateway. They are magnificent clinicians. They have a critical role in the overall health of our patients. Probably don't have the same understanding of pharmacology and management of treatment-resistant depression as we do, so any tools and alerts or individualized care guidelines through learning health would be a tremendous advantage for them. But we get to this point of really wanting to have precision medicine, which one definition is, we go beyond a placebo-controlled trial. Best example I can tell you is very often when I start thinking in cases where pharmacogenomic testing may have some merit, I say something to the fact of, I can write you a prescription for an antidepressant that I know is in the drugstore and that you haven't been on, and maybe there's some symptoms here that would suggest you might do well with an SSRI, anxiety, for example. There may be merit in actually having a little more precision to the choice of medication based on your biology, meaning how you might metabolize medicine or what your brain receptors might do with these medicines. And that can, for some, be a value-add. I'll be the first to say that that type of decision-support tool, the large-scale studies, are negative. It is not ready for prime time. I would argue we need to do those studies differently, but we can talk about that later. But the general sense is, let's use more than an FDA approval in what we know is available in the drugstore to guide selection of a pharmacotherapy for you. So depression, lots of FDA-approved treatments, no precision. Bipolar depression, very limited FDA-approved treatments, even further lack of precision, and as that is highlighted here, the exception of one medicine, the combination of elansipine and fluoxetine, all antidepressants are off-label. This is a policy pitch that can often get me irritated, and I will promise I'm not going to spend a lot of time on this, but I would posit to you that clinical trials for bipolar disorder are fundamentally harder and more complicated than trials for schizophrenia and trials for major depressive disorder. Either poll, if you are trying to treat mania, there is a post-manic depression that is a potential adverse event. If you're treating depression, there can be an induction of mania that is not the adverse consequence that has to be thought of and designed in studies and other major mental illnesses, and that complexity, frankly, is why we have so few treatments. That is why antidepressants have been looked at so desperately, and the data you see here really is pretty unimpressive. When we look at meta-analyses, antidepressant versus a placebo, no significant difference in response or remission. There is some data to suggest that maybe that antidepressant, when with an atypical antipsychotic, might be better, but I suspect those benefits are modest. Last bullet, if we look at really current data as to the extent of antidepressant use in this country, this is the most up-to-date national ambulatory medical care survey, which showed from 2013 to 2016 that over half of all the visits for bipolar patients involved in antidepressants. And yet we have very little data to suggest that they are helpful. Two exceptions. I think there is enough evidence, and I'm very comfortable thinking about antidepressants for patients with bipolar type 2 disorder. This data is really showing venlafaxine, sertraline, some other data with fluoxetine look very effective. They are effective and they are safe, meaning no induction of hypomania, no induction of rapid cycling. And this is the second sort of clinical variable that I think it's important to spend some time with. One of our studies from many years ago really looked at what happens to antidepressants in bipolars who actually respond to the drug. On the left part of the slide, I'm showing you a survival curve that as the red or green lines go down, bipolar patients are relapsing back into depression. This is really a group of bipolar patients depressed on an anti-manic mood stabilizer, got on an antidepressant, didn't get manic, didn't have side effects, achieved remission, and stayed on the medicine for at least a six-month period of time. You can see that the group that had the antidepressant removed within six months' period of time is in red, and they are relapsing far faster into depression than those who are the green line, which was staying on the antidepressant. The slide on the right is really breaking that down a little further, and you can see that the longer you stay on the antidepressant, the longer you stay well. This really is anathema to how many of us were taught how to use antidepressants in bipolar disorder, which was basically very sparingly, once they're better, take them off. I would argue there is a minority of participants, but a group where an antidepressant probably for them is a mood stabilizer. Our field was getting stuck to saying these drugs should be used, they should be banned, and I think the better question is for whom should they be used and how to work with those medicines to really achieve the best outcome for them. I'm showing you a publication where we looked at r-modafinil. This was a drug FDA approved for daytime somnolence and fatigue, and it has some very good evidence base in bipolar depression. When the company got the r-enantiomer of modafinil called NuVigil, you see large study. Number one was positive, large study number two was positive, large study, sorry, large study number one positive, two negative, three negative. The second you hit two negative studies, most industry sponsors say we're done. I was frustrated because this was really the first compound that was not an anticonvulsant or an antipsychotic that was being studied for the depressive phase of the illness. The designs from a regulatory perspective for these types of compounds, which we didn't know a lot about, was a bar far higher than any FDA approved treatment for the depressive phase of illness. So we went ahead and kind of put all those studies together. Nicholas Nunez published this a couple years ago where it really does suggest these types of stimulant compounds are effective in treating bipolar depression and do not have higher rates of inducing mania. So this gets to Phil's earlier comment about old drugs or failed drugs, can we repurpose them or study them with greater precision as to who might be considered to be a potential candidate for a treatment like that. So antidepressants, do they work? The large body of evidence would say no. I would argue it would be of merit to study for whom are they helpful. The second question is are they safe? And we like to kind of think of actuarial risk factors. Let me tell you a story that happened many years ago but it hasn't left me for reasons you probably will understand. I have four siblings. My father was called into the office of the auto insurance company that was providing the family insurance and they said, Dr. Frye, we're no longer supporting or covering your family. He said, why? And the officer said, well, a blazer blew up, the Jeep went in the ditch, a Saab split in half, then we went to the cheaper cars, the Plymouth Horizon dented all over the place, the Ford Pinto, who knows? And by the way, lots of tickets that seemed to come out of everywhere. That insurance actuary list was saying, four boys, you know, all under the age of 25, these tickets, that accident, we're not supporting you any further. There's still some hard feelings about that. But the concept I think is really useful here. If there is a depression and I've tried conventional treatments and we're thinking about antidepressants, I literally go through that list and look at potential risk factors, gender, age, bipolar one, number of prior episodes of depression, anxiety, alcohol, comorbidities, subclinical hyperthyroidism, tricyclic antidepressants. This is all the clinical data that is amassed over 20 years. And we can use that in addition to some of the newer data that we're going to talk about and do something like this, a risk stratification. If we're going down this road, what's the risk of an adverse event? And let's make sure everybody is aware of that. Generally speaking, if I get three or four of those, I'm just kind of moving on. But I'd rather have a large-scale learning health system to tell me a little more specifically what to do. So let me quickly run through some new omics, biomarkers that are not in your clinical practice. They are not ready for prime time. But if we're building this resource, I do not want these biological metrics to be bypassed. And we really focus on electronic health and conventional ratings. We have to do more. The good news is we can. You see seven or eight studies that looked at the relationship between the serotonin transporter, which is the gene that encodes the protein where a molecule of serotonin or a molecule of fluoxetine will sit. The hypothesis was if you have less serotonin transporter, in this case the S allele, are you at higher risk of having an antidepressant-induced mania? That answer is yes. That is nominally significant. And I want you to know that I don't find this to be the most important piece of this slide. What I'd rather you take under your belt for further consideration is that upper right corner, which says there may actually be a serotonergic haplotype, which means several variants that protects against antidepressant-induced mania. But if we're really going to expand our work, we need to be thinking about the norepinephrine transporter, dopamine transporter. More broadly, we need to stop candidate genes and really look at the whole genome. And if we see a signal, then actually study that in greater depth. If we were to think about a creative way to do a study that takes pharmacokinetic and pharmacodynamic information together, this would be the one. This was a postdoc of ours from a couple years ago who took two very large data sets. PGRN is the Pharmacogenomics Research Network, where there were 800 individuals at Mayo treated with citalopram or escitalopram. STAR-D was one of the studies NIMH funded looking at a stepwise way to treat depression. These were prospective trials. And for those individuals who failed the SSRI, there was an opportunity to offer them an SNRI, in this case, either duloxetine or venlafaxine. The main point here that I'm showing this bullet that you see in the figure to the right, this was not about looking at how a drug is metabolized, just the pharmacokinetics or the PK. This was not about what happens in the brain or pharmacodynamically or PD. Actually looking at them is an interaction. We have to stop doing these silos of just looking at one genetic variant and the other. I'd like to see what happens when they interact with each other. And you can see individuals who failed an SSRI, if they were an ultra-rapid metabolizer at 2D6, and they had the long form of the serotonin transporter and the long form of the equivalent norepinephrine transporter, remember, duloxetine and venlafaxine have a noradrenergic piece, those individuals had much higher rates of achieving remission. So if we had that data from the very beginning, we would bypass the SSRI. We would not even start that medicine and we would go to the second. That's the sort of thing a learning health system can do. Phil had mentioned taking other diseases or other data sets and how could that inform a bipolar learning health network. This is data from our group that, again, is focusing on antidepressant-induced mania, but it's taking advantage of a genetic technology where we can quantify polygenic risk scores. And as you can see with that second black bullet, this is really what is a computation of genetic risk from large GWAS studies that relate to either a treatment response or a diagnostic category, if you will. So from our biobank, we used as the comparison to generate that PGS a very large study published in the literature where over 5,000 individuals with major depression were given an antidepressant. We did not find a genome-wide significant hit, so we're not focusing on a single variant, but when we look at the polygenic score, the overall genetic configuration of those individuals who responded to antidepressants in major depression, they look like the bipolars who are getting manic from those same antidepressants. So this starts to look like maybe there's a spectrum of mechanism that works well for major depression responding nicely to the antidepressant, but is on steroids, and that's what we see with the induction of mania when given to individuals with bipolar disorder. Comparison samples, thinking mechanism trans-diagnostically, learning health systems can do that. The last novel marker we want to see move forward is the concept of suboptimal mitochondrial function. This has been a term operationalized by my mitochondrial medicine colleague at Mayo, Thomas Kozich, really thinking of points in the spectrum of how mitochondrial functioning or vulnerability in mitochondrial functioning could be associated or at least studied with various points of illness along the mental illness spectrum, and what caught our attention is there a mitochondrial vulnerability that contributes to relapse in established disease. So let me walk you through figure A. This is a visual of an electron transport chain in mitochondria, and it is increasingly recognized that the very pathology of bipolar disorder may in fact be driven by a primary energetic dysregulation where quantifying this in mitochondria has merit. Going from left to right, that orange little bar is complex I. We know from work of Anna Andreazza that prefrontal cortex mitochondrial enzyme or mitochondrial genes for complex I are significantly reduced in comparison to control, so the very beginning of the electron transport chain, we see differences in bipolar individuals versus controls. Sue Tai, who had been at Mayo Clinic, is now back in Queensland, has an animal model of mania that when related with an antidepressant, when you compare that group to those animal models where mania wasn't present, she identifies complex IV, which is this brown one right here, and Hilary Bloomberg is very focused on the last stage of electron transport chain where energy is being produced, specifically looking at complex V. When Sue identified this mitochondrial connection to her mania model, we were lucky enough to have a postdoc who was reviewing all of the preclinical literature and how psychotropic drugs impacted mitochondrial functions. You can see on figure B various types of psychotropic drugs. The columns are literally going left to right, exactly what you see in figure A, those key points along electron transport chain. The green squares just mean that drug increased mitochondrial function at that point in time. The red decreased, yellow data wasn't very clear. Two points I want to bring up today. If you look at omipramine, across the board, with one exception of complex I, there are green squares. This drug is the gold standard of an antidepressant induction into mania. The tricyclic antidepressants have the highest rate for doing this. These studies are done 25-30 years ago. Here's a drug that induces mania. If we look at the preclinical profile, it is overwhelmingly green. Second comparison, look at paroxetine, SSRI, preclinical function, green, increasing mitochondrial energetics. Here is a sister drug, escitalopram. They're in the same class. We think of them very similarly, escitalopram does the exact opposite. Thinking about drugs as to what they do with serotonin, norepinephrine, dopamine, those days need to go away. We need to think of these compounds in different ways. We would argue that mitochondrial energetics or mitochondrial function has merit because when we take that preclinical literature and look at the left part of this slide, that's our predoctoral group taking all that information and just categorizing antidepressants that increase mitochondrial energetics in green, decreasing mitochondrial energetics in blue. They don't know that Haldol is not an antidepressant, so we have some collaborative correction work to do, but I'll have you focus on escitalopram. Part of the right part of the slide is what really just kind of blew us away. Manuel Gardea looked at our very careful treatment emergent mania phenotype from our bipolar biobank, and we pulled just those antidepressants that are in green or in blue, and we asked if the rate of antidepressant induced mania or treatment emergent mania was different, and it was substantially. If you had exposure to a mitochondrial increasing energetic antidepressant, your rate of antidepressant induced mania was almost 25%. If you were on an antidepressant that was shown to decrease mitochondrial energetics, that rate was significantly less. We controlled for as many clinical factors as we could, and we saw this robust finding. We would like to keep thinking about pharmacogenomics, mitochondrial energetics, biological variables that need to be part of this network going forward. How about clinical pieces that are already in the electronic health record but we need to prioritize, and they get on that screen or the platform visually prominently, and it's something we track carefully. From our bipolar biobank, we tried to think what are clinical measures that don't take a lot of time and could easily be tracked by a psychiatrist who didn't, you know, who's 25 years from internship. Here's what we found. I'm showing you our bipolar cohort where our bipolar individuals in blue are at significantly greater risk of having an elevated BMI. Our bipolar females in particular are at significantly elevated risk, greater risk of having a greater central adiposity, elevated systolic blood pressure, elevated diastolic blood pressure. We have these metabolic markers that are right in front of us. Why is that significant? Manuel Gardea, again, was very interested in looking at the quality of diet and how that might contribute to these cardiometabolic markers and actually depression. So it's a little bit of a complicated story, but let's go from the upper left and really focus on the blue arrow. As your diet quality goes down, your BMI goes up. As your diet quality goes down, central adiposity goes up. As your diet quality goes down, current symptom severity of depression goes up. As your diet quality goes down, disordered eating goes up. And I suspect that what that's really tracking is kind of nocturnal sort of binge type of symptoms. I can't think of more powerful evidence to show what you're eating right now is contributing to depressive symptom burden and disordered eating. So we've got to figure out a way, how do we get rid of clearly a diet quality that is very poor and is contributing to metabolic burden? There is a lot of interest in thinking about precision nutrition, in particular, therapeutic ketosis. And I would encourage you to come to our ISBD meeting in Chicago to see some of that early data. Why might that be relevant? And to the comment about maybe we just look at aspirin, Tylenol, see who gets hospitalized, why don't we actually look to see how many of our bipolars are on metformin, loraglutide, GLP-1 agonist insulin, and let's see what happens to symptoms of depression, BMI, overall health. A fascinating study by Kalkin and colleagues that was looking at targeting insulin resistance in bipolar disorder to see if reducing that insulin resistance burden was associated with mood improvement. And indeed it was. Methods here are important to really appreciate with regards to the rigor that they have. So this was bipolar patients who were insulin-resistant positive, and you got to that by a blood examination of your fasting glucose and your serum insulin generating a HOMA-IR score. If you had established diabetes, you were not part of this study. We're not looking at end stage, we're looking at early signals of insulin resistance. The group was randomized to metformin versus placebo. And in those individuals who were IR converters, by that I mean being randomized to metformin or placebo. I think one of the placebo patients actually converted. But the intervention took you from insulin-resistant positive to no longer being insulin resistant, or what we would call IR negative. That was associated with a significant drop in symptoms of depression, as measured by the MODRIS, which you can see over a 26-period of time. So this is not an antidepressant. This is actually taking an intervention that our diabetes colleagues use all the time and saying if we convert and turn the insulin-resistant patient into someone who is no longer insulin resistant, we're seeing improvement in depressive symptoms. So last couple of minutes, let's talk about a transformative initiative going forward, and that is a roadmap for bipolar disorder research and improved care. This is BD2, Breakthrough Discoveries for Thriving with Bipolar Disorder. Through the Strategic Institute for Milk and Philanthropy, there are three families that have committed to really changing the landscape nationally and internationally in how we work in this clinical space of bipolar disorder. The Buzuki family, again recognizing Jan Allison Buzuki, the Daughton family, and the Brin family. What I love about this program is it is highly integrated, and each of these four colors is contributing meaningfully to the bottom left, which is the integrated network. There is an initiative going forward, large-scale genetics work, large-scale brainomics work, discovery grants really looking at mechanisms of bipolar illness staging or progression, and I would have you think about all of that as informing an integrated network with the goal of improving outcomes and wellness by optimizing care and discovery of new biological and clinical targets. A wonderful, renowned leadership team, again, I've had the honor of working with Dr. Burdick and the steering committee to really shape this early initiative. The integrated network, that's where our learning health network is going to be, but it's more than that. We have a longitudinal cohort protocol where we're going to follow at least initially 4,000 bipolar patients for five years with key metrics of neurocognition, symptom burden, MRI volume, digital monitoring, EMA assessments, but we're going to take all of that biological data and embed it in a learning health network. So our learning health network is not just EPIC. Our learning health network is bringing in biological measurements, cognitive measurements, imaging measurements. We want as much information as we can to, again, identify those new vectors of association that help us design better care. I think this visual really emphasizes it nicely. So if you start with the large group of very diverse-looking people who have different symptoms, different biotypes, we are going to be collecting in that orange top circle clinical care data, structured diagnostic interviews, symptom burden. All of that's going into a clinical coordinating center. We are going to be collecting EHR data. Anything else that's local that we can get our hands on, we're going to put in and we're going to integrate it. We're going to have in-depth biological data. So clinical coordinating, in-depth biological data, all of these in parallel processes that meet in a data coordinating center. This is where we're going to create a platform that individuals, clinicians, researchers will have access to that to look at their data, look for vectors of association, identify new discoveries that we then try to scale to a larger network and reinvest back into the practice. So depression treatment's limited. Biomarkers have the potential to inform our practice, I would argue, as our learning health network goes forward. They need to be embedded in that. And we will finally get to a place where we're generating new knowledge and driving innovation to better individualized care through a learning health network. Thank you very much. We have a couple moments for questions if people want to go to the mic. Hi, I'm Ralph and I did a post-doc with Terry Ketter at Stanford and he said anything that Mark Frye does, you should pay attention to. And I was an engineer, I went to medical school when I was 40, I did a Ph.D. in control systems and modeling of systems. And I think it's, this is great. I saw a presentation by Dr. Marmer and I wondered, it seems like that's sort of a subset of what you're trying to do. Where he, at NYU, he's done an analysis of 350 biomarkers and then decoupled it and find the ones that are whatever, eigenvalues, whatever it is. How does that relate to your plans? So I'm aware of Dr. Marmer's work. If memory holds true, that's primarily PTSD. But I think the concepts are the same. Really having large data sets with as many different types of data that are aggregated to then look for vectors of association. This initiative is just getting started. The definition and the goal is to bring in as many collaborators and as many data sets as positive to really harmonize and then really harness that information for better care. So we look forward to speaking with him or hearing about that from you at some point soon. Yes. So I remember as a young attending working at the Brooklyn VA and I had a patient who was in a stuporous depression. I put him on fluoxetine and he went into one of the most violent manias I've ever seen. And I always remember that experience. But this morning there was a meeting on bipolar depression and Dr. Gitlin offered the idea that maybe for all the anguish that psychiatrists feel about flipping people into mania, that it doesn't happen that often. And he cited a Swedish study that indicated bipolar patients exposed to antidepressants had a low rate of switch and that the number needed to harm was something like 112. And so although it obviously it does happen and when it does it can be very distressing but that doesn't happen that often. And so I'm just throwing that out there is that is this something that we really need to be deeply concerned about. So I might have you let's see if we could have that patient from the Brooklyn VA answer that question because I have a feeling they will say yes. And I but you know if I I'll I'll join your misery. You never forget these cases a patient of mine and I remember really focusing on lithium because she was a hairdresser and we were really talking about the tremor but she got she was getting more and more depressed put her on bupropion and within two to three weeks she missed her appointment. I called her parents they informed me that she got manic was in a high speed chase was pulled over by law enforcement and she went to their car as opposed to the law enforcement person coming to hers that got her promptly arrested and she sat in jail getting progressively more manic for several days. So you never you never these experiences never leave you. And I think that's exactly what fuels this initiative. Is this a is this something that's going to happen on a regular basis. Probably not. Can we learn more about how to prevent it. Absolutely. All of our cancer chemotherapy docs they have rare events but they spend a lot of time thinking about them because they are so disastrous and sometimes they can be lethal patient of mine. Early days of LaMotrigine presented on rounds Monday morning with sunburned ears. The next day she had a rash over her entire body. The next day she was in a burn unit. I mean these you these never leave you. And I think learning health systems allow that end of one to really make a difference because I have a feeling if we start to look more broadly these rare events aren't going to be so rare. I think just a personal interest for me is I think we're starting to understand that the induction of mania with an antidepressant is that drug. But at some level it's also the underlying biology of the disease and if we can study them together we may better understand what it's all about. So I would say these infrequent events have meaning and we finally have a technology and in a learning environment we can probably better understand them in ways that we haven't before. Yes. Thank you very much for the presentation. I was very curious about the mitochondrial data that you presented and do you think that based on that that there could be a superiority of one antidepressant instead of another like you showed that acetalopram is an inhibitor of mitochondrial function instead of peroxetine being a stimulator. So do you think that perhaps studies could come that would prove specificity of one SSRI instead of another in terms of better improvement in depressive symptoms or less risk of manic switches. Great question and I think this is this question really thematically thematically is linked to when does discovery data change the clinical practice and I would argue not right now. These data need to be replicated. I simply bring it forward because we need to start thinking that these drug therapies work very different than how we conventionally describe a selective serotonin reuptake inhibitor. That's what that pair wise contrast was all about. So we would say let's actually replicate this finding. We have some mitochondrial medicine colleagues that have taken those very antidepressants that were of concern for us and are now doing very detailed mitochondrial assays to look specifically as to what's impacting electron transport chain to better understand what that might be mechanistically. Would I not prescribe an antidepressant right now because of that mitochondrial finding. Probably not. OK. Thank you very much. Hi. Hi. I wondered if you could touch on you refer to ketosis as a possible treatment. Yes. So there is great interest in looking at therapeutic ketosis for bipolar disorder. The literature historically really goes back to 1921 with the first publication in the Mayo Clinic proceedings of therapeutic ketosis being a treatment for pediatric epilepsy. And the time course of anticonvulsant effect was short. And when that young patient population went off of a ketogenic diet seizures returned. So if you start thinking about neuronal stabilization in the in just the broadest way. And we have lamotrigine diva perk sodium electric convulsive therapy is actually a changes neuronal membranes as well. There was interest in starting to think about the potential role of therapeutic ketosis. So there are a number of anecdotal observations. One is actually Matt Suzuki who had a and this is on in the public domain but had not responded optimally to conventional pharmacotherapy for bipolar disorder and has achieved remission with a therapeutic ketosis intervention. There are a number of open trials I believe five that are nearing completion and will be presented within the next year. I believe the first will be at the ISBD meeting in Chicago. Stay tuned. So Phil I've got a question for you. I think the people who are not who are very new to this area that are novice in their kind of read of this literature. Help us figure out what the lexicon is the vocabularies is not intuitive to many. Anything you might recommend reading or thinking about to better address that. I think all of the challenges have come up you know I mean it's very interesting to see you know the state of the science you know and how would a learning health system or network help. And I think you know it's all been captured in many of the comments. You know today we have a lot of anecdote right. We have a lot of small studies and you know they're often underpowered and they force us to you know find look at candidate gene variants or you know take our best guess at a mechanism. Right. And I think the goal of a learning health network you know ultimately is being able to just aggregate all of our collective knowledge and data so that we can actually you know not do candidate gene studies do unbiased you know analyses of you know the whole genome. Could we you know invest could we do systems biology you know and not look at a single you know level or dimension you know of data but look at you know interactions across levels you know data that that's only going to be possible if we integrate you know high dimensional data on a large number of people capturing real world interventions and the outcomes from them. And I mean I think so you know in that kind of simplistic way I think that's that's literally what the learning health system concept is designed to overcome. It's it's kind of taking our research efforts which may be small non generalizable looking at a single factor a single level of data you know and just kind of bringing it all together in kind of a holistic fashion. Excellent. Well you have stayed until the bitter end. We are grateful. We hope you've enjoyed your APA meeting in San Francisco and look forward to seeing you next year in New York. I think right. Thanks everybody.

patient-centered research

bipolar disorder

clinical practice

learning health systems

Dr. Mark Fry

Dr. Philip Wong

molecular medicine

mental health treatments

deep learning health system

real-world data

machine learning

BD Squared initiative

genomics