Episode Transcript
[00:00:00] Speaker A: The reality is we have a workforce shortage, we have an increasing demand, and sadly, many people who want access to mental health care are not getting it in a timely way.
[00:00:17] Speaker B: Well, Jordan, thank you so much for having this conversation with me. I've been looking forward to it for quite some time. You're professor of psychiatry at Harvard. You are the chair of Psychiatric Neuroscience here at Mass General. And for my understanding, maybe for the audience, what's the difference? What does that mean?
[00:00:34] Speaker A: Well, an endowed chair usually comes with a name either of an individual being honored doesn't mean you're chair of the department. It means it's an endowed chair that is sort of an honorary thing that also comes with certain other perks.
[00:00:52] Speaker B: Certainly an expert in the field of precision psychiatry and neuroscience here. So thrilled to be having this conversation with you. Maybe I'll give you just a moment to introduce yourself and maybe a thumbnail sketch of your career and how you kind of got to where you are today.
[00:01:07] Speaker A: Sure. Well, I'm currently the associate chief for research in our department of Psychiatry here at Mass General Hospital. I'm the director of our center for Precision Psychiatry. I also direct the Psychiatric and Neurodevelopmental Genetics unit in our center for Genomic Medicine and I co direct the center for Suicide Research and Prevention at MGH and Harvard.
It's been a long, long journey.
I think I got very interested in these areas in college as as a major in psychology here at Harvard. And then I took a long time trying to decide what I was going to do with my career. Ended up going to medical school after working in a business and doing some other things. And then I was debating psychiatry versus other things. I ended up doing my medical internship and residency in psychiatry. And then I wanted more research training. And so while I joined the faculty, I also got a master's in epidemiology at the Harvard School of Public Health and then a doctorate in epidemiology and did some postdoctoral work, et cetera. So I actually ended up finishing my degree 18 years after graduating from college. What that ended up doing was really giving me some grounding in the very broad range of domains that I actually now use all the time. So psychology, psychiatry, medicine, computational methods, biostatistics, epidemiology, molecular genetics. I ran a lab, molecular lab for a little while and now increasingly leveraging those to better understand machine learning and AI and so on. So it was a long journey, but it's sort of. And partly that was just because I couldn't make up my mind, but I kind of accumulated A lot of domain knowledge and experience and that allowed me to have the career that I have now, which I love, wouldn't trade.
[00:03:21] Speaker B: I love that.
Now I'm hearing a lot of chatter about this idea of precision psychiatry being the next frontier. How, how long ago were you tipped off on that and what gave you the confidence to really make that your platform?
[00:03:36] Speaker A: Well, it's interesting.
I think I really kind of came to what we would now call some of the methods.
First of all, I started doing psychiatric genetics more than 25 years ago, but there wasn't much going on until after the human genome was mapped and then sort of an explosion of biomedical and omic data.
But I also got involved with biomedical informatics in a project that was using broad scale data to try to predict who was going to be treatment resistant when it came to antidepressant treatment. This was back in probably 2000. And that led to a whole series of research of using these kinds of modeling approaches for prediction and for large scale phenotyping or diagnosis in lots of different areas, suicide, mood disorders and so on. And then it sort of built from there and, and I also was involved in, in helping with leadership of our biobank here and then some big precision medicine studies outside of psychiatry.
And it was, it became increasingly clear that this was really a new set of opportunities that needed a lot of focus but had a lot of potential help.
[00:04:57] Speaker B: Just lay the groundwork here. What is precision psychiatry? How would you define it then? Maybe for patients at home, how should they be thinking about that impact on them?
[00:05:09] Speaker A: Well, I think about it in a broad sense as the application of precision medicine to psychiatry. And precision medicine is essentially the idea of accounting for individual differences, individual variation in trying to improve how we diagnose and treat and prevent illness. And that's the emphasis is sort of not moving beyond a kind of one size fits all. And in psychiatry, often a trial and error process of finding the right treatment for the right person at the right time.
It's a pretty broad umbrella, I think, but when I think about, involves several main areas, main tools and resources. So one area is improving how we identify factors that put people at risk of an outcome or that are resilience factors that are modifiable factors, and using big data and machine learning often to do that at scale in a way that we couldn't do before. So there are many outcomes in psychiatry that sadly we don't have a very good way of predicting or even sort of knowing who might be most at risk. And that makes it hard to apply Prevention in an effective way. Another area is matching people to the right treatment at the right time or treatment stratification. We sometimes call, and again, in most cases, when we have effective treatments for psychiatric disorders, and they're helpful for lots of people, sometimes life saving, but for too many people, they don't work as well as we would like. And some of that is we don't know what's going to work for you as an individual, but now with some of the tools and technologies and methods that are available, that's possible to do so again, moving besides beyond this sort of trial and error approach, and frankly, people do.
If people are seeking treatment for depression, let's say, and we decide together that an antidepressant is indicated, I don't really know ahead of time which one to look for, what's going to work for you. I know what on average works, but on average they all kind of work the same. And so people sometimes spend weeks, months finding the right treatment. And meanwhile there's a lot of impairment and symptoms and suffering going on.
[00:07:38] Speaker B: Side effects.
[00:07:39] Speaker A: Yeah, side effects.
Another area is actually using, partly for treatment prediction, but using other tools. We rely on this dialogue, symptoms, what people say. But there are other tools that could be used that might give us insight into either what somebody might be at risk for or what treatments might be effective. And those are things like clinical neuroscience tools like imaging or EEG or some of the technologies that are now available.
And then there's the area of developing new treatments. So many of the treatments that we have, certainly on the drug side, although this is really true on the psychotherapy side as well, are based on insights that are decades old. And only now are we starting to see some fundamentally new treatments. But even then, most of our treatments are not based on what we know about the etiology or causes of illness. They are found by chance or a savvy observation that somebody made.
And the idea now is that we might be able to better target the development of new treatments and who might benefit from a particular treatment based on what we know both about biology, neurobiology, but also the other characteristics that people bring with them to treatment that might make them more or less likely to respond to a particular mechanism or avenue of treatment.
So it's a lot of possibility, it's still early, and we're always sort of threading that needle of conveying the hope of it without falling into hype.
[00:09:24] Speaker B: And we've seen a lot of hype in the space. So maybe take me into the practice. I'm a patient, I Come in. What is the current state of what is possible and what we can do?
[00:09:34] Speaker A: Well, it depends what you're coming in with. So if you are coming in with a, let's say an anxiety disorder. We have actually pretty good treatments for anxiety disorders. In fact, the psychotherapy treatments for things like panic disorder or social anxiety disorder or phobias are remarkably effective for a lot of people. Not everybody. We have medications that we use. Again, most of them were developed serendipitously and we don't know which one's going to work best for you. There is some evidence that combining those two is more effective. Similarly, for depression, we have a set of first line antidepressants, for example, I don't know which one's going to work for you best. And on average, they work about the same. On average, about 50% of people will respond to an initial trial of an antidepressant. And actually, on average, their effect is pretty modest.
But we know that some subset of people will respond remarkably well to a given medication. Others won't. We just don't know why or who. And so people, as I said, will often spend a long time through this trial and error process of trying to find the right treatment for things like psychotic illness, schizophrenia, et cetera. Again, we have treatments that can be very helpful for people. Many of them have a lot of side effect burden. They're often difficult for people to stay on. And again, we don't know at the beginning what's going to work best for you.
And for bipolar disorder, another important illness that we treat, again, we have effective treatments. Same story, though we don't exactly know what's going to work best for you there. We're trying to often treat both the mania side of bipolar illness and at other times the depression side. Those are different treatments sometimes.
So it's complex and you know, in terms of what we have to go on as patients or as clinicians to optimize treatment, I have to say it's still fairly limited.
[00:11:49] Speaker B: Well, this is a great jumping off point for us to talk about some of the tools that you hope to be developing. It sounds like we're in the early stages of understanding really what's possible and how it might apply. But maybe talk me through your current thinking of how it should all work.
[00:12:04] Speaker A: Well, I think that that partly depends on which problem we're tackling.
One thing that, as I said, I direct the center for Precision Psychiatry here, and our kind of motto is innovation to implementation. So let's ask and try to tackle questions where we're bringing something new, but there's a path to actually making a difference and for patients, for clinicians, et cetera. And that can differ. So one thing that we think has a lot of potential is doing better at identifying people at risk. And we've done a lot of work in this area with the problem of suicide, which, as you may know, is a sadly growing problem.
Rates have increased, relatively speaking, by about 30%, 5% over the last couple of decades, and maybe almost twice that among young people.
And so we've spent a lot of time using approaches that might do better than what clinicians like me are able to do because they can integrate a lot more information, they're more powerful things like using machine learning with data sources that could be electronic health records, could be maybe digital or wearable kinds of data streams as well. We've been able to develop algorithms that seem to perform quite well in a variety of settings. And we're now in the process of doing a randomized controlled trial to look at whether that actually does make a difference.
Similarly, we've done similar work using AI to try to get around this problem or tackle this problem of what treatment is likely to work for you could we know and provide decision support to clinicians early on so that we don't have to go through this long odyssey of picking a treatment. And again, you know, we've had some success there.
And one of the things there that I think we keep in mind is some of these things don't have to be perfect perfect. They have to be better, you know, distinctively better than what we're doing now. And I think that is within reach for certain questions incorporating genetics and genomics is and other omic data. So that could be gene expression data or protein data or those kinds of things.
Is a part of this where I think the genetics is likely to play the biggest role is in helping us identify new treatment targets, validating them, speeding clinical trials, because you're basing the trials on biology that is related to the illness, the target illness, and even doing it in a way that's not a one size fits all way, but is actually tied to mechanisms that may not be the mechanism for everybody, but may for a subgroup of folks. And that could help us do much better. So, you know, that doesn't obviate the need for the whole drug development pipeline, but it puts you on more solid ground. And now we have data that and assays and biology that we couldn't have looked at in the past through things like, you know, induced neurons and Organoids and the use of single cell sequencing and much more fine grained biology for the target organ of interest for us, which is the brain. I see this working on many different fronts in parallel. But then you ask, what about today? So which of these things has really changed the game so far? I would say not much so far. We're at the beginning now. The one thing that has been more widely implemented in practice that's along these lines is pharmacogenetic testing. So there are recognized and in fact some of these genetic tests are in the labeling of medications. There are recognized polymorphisms or genetic variants that seem to influence especially the metabolism of common psychiatric drugs. And there are platforms and companies, commercial interests that are marketing and making these available to clinicians. And again, most of that is based on a fairly limited set of genetic tests, which are those for, for those who are interested in this area that we think of as pharmacokinetic targets, meaning they have to do with how different people metabolize drugs. And so that might mean that you in theory need a higher dose of this drug than the other somebody else because you're breaking it down, let's say more quickly. Or you might need a lower dose because you're not breaking it down as quickly as other people.
[00:17:08] Speaker B: So supposing it works for you, you can kind of modulate on dosing, right?
[00:17:13] Speaker A: And maybe it's just more effective when you have that information because your clinician, let's say, is guided by it there. I think the evidence is mixed, frankly. And on average, if you put studies together, there does seem to be a benefit in the controlled studies that have been done to date in the area, at least of antidepressants, of using these tests. There's a statistically significant increase in some outcomes that desirable outcomes like remission of depression, let's say when you compare using the clinicians who use the test to those who were getting treatment as usual.
But many of the professional societies and folks who have looked at this in detail have overall come to the conclusion that it's not ready to be a routine part of clinical care at this point because the effects are, you know, the benefit is still not totally clear.
And so that's not to say that it can't be very helpful in certain cases, but it's still a fairly limited set of information that people are getting. And also how it's delivered can affect its usefulness. For example, sometimes, you know, it's portrayed as, you know, this is a drug you should stay away from based on genetics and These are the drugs you should use. And yeah, that's not, it's. Since it's not the full story that can be misleading. So there might be a medicine that's on the use with great caution list, but actually it turns out the person's already on that and they're doing well or something like that.
[00:18:55] Speaker B: So you've given us maybe an interesting roadmap to double click into. I heard maybe three things on it. The first one around suicide prevention, risk mitigation, probably nearest to term, ability to implement and translate. The second one around treatment planning, and then the third one around drug development and how to even think about looking for new frontiers. I'd love to get into each of them. Talk to me a little bit about just some of the things that you are starting to look at data wise and the models you're trying to build around flagging risk.
[00:19:30] Speaker A: Well, one of the things that we're very, that we've spent a lot of time with are real world health information data, for example, electronic health records. Because that is a very large, ever growing data set. Because every time we see a clinician, all kinds of aspects of that clinical encounter, about a real world clinical encounter are captured in the electronic health record.
It creates a very large resource of data that is obviously clinically relevant. In the past, it would be hard to make use of that level of data in some kind of really comprehensive way. The reason that we are in part particularly interested in that in the realm of suicide risk and prevention is that most people who attempt or die by suicide are seen by a clinician, a healthcare provider, in the month leading up to it. So that means that we have an opportunity, maybe an obligation to try to take that opportunity to do better. Many people do not disclose those kinds of thoughts and those, those thoughts or intents are often fluctuating. So they may not actually be happening at that moment. But with the help of machine learning, artificial intelligence applied to large scale data, we have shown that we can build models and it actually works with a variety of different model strategies. So it could be deep learning AI, or it could be more traditional machine learning, or it could be even certain forms of regression analysis. They all can do pretty well because they're making use of way more information than a clinician could make use of. And especially when you pair it with, for example, something at the point of care when you're seeing the patient, like a, you know, a measure of how they're feeling or to what extent are they experiencing these things. It does seem to do well at Least in the studies that have been done, we need to test it in a big randomized trial. And that's what we're endeavoring to do now.
But I think that opens up a new area. Again, this is one of the most obviously serious and tragic outcomes in our field, and yet we have not been able to do much better for a long time. And we've also shown that using these models can be cost effective when implemented in clinical care, especially when coupled with evidence based prevention strategies. That's another area where we need some work about what are the things that are most effective for an individual patient. That's another thing we're looking at. So, you know, developing what one might call precision treatment rules. Meaning given this set of history or characteristics or personal, you know, profile, what's the most effective thing to do for that person?
[00:22:36] Speaker B: What is the predictive power do you think that we have today for, for this kind of risk?
[00:22:45] Speaker A: Well, that's an interesting question because predictive power is a perhaps more complicated phrase than it might be because it partly means what do you, what's your metric of prediction? So often in this field people use things like the sensitivity, specificity or the area under the receiver operator curve. Curve, which is a measure really of how good is your model at distinguishing people, let's say, who are at high risk or go on to have the event versus don't.
But clinicians care about other things like what's the positive predictive value? Which means given that the model, let's say, thinks that this person is at high risk or whatever it is you're trying to detect or predict, what's the probability that they really are?
Which is a slightly different question. And similarly, the negative predictive value, that is the model doesn't think this is someone at risk. What's the probability that they're really not?
Another extension of that in some ways is what people think of as decision curve analysis or net benefit. Where is it? Where in the risk threshold is it worth using a model like this?
[00:24:04] Speaker B: Presumably you're solving for minimizing false or actually you'd be okay with false positives. And the thing you'd be solving for is minimizing false negatives.
[00:24:15] Speaker A: Yeah, yeah. Although not always. Not always. I mean, false positives and false negatives both have their real downsides. And so if you think about the false negative side, if it's a really, you know, tragic outcome, you certainly would, you might err on the side of having a, you know, more false positives be okay than missing something that's really terrible. On the other hand, the false positives, depending on what it is, can also have their burden. That is, somebody might be inappropriately hospitalized, right? Or somebody might be, you know, there might be stigma attached to thinking that one has been designated at risk for, let's say, something some psychiatric illness or whatever. But you're right that I think mostly we want to make sure that we don't. Well, two things. We want to feel confident about the positives and confident about the negatives. Because if you know that somebody really is at very low risk of the thing you're worried about, that has a real impact too in decision making and, and can be very helpful. Now the thing about these positive and negative predictive values, and the reason it's sort of complicated is they depend in part on the base rate of the thing you're dealing with. So let's say it were antidepressant response. So there the base rate, the prior probability, let's say, is maybe 50, 50 that you will respond or not. So it's pretty common. Both of those are pretty common for a suicide attempt or death, especially over a given window like the next six months. While that is a huge problem and tragic outcome, it is relatively uncommon. So it might be half a percent. And because of the way these predictive values are calculated, you're always going to have a relatively low absolute risk if the thing itself is pretty rare. So the question isn't, would you, in those circumstances, would you use this to as the determinative factor? At least we think of it as you would use it to alert clinicians that this is somebody who might require further evaluation. Interesting.
And it's always going to be coupled with clinical judgment, the human in the loop kind of thing, because these have error rates and they are also not necessarily tuned. Even if you use a kind of point of care assessment, you're asking people how they feel right now. Not just all this historical data, some of that doesn't translate to the person's individual circumstance, which the clinician may have a greater insight into.
[00:26:56] Speaker B: It sounds like you're starting up or you're in the process of doing larger scaled controlled trials on this.
When might you expect us to have definitive answers on a path forward here?
[00:27:09] Speaker A: Well, we're hoping that this study, which is part of our center, which was funded by the nih, is going to be, is going to have some answers within the next two to three years. But we're not the only ones doing this and there are many studies going on now and we are in fact part of a network in this center with a lot of really smart, dedicated people tackling this problem. And it's hard to imagine that it's not a top priority for not only those of us in this field, but everyone.
[00:27:42] Speaker B: I'd love to hear a little bit more. I know you just published a paper on ways to do treatment planning based on depression in particular. Tell me a little bit more about that and the findings.
[00:27:52] Speaker A: Well, in that study, we were applying similar kinds of ideas using artificial intelligence to assist with trying to predict which of the four first line kinds of antidepressants people are typically offered would be most effective.
And in that case, this was led by Ehan Hsu in our center.
In that case, we used the large amount of electronic health record data. In this case, 17,500 or so patients who were started on an initial antidepressant. And the first line antidepressants that are sort of standard are certainly the SSRIs, which we're all familiar with. Probably the, you know, Fluoxetine or Prozac or Escitalopram or Lexapro, those kinds of medicines. The Snris, which would be like Venlafaxine or Effexor is the brand name, or Duloxetine, Cymbalta and others, Bupropion, which is also known as Wellbutrin, and Mirtazapine, which is also known as Remeron. And again, if you look at the studies that are available, it's about equal in terms of what, on average, people respond to.
In this case, we built a variety of different models so that we could compare what would work best that used electronic health record data, that used natural language processing, which is another way of extracting information from the records.
In some cases using deep learning, in other cases, more simple methods.
What we found was that we could build models where the positive predictive value was about 75%. So 75% of the time, the model could make a prediction that was accurate about what a person would respond to.
The thing is that what you can now do is not simply look at, does this person respond to an ssri?
But which of the four. Which is the real clinical question? If you tell me. Well, they're not. That's the relevant question. I want to know what should I do?
[00:30:08] Speaker B: Right.
[00:30:10] Speaker A: And, you know, if you look at what the model was looking at, there were a variety of factors depending. Yeah, yeah. So it. And it depended on the model. But you can look at the importance of different factors, and there were some, you know, more or less Obvious ones and others that, you know, were probably not so obvious.
The thing about these importance factors, and this is an issue we come up to a lot in these kinds of predictive models. We are building these models to optimize prediction. They don't always represent causal factors.
So the thing that might be a strong predictor might not actually be the causal thing that you would intervene on. It might be correlated with a causal thing that. That just didn't have the power to, because of its frequency or something like that, to show up.
Nevertheless, we can get a look at that, and we can see how it would change clinical decision making. Because for some people, the SSRI was clearly the best choice initially, and for others, it was one of these others that might not be everybody's first impulse.
[00:31:28] Speaker B: Can you give me maybe an example of something you found, like a set of patterns that led to a specific recommendation?
[00:31:34] Speaker A: Yeah. And let me also caveat this by saying this is all sort of in silico, meaning it's all using data that exists already. It is not actually implementing these things yet in clinical practice. But if you look at some specific instances, we see a patient who had a really high probability of responding to an ssri, lower to the others, and another patient who, you know, again, off the bat, we would have said they're probably going to respond to any of these about the same, had a really low probability of responding to an ssri. And mirtazapine, for example, comes up as the top choice that would make a difference if you immediately reached for the one that was more likely to work. And in terms of what was. You know, there were some factors that have to do with, like, the severity of depression or the comorbidities with other illness. But some of them were things like marital status. You know, it does allow us to get a little bit of a look at what are some of the factors. There are other methods that can be used to sort of enhance the explainability or the causal nature of what you're seeing. And we didn't do that in this case.
[00:32:43] Speaker B: Explainability with deep learning is big issue. It's a big issue.
Okay. So that. That helps me start to paint a picture on SSRIs, SNRIs, but then choosing between the brands, how does that happen?
[00:32:58] Speaker A: Well, it would happen conceivably if you had enough data, right? We didn't have enough data to get there. Yeah. And arguably that's less of a significant choice because you would think, you know, the way we make these decisions currently, if you're starting with an SSRI is often things like what's the, Are there some differences in expected side effect profile? And let's say that for one medicine there might be a little more weight gain. And here's a person who's depressed and has lost a lot of weight. Maybe we go with that one or somebody who has, you know, gained a lot of weight, which also can happen with depression. We might go with another one.
So those are some still fine tuned things. But if we had a sense of which class of medicine we should start with, that would be a big help.
[00:33:51] Speaker B: And the pragmatic answer is that we need to start there before we can get any fancier. Now maybe the kind of third leg here is around drug development.
What are you hoping to see there?
[00:34:03] Speaker A: Well, what I'm hoping to see is really a kind of comprehensive and systematic approach to this. First of all, we're hoping that pharmaceutical companies remain engaged in this space because there was a period where there was a real retreat. It's a little bit of a renaissance. And that's good because the fact is that actually developing and bringing a drug to market is not something that academics alone are likely to do. So. So we have again some real opportunities I think that have played out in other areas of medicine and precision medicine. Using genetics, for example, to de risk to identify and de risk targets, meaning in the past to figure out like what compound would be worth developing. There might have been an observation, there might have been an animal study or something like that. We might have a hypothesis about what's going on. Maybe depression is too little serotonin or something like that.
Those were always oversimplified.
Now we have through genetics, a more direct look at what is the underlying biology of many diseases and each genetic variant or the pathways that it illuminates. As you accumulate more and more insights about the genetic architecture, we would call it, of the illness starts to point to specific human biology. And there's evidence now that it's actually been for a while, a decade or so that if you start with targets that have genetic evidence behind them, you effectively double or maybe triple the likelihood that they will succeed all the way through to phase three and then marketing.
And that's because you've grounded it in real biology. You can also, when I said de risk it, part of that is just feeling more on solid ground that it's a real target. And part of that is because there's variation in genetics. You can look at what's the effect of more or less of this gene product, let's say, right? So that could be helpful in knowing first of all, is it, do you want to develop an agonist or an antagonist? Also, what about the off target effects? If you, you know, you have the biology that suggests that you would be protected from the illness, let's say, or it would be a good thing. But maybe that also incurs problems in other organ systems or other aspects of your physiology that would make it kind of too risky to spend a whole ton of money and time trying to develop it. And maybe we should pick another compound that does that targets the same thing, but doesn't have that component. So that whole thing I think is great. And then on the big data and AI side, I think people have been advancing quickly in using AI to understand the biology of proteins, the biology of gene expression in ways that again were not really possible, and to screen now large libraries of either actual compounds or designer compounds that could be imagined and that could short cut some of, of the extensive work that goes into even beginning to develop a trial.
Also we can imagine now using some of these precision techniques or techniques that help us parse the heterogeneity. The fact that people, for example, with depression or schizophrenia or whatever it is, that's not one thing, it's many things. And there may be many different components to it and we can now use potentially, and there aren't a lot of these. But the hope is that there would be more biomarkers or genetic predictors that might suggest it would be good to target a trial to folks who appear to have this as the basis of their illness. And that again might make trials much more efficient and then along the way, companion biomarkers that might be predictive of outcome. And we've seen some of this in academic and also industry efforts to use brain imaging or electrophysiology to predict is somebody likely to respond to, let's say, an antidepressant or maybe one of the other new therapies that we have, like transcranial magnetic stimulation or something like that. So there are a lot of fronts to move forward on.
Still hasn't happened in clinical practice yet.
[00:39:04] Speaker B: Though I'm sure there'll be many years before some of the targets discovered this way make its way into clinical practice. But we do have a fair amount of industry folks who watch this show. What are the targets you're most excited about?
[00:39:17] Speaker A: Well, I think that, you know, if we can, and this has happened now, especially in say something like schizophrenia, where specific genes and both common and rare variations in a certain set of genes are fairly convincingly associated with the biology of the disorder. Now, some of those involve systems that we kind of already were pretty interested in, like glutamatergic synapses or things having to do with the structure of the synapse or the function of the synapse, let's say. But these, I think, are. Are pointing to biology, which now has a pretty good set of evidence behind it. That is, it's not just one target, but it's. It's a set of targets.
And those again, are opening up new areas. Some of them were wide, were genes that people hadn't really particularly thought about.
Another thing that we're finding, which I think is fascinating in this area, in the area of psychiatric genetics, and not surprising in a way, is that the genetic contributions to many disorders are actually pretty overlapping.
So the genetic basis of, let's say, schizophrenia and bipolar disorder, which we think of as two very different things, things, in some ways, those were the two main disorders that kind of started the whole classification system.
They're mutually exclusive. You can't have both. And yet at a genetic level, they're very highly correlated. That is, the genetic variations that are contributing to one of them are largely overlapping with the other. The genetic correlations are about 0.7. It suggests that the biology is shared. It also suggests, and we see this with other disorders as well, that there are some, you know, when we model this, all of the genetic data across disorders, you see some fundamental underlying genomic factors or latent genomic factors that capture the variation across whole groups of disorders. So schizophrenia, bipolar disorder is one. Then there's the realm of kind of depression, anxiety, post traumatic stress disorder. There's the realm of compulsive disorders like OCD or Tourette's syndrome or even eating disorders.
So that suggests that our way of classifying psychiatric illness, as you may know, is based on a consensus of experts trying to put together the best evidence that they know of and coming up with criteria which are not at all based on cause or mechanism. In fact, they're explicitly not supposed to be. They are, you know, this set of symptoms or signs, and if you exceed a certain threshold and it's associated with impairment, you have that disorder. But even from the beginning, people cast those as kind of provisional descriptions, even though they've become sort of, you know, the basis of drug approval and diagnosis and our popular culture view of these things. But, you know, the chance that people sort of came together and settled on a list of things that are actually the relevant boundaries is, you know, pretty low. Yeah. And for a long time, you know, people have known that, you know, any given person may not fit the criteria, even though they clearly have an illness. And it's got certain aspects of this and certain aspects of that. So in some ways, it's not surprising that there's genetic overlap in the categories that we've defined. But it also suggests that there may be biology that that is broadly applicable and that itself could be the target of new therapies, often in ways that one of the limiting things about treatments for disorders and the way that drug approval, for example, is set up now is that it is based on these diagnostic categories, dsm, ICD categories. But really what affects people, people's lives often are the symptoms, the prognosis, the things that transcend a diagnosis like cognitive impairment or psychosis or depression, which is not limited to a single diagnosis.
And this is consistent with that idea that we should also think about these sort of transdiagnostic phenomena as treatment targets, as ways of thinking about what is the clinical formulation for any particular patient.
[00:44:08] Speaker B: Yeah, and I know you've done a lot of work on this idea of transdiagnostic dimensions. How should we be thinking about that? I know you're challenging almost the kind of classic diagnostic paradigm. What's your view on how it should happen?
[00:44:25] Speaker A: Well, I think it's a really fascinating. I mean, it's one of the most fascinating questions to me, but also a challenging question 1. To me, it makes sense that, you know, the brain is the organ of, you know, psychiatry and neurology. Right. And our brains were subject to evolution. They evolved to, you know, solve certain problems. And we have systems and circuits that in some ways are tuned to solve certain problems. And many of the things that I think we think of as psychiatric disorders are variations along that degree of functioning of a given system. We don't have a clear map of that, but we have some of them. In fact, one of the counter proposals to the classic DSM diagnostic system is something called the rdoc, or research Domain criteria, which is the idea that there are some fundamental underlying dimensions or domains of brain and psychological functioning and social functioning, that what we see and call disorders are maybe reflections of that. But if we actually could go and understand what was the fundamental nature of these things, that would be very helpful. Now, that's not an easy thing to do. You. You have to know what to attend to. Where do you decide that these systems are? And in some cases, I think we've gone further than others. For example, fear and anxiety. People have understanding of some of the core components in the brain and in behavior that go into phobic behavior or fear behavior. And to the point that it's informed how we understand things like post traumatic stress disorder or, or phobic disorders for other things. We have very limited understanding at this stage. But I do think it's something that we need to pursue as a project to understand at that level beyond simply the surface convention diagnosis level.
[00:46:37] Speaker B: So is the vision that you might get a spider chart across these dimensions for an individual? Like if I did a personality test, I might get like a spider chart. But. And then as a, as a clinician, you'd be able to look at therapies, therapeutics that target specific dimensions of this as opposed to having a categorical.
[00:46:59] Speaker A: I think that is one of the visions, but that sounds easier than it might be to get there. And the reason in part is while these diagnostic categories are kind of construct, they are constructs as defined specifically.
They're also useful in the real world until you have something better to replace it with. And it might be that what you just described, that would maybe be more actually accurate, precise, but you'd need a lot of data to validate those. It's not impossible, and I think we're getting data of that scale.
But the reality is when somebody comes in today and wants to understand what's going on, what are the treatment options, etc. Diagnosis is pretty essential. And diagnosis is often a categorical thing. It's not always because people talk about blood pressure being elevated.
We do have a category of hypertension, but we recognize that's a little bit arbitrary, but it's still useful because that tells me when I maybe feel obligated to be treated.
So I'm a little more reserved about some of the speculation that we could simply, and I don't think really anybody is at this point entertaining this. Simply throw out the diagnostic system and put in these new, more scientific dimensional approaches at this stage. Because I don't think in the real world that's an easy thing to do.
[00:48:36] Speaker B: Or is it a layer underneath or a layer on top? I know that there's some work being done in cognitive impairment with schizophrenia.
[00:48:43] Speaker A: Right. I think that's right. And I think that that's a more near term opportunity. That is, we know there's some core symptoms which probably reflect core biology in the brain, for example, that should themselves be targets of treatment. And for example, our current antipsychotic medications don't do very well with that component with the cognitive impairment, which means a lot to people.
And so recognizing that this Is a thing characterizing it and then targeting treatments to the ones that are really important for people's quality of life and functioning is a way forward, I think, rather than the sort of whole scale. Will we reject the categorical diagnosis?
[00:49:27] Speaker B: One of the things that has garnered a lot of attention recently is the application of psychedelics in this space. I'd be curious how you view that.
[00:49:38] Speaker A: Well, I haven't used psychedelics in my own practice, but I have patients and obviously I've read the literature and I would say that it's exciting that there are new options that are based on new mechanisms and some of them. And you know, when you say psychedelics, sometimes that is used narrowly or broadly. So sometimes it includes also things like ketamine or anesthetic, you know, that family of compounds. And sometimes it's more focused. People are talking about psilocybin or classics. Yeah, the classics.
And I think those are maybe different structures, stories so far in terms of what we know about why they would work or that they work.
I'm pretty.
I feel good about the idea that ketamine has an antidepressant effect that's different from what we've seen before and that for some people, again, can be really helpful. Again, for other people it's not. And there's no reason to think that there'd be a one size fits all approach. But that's exciting, especially when we think about folks who have tried multiple treatments and those treatments have failed them.
The more classic psychedelics like psilocybin, again, encouragingly, some clinical trial data, even head to head data suggesting it's not inferior to some of the treatments that we think of as really pretty effective and in some cases maybe even superior.
I think we have a lot to learn still.
In part that's because they were off limits for research for a long time and there is a lot of activity and a lot being learned. I think it does provide hope that we have a new way of approaching these things that might both be quicker or more effective for some people, for whom I don't know what's the optimal way of delivering it, the optimal dose, what other compounds might be safer but still have that effect.
And those are all really important questions. One thing that worries me a little bit is the fact that there's so much excitement about it that people have kind of taken it into their own hands and are experimenting outside of, you know, outside of trials or outside of, you know, settings that would be safe. Yeah, and that can be really problematic.
And I worry about that. I worry about the, the broad, you know, unregulated potentially and sort of just highly variable way that people may be using these things.
[00:52:32] Speaker B: Now I know that in conversation with a lot of the developers of some of these therapies, there's a question that always comes up around, who should be considering this? For whom would this be the better choice?
When do you think you'll start being able to incorporate this into the treatment planning research you do?
[00:52:53] Speaker A: Well, you know, a treatment that is as different and potentially powerful in a certain new way that we don't have our heads around is often thought of when it's shown to be effective as kind of like a second or third line option for people. Right. If you, if you didn't have a good response to the tried and true, you know, typically low toxicity treatments.
But that can change if the evidence is strong enough, especially head to head evidence and especially in the right population, like if there were evidence that it was superior to or could be used first line with better outcomes, that's viable. I don't know when I don't know the answer to your question.
[00:53:40] Speaker B: I know the Mindmed team is making a strong argument in terms of relative effect size and anxiety.
[00:53:46] Speaker A: Sure. I'd like to see more data and we have to recognize some of the safety issues and those kinds of things.
[00:53:55] Speaker B: So I'm hopeful and maybe talk a little bit more about emerging technologies that you're excited about. I know vocal biomarkers seems to be floated around a few papers published recently on the application of them. How are you seeing those start to.
[00:54:14] Speaker A: Hit the road again? Those are areas of, I think great interest and in part because the technology has arrived at a level where you can imagine them being more useful than what we used to be able to do because of AI and because of the ability to make sense of massive amounts of data and data that's of a different format than we often think about.
There are a couple of things I would say about it. One is I like the idea that we are thinking about real world data that also has the opportunity to be tied to people's, to variation over time in how people are doing because we don't have a lot of that. We have questionnaires you could administer to people, but that's not capturing their real world life.
So whether it's vocal or it's wearable sensors or those kinds of things, those do have the advantage of that and I think that's a really promising thing.
I think that there are two things that on the other hand are challenging. One is how do you think about implementing this in the real world? When would this actually happen? Would clinicians, patients do the thing that you think is really a high tech cool thing? Right. And you know, this is related to the third thing that I've seen both in this area and in AI and healthcare as well, which is people often come at this from two different directions. One are the computer scientists, engineers who know a ton about the possibilities of engineering models, devices, things that can be useful but don't know a ton about the clinical practice issues or the reimbursement issues or any of those issues. Then you have the clinicians who know the urgency of getting answers to these questions and developing something better and delivering something, something better. They don't know a ton typically about the, you know, the underlying assumptions or nature of some of the models or devices. Some of them do.
And both of those are sort of incomplete. And so clearly you need both. Right. You don't always have both though. And so you see a lot of things come out where super cool. But no, it's probably, I mean, that's a great impractical and it's not clear what the path is for that or things that underutilize perhaps the potential.
I think LLMs are kind of a little bit of a meeting point for. Yeah, it is pretty interesting because, you know, they have an underlying highly complex modeling structure, but they're also highly accessible in terms of what they're delivering and for clinicians to understand and patients to understand what you get out of them.
And so that's going to be really interesting to see how that plays. And of course that technology of these massive models can be used not just for text, but for other things. So that might be a meeting point, but still you need both because for example, let's say you're using LLMs to extract clinical information from something. We find all the time that there are certain interpretations of what that information is. For example, information in medical records that you really can't get unless you have a deep familiarity with it. I would love to see a cadre of people who are really trained in both of these worlds. And we're seeing some of that, which I think is great, but short of that at least having both and more than that. I mean, but those are two obvious fundamental components of it. I mean, you also want people who are savvy about bioethical issues and legal issues and all those kinds of things, but at the table, otherwise it's just not going to work very well.
[00:58:26] Speaker B: It sounds like kind of inherent in your Answer.
[00:58:28] Speaker A: There's.
[00:58:29] Speaker B: There's the monitoring use case longitudinally and potentially a diagnostic use case up front. Is there one or the other you're more keenly excited about?
[00:58:42] Speaker A: No, I think both of those are really potentially valuable because diagnosis is very tricky.
Sometimes it matters a lot. Sometimes it doesn't matter that much as people think it might, because it doesn't alter the treatment. Treatment options very much. And whether it's this flavor of an anxiety disorder or another.
I mean, people attach a lot of importance to the labels, but in reality, that's not the most crucial thing. But sometimes there are some real important decisions. For example, most people who have bipolar disorder first present with a depressive episode.
And you don't necessarily know during that initial episode is the. This unipolar depression or is it bipolar disorder that hasn't. A person hasn't had a manic episode, let's say. The reason that matters is that if you were to just treat with antidepressants, there is some increased risk of worsening the course or triggering a manic episode.
So those kinds of decisions are often really useful. To have better ways of detecting the monitoring thing, though, I think is really important because, again, we don't have real world monitoring. Monitoring sounds sort of 1984. I don't mean it that way, but I mean that the clinician, the patient, and the family can kind of collaborate on how are things really going in the real world. And they may be things that you aren't even aware of, like, are you sleeping? People are often not terribly accurate. I'm not. I'm sure about how their sleep is really going.
So to actually have data on that and be able to deliver it to the care environment could be incredibly useful in really tuning treatment or selecting treatment and also in forestalling things from getting worse. If things look like they're really deteriorating for somebody, they need care. And you might not know that the person themselves might not realize.
[01:00:41] Speaker B: So we've. So, Jordan, we've spent a lot of time today talking about the current state, the future state of where you hope to see things go. My kind of traditional closing question is a magic wand. I give you a magic wand and I ask if you could change anything about how, let's say, psychiatric care is practiced today, what would you change?
[01:01:00] Speaker A: Oh, my God. There's a lot of things. You know, I think partly I would change the access problem.
[01:01:09] Speaker B: Okay.
[01:01:09] Speaker A: I mean, the reality is we have a workforce shortage, we have an increasing demand, and sadly, many people who want access to mental health care are not getting it in a timely way. Many of the folks who are most deeply affected, either with serious mental illness or who are unhoused or have the resources are not getting care or the quality of care is less than what it should be.
That's a difficult problem to deal with. I mean, sometimes it's been dealt with, but it's a complex sort of web of political priorities, reimbursement and third party payer kinds of things.
So people are looking at this is another place where actually digital health is potentially going to make a difference, where people can access health care without having to get themselves to an office park, get daycare for their kid or something like that.
Telehealth of course, has expanded tremendously and that has been helpful for some people.
But the reality is just a lot of people are not getting the care they need and we're seeing, you know, just increasing distress and as I mentioned, suicide and other things, especially among young people.
That's just really a problem we have to tackle in a more systematic way. So if I had a magic wand that could fix that. Yeah, I would.
[01:02:50] Speaker B: Okay. Yeah, well, we'll close on that. I think the audience can feel assured that you're working on some of these problems and helping bring some of these technologies, techniques into the forefront. It'll take some time, but it feels like we're at a turning point now.
[01:03:04] Speaker A: We often say that and then five years later, where are we? So you never know. But I have to say I'm very optimistic.
[01:03:13] Speaker B: Bill Gates says that we underestimate or we overestimate what can happen in a year and a. We'll see where this goes. Well, thank you so much.
[01:03:18] Speaker A: Thank you.
