← AI and I
AI and I
The Future of AI in Medicine: From Rules to Intuition | Awais Aftab, Psychiatrist and writer
00:53:02
Loading summary
Transcript49 lines
- [00:00]
Dan Shipper
I have ocd, and it was a mess for me to even figure that out. It took like 10 years. And I went to a bunch of different therapists, and eventually, like, I was just reading a bunch of stuff, and I was like, I think I have.
- [00:09]
Psychiatrist Guest
OCD in my own clinical practice. One of the diagnosis that seems to have been missed by other clinicians a lot is obsessive compulsive Disorder. Ocd. As far as we know, it's not one thing. In fact, you know, even from a symptom standpoint, it has fuzzy boundaries. So we have to respect that heterogeneity, but it offers us a certain descriptive potential. We can about it, we can identify it, we can use it for guiding treatment. And so this would be a good example of something that's very helpful pragmatically. It's an entity that exists from a pragmatic point of view, but not from an essentialistic point of view.
- [00:44]
Dan Shipper
There's a lot of pressure to be able to reduce these things down into a checklist. Right. One of the things you're saying is, yeah, you can do the checklist, but really a good clinician has a little bit of, like, a smell for it. Observing clinicians and then reducing how they operate to diagnostic rules. That's exactly how machine learning researchers started in AI and exactly what didn't work, what ended up working is deep learning. Avice, welcome to the show.
- [01:22]
Psychiatrist Guest
Thanks, Dan. I'm. I'm happy to be here.
- [01:24]
Dan Shipper
Happy to have you. So for people who don't know you, you are a psychiatrist. You're also the clinical assistant professor of psychiatry at Case Western Reserve and the editor of Conversations Critical Psychiatry. And for people who are listening, who are watching, who are like, why are you on AI and die? I promise there's an explanation in particular for anyone who's been a listener of the show for a while. I've been really obsessed with how AI might change science, and in particular, how it might change psychology, psychiatry, neuroscience, more complex areas of science where science has historically struggled. And while you don't write specifically about AI, I've been reading your sub stack for a while, and we'll link to the substack in the show notes. And I love it, and I love your perspective. And I just feel like there's a lot of overlap in some of the things that we've been thinking about. And I wanted to understand more about your ideas and then talk through some of the AI stuff and see if you think I'm crazy or not.
- [02:30]
Psychiatrist Guest
That. That sounds great.
- [02:32]
Dan Shipper
Cool. So I think one of the, If I had to, like, pinpoint one of the things that I like most about your work is I think most people who work in psychology or psychiatry doing research, want to come up with a single theory that's like depression, is anger turned inward, or, you know, it's serotonin imbalance, or it's this or it's that. I think those kinds of single theories are, are, Are quite valuable. I think they're, they're, it's, it's the intellectual lineage of them, which we can sort of trace back to, like physics and stuff, makes a lot of sense, but they haven't worked that well in, in the domains that you care about. The domains that I care about. And I think you, you tend to advocate for something called explanatory pluralism. So the idea that when we're, when we're working in a domain like psychiatry, you have to bring to bear multiple different perspectives to, to really get a sense for the thing that you're trying to understand. Can you, like, unpack that a little bit more for us?
- [03:30]
Psychiatrist Guest
Yes. Yeah. And, you know, I'm glad to hear that this is, this is an idea that has resonated with you quite a bit. And one of the things that I've been trying to do is to promote more pluralistic thinking in psychiatry and clinical psychology. And I think that there are two ways we can understand this idea of pluralistic kind of thinking. One is kind of takes inspiration from this general philosophical attitude, what we call scientific pluralism, that things can be explained even within the realm of science through a variety of different theoretical perspectives that rely on different assumptions, different background kind of ideas, and that each perspective might offer certain theoretical advantages or empirical advantages, but there isn't necessarily one single correct true perspective. But things are multifaceted, and that multifaceted nature is not captured by one theoretical kind of perspective. And this applies particularly in situations where phenomena do not possess a singular essence. When something possesses a hidden property that makes it what it is, kind of often identifying that essence is the single best way of characterizing that. I mean, I think a good example of that would be the pretty periodic table of elements in chemistry. It captures something really real and objective about the structure of elements. Their atomic, the way their atomic nuclei constituted has predictive, tremendous predictive power. And if different kind of chemists at different points in time start out from different theories of what chemistry is, if they do science well enough, they will all converge onto that, that kind of thing. But there are other phenomena that, that don't have that kind of singular, you know, hidden essence behind them, and they can be theorized and conceptualized differently. And this is particularly the case in, you know, in psychology and psychiatry, where these phenomena are really complicated. And so the one, one idea simply is that, you know, there is, there, you know, there isn't a hidden essence. You know, we have different ways of describing, explaining, understanding these ideas, and they come with advantages and disadv. But one is not a, you know, knockout winner. One doesn't displace the others. I think that that's the, you know, that's the fundamental sentiment behind explanatory pluralism. You know, as an implies in science, this is particularly true because we are dealing with the mind body split too. So there, there are theoretical perspectives that are more grounded. For example, in neuroscience, you know, trying to look at brain circuitry and, you know, receptor actions and, you know, other biological process. And then there are explanatory perspectives, more that utilize cognitive terms or utilize kind of psychological forces and unconscious psychological dynamics, et cetera. So we are dealing with that split kind of thing. A related idea that often comes up is that of levels of explanation, that things can be described at different levels of organization, different levels of coherence, and that different levels invoke different kinds of relevant concepts. So there is the physiological kind of level, neurophysiology, biology, there's kind of cognitive stuff, there's psychological stuff, and then there's kind of, there's social, interpersonal stuff which happens because of multiple brains interacting with each other. And so there's this idea that when things interact, new properties emerge that cannot be described adequately using concepts of the lower level. So sometimes people talk about pluralism in the context of psychiatry and psychology, just in the sense that these phenomena transcend multiple levels and can be described at multiple levels of organization.
- [07:45]
Dan Shipper
That's really interesting. You said a lot there that I want to unpack. So I think the first thing that made my brain go ding, ding, I want to talk about that is you're talking about the periodic table of elements and the idea that there are these sort of objective things that anyone who's doing science well is going to find. So, so periodic table of elements. You can, you can organize the elements at such in such a way that, you know, when you're going to transform an atom from one element to another, and basically anyone's going to find that. But there are other things that can only be described in, with different ways of looking at them in different, different perspectives. I'm curious about your Thought and this is maybe getting a little philosophical really quickly, but there are a lot of philosophers of science, for example, who might argue that even the periodic table of elements, that's one way to look at carbon, for example. And if we had a different set of requirements or a different set of intuitive ideas about how to interact with reality, we might find something else or structure things differently. Are you saying that or what are you saying? How far does the pluralism go?
- [08:59]
Psychiatrist Guest
Yeah, so I mean, even in things like physics and chemistry or it's a situation like periodic table, it doesn't necessarily mean that we have to rely on the standard periodic table description at all times and other things have no role, but rather so it's very context dependent. So there can be other theoretical ways of looking at how to think about elements or how to divide them or map them up. And there might be certain contexts, a certain situation in which you want to achieve a certain goal. And that goal is best achieved using a classification of elements that is not the one that comes from periodic table of elements. Right. I don't know enough about fundamental physics, theorize what they would be, but I can imagine, we can theoretically imagine situations where an alternative classification offers some advantage or benefit in a particular context. Right? And so we can imagine that. And in those contexts it would be useful, you know, it would make sense to utilize the alternative one. But you know, the priority of elements possesses such immense explanatory power that the majority of situations and the majority of contexts is going to be the useful one to use versus there can. There are other situations where there isn't a single perspective that dominates it in the same kind of way, you know, maybe like 60% of the time or something. But, but there, there are going to be significant cases, you know, in, in clinical work or scientific work where we are relying and using other perspectives.
- [10:42]
Dan Shipper
But aren't we sort of like sliding in something that's, that's unscience Y? Basically I agree with you, but I, I want to present the other side of it, which is by saying that aren't we sliding in something that's like fundamentally unscience y? Because science is really about finding those context free, ultimately generalizable rules or facts about reality. So say, well, it's contextual. Is that really, is that really the goal of science?
- [11:09]
Psychiatrist Guest
Philosophers of science have been thinking about, about, you know, about these things, you know, for, for, for decades and you know, if, if not centuries, and ultimately it all, it all comes down to, you know, some form of context or another. The, the reason Is that we don't have direct access to whatever, you know, mind independent reality there is. So what, whatever, you know, we don't have direct access to that, you know, we only have access to what we perceive or what we detect by means of other instrument, you know, or by, by means of experiment, you know, experiment with you. So we have to make inferences based on the, the data that we, that we are gathering. And any kind of data that we are gathering, it is taking place in a certain, you know, research context or in a certain scientific context. It is, it is guided by certain theoretical assumptions, you know, so all observations, so philosopher talk about all observations being theory laden that there is no theory free observation because any observation you're making, you're making that within a certain theoretical context, within a certain framework, within certain goals and they can change. Now some of the most fundamental theories that we have in science, they work really well in basically all contexts that we can apply them to until, until we start encountering situations in which they don't apply. You know, for example, like quantum, quantum physics, you know, we, there, there's so many aspects of it that we don't understand. Huge debate. But as far as empirical predictions of quantum physics are concerned, you know, you know, so far they have been unsurpassed, you know, like, you know, they have repeatedly proven true. True again and again again. So, you know, so that's a, that's a really, you know, powerful theory that, you know, gives us a lot of power. So there are, the fundamental theories tend to apply in all situations that we currently possess, but there would be new future situations in which they may not apply versus then there are theoretical models that are more circumscribed in scope that work well within certain contexts, but outside of that they don't work very well. And this may be because they approximate something about that larger theory. Consider Newtonian kind of laws of motion. They work very well within ordinary speeds, but once we cross certain thresholds, they're not as applicable. So one way of thinking about this is in this respect. And then there are other situations in which we're simply dealing with very different kinds of phenomena. So in physics we have the theory of general relativity at one end and we kind of have quantum physics on the other end. And at the moment physicists kind of juggle and they move back and forth between the two, depending on whether they're looking at galaxies or whether they're looking at. And right now we don't have a way of unifying them. Probably there may be a theoretical way of unifying them, but we don't know. And in a similar kind of way, we can think about theories that are examining the psyche from the perspective of. From the cognitive framework in terms of predictive processing of beliefs. Again, these are things we are inferring because we can observe them or we can kind of look at these things from a brain circuit perspective or neurotransmitter perspective. So some of it is methodological as well. Can we unify them at some point? Perhaps we would want to, and what that would look like we don't know, but we're not there yet.
- [15:10]
Dan Shipper
Yeah, I agree. And it seems like probably the history of science is like, if you go back to Newton is finding something that we think is super general and then realizing that it's actually contextual, we just didn't have a way to access the context in which it wouldn't hold.
- [15:26]
Psychiatrist Guest
Yes, yeah, exactly. Yeah, this is. Yeah, and there's this again. You know, we hear these term, the terms realism versus anti realism in philosophy of science, which refers to the idea that, that these unobservable entities that we are hypothesizing, what kind of status do they have? And realism, at least the kind of naive realist view is that the unobservable entities really exist versus kind of the anti realistic and anti realist kind of view would be that these are perhaps useful fictions that we don't know what kind of reality they possess. We just know that they are useful enough that they guide our empirical work and they make predictions that turn out to be true.
- [16:12]
Dan Shipper
Where do you fall on that continuum?
- [16:15]
Psychiatrist Guest
I take a more pragmatic kind of view towards these things which in some ways bridges the two, which it has kind of one leg in realism and one leg in anti realism.
- [16:26]
Dan Shipper
And you mean philosophically pragmatic, not like pragmatic. Well, like the school of philosophy. Pragmatism.
- [16:32]
Psychiatrist Guest
Yeah, yeah.
- [16:33]
Dan Shipper
Can you just tell us, for people who don't know what that is, like, generally what that is?
- [16:37]
Psychiatrist Guest
Yeah, it's basically, you know, pragmatism by itself is a. It's a pretty broad, you know, broad term, but generally what it does is that it shifts the focus away from certain metaphysical questions towards our ability to make use of concepts in the world. It asks questions like what it is that we can do with a certain theoretical concept, what are its uses, how are we applying it? And it judges what a theoretical perspective is offering or how we understand, you know, any particular phenomena based on how it is being utilized in practice.
- [17:19]
Dan Shipper
Yeah, so it's Sort of like, instead of asking, is this true? It's asking, like, is it useful for me to think of it as true and what can I do with that if I. If I do?
- [17:30]
Psychiatrist Guest
Yes. Yeah. And that's a simple. Yeah, that's a simplistic version of that.
- [17:34]
Dan Shipper
Yeah, there's a. There's a lot in there. But I want to get into, like, psychiatry, psychology. And I think, like, the reason that I'm so into this, other than, like, it's just fascinating, is just like, I have ocd. And it was a mess for me to even figure that out. And it took like 10 years. And I went to a bunch of different therapists, and eventually, like, I was just reading a bunch of stuff, and I was like, I think have ocd. And, you know, there's a lot of debate on. About whether that's even helpful. Right. Like, for some people, the label is bad. For me, it was quite freeing because I was like, well, you know, for ocd, typical talk therapy tends not to work that well until you're much. Until treatment is much better managed. And even once I had gotten into the correct treatment for it, so, like, exposure and response prevention, that it took me a while to get to someone who actually even practiced that well enough for it to work. And then it. It was a nightmare to get on medication. But once I did, it's been amazing. It like, literally changed my life. I'm like a completely different person. And during that whole time, I've also just always been interested in philosophy of science stuff. I was like, the way that this works is fundamentally broken and bad. And it's bad philosophy too. Ye. Talk to me about that.
- [19:03]
Psychiatrist Guest
Yeah, and actually, I think a lot of people will relate to what you're saying. And in my own clinical practice, one of the diagnosis that seems to have been missed by other clinicians a lot is obsessive compulsive disorder, ocd, where people will have histories of being treated for depression, anxiety. And yet, can I do a comprehensive interview? And I'm like, I think you may have OCD and suddenly are, you know, clicking into. Into the place for them. And I think the reason is that for, you know, for one reason, one, I think, you know, it's not as popular of a diagnosis in. In the cultural imagination as things like depression, anxiety, or adhd. So oftentimes when, you know, here's the thing, you know, in. In. In the psychological realm, you know, even when we are describing, characterizing, talking about a phenomena, it is very sensitive regarding the language we use to Even talk about it. So what happens is that if someone, for example, if a person has no idea of what OCD even is, they have never heard of it. They're not familiar with obsessions in the technical sense. They might notice that they're getting really anxious. They might notice that they're having a lot of repetitive intrusive thoughts that kind of make them uncomfortable. One person might just think that, oh, I'm just, I'm just a really nervous person or I'm just, I just have really bad anxiety. And this anxiety and my anxiety is just causing me kind of like, you know, these, these kind of, you know, really nervously anxious thoughts. Another person might think that I'm, you know, I have some kind of paranoia. I'm scam. Getting these really weird, scary thoughts that make me kind of, you know, uncomfortable, make sufficient, you know, I may have paranoia or something like that. Another person just might, you know, interpret at them like, you know, I just, I get these racing thoughts, you know, maybe, maybe I have some kind of bipolar. I'm seeing, you know, I'm getting these weird racing there. So depending on, you know, how someone, the language someone might use to describe them, it might shift them in one or a different diagnostic, you know, diagnostic direction. So it takes familiarity with a number of different, you know, you know, diagnostic, you know, categories to begin to recognize, you know, which direction going. So one oftentimes patients themselves not tell the clinician that they, oh, I'm having obsessions because they don't use that language. They'll say, I'm anxious, you know, I'm having nervous thoughts, et cetera. And it would take, you know, a somewhat astute clinician to kind of do a good interview and pick up on the fact that someone is having intrusive thoughts and then do, you know, ask further questions. So I think that's one big difficulty. And you know, and. But when you do have the right match, when you can match someone's presentation with this idea of ocd, you know, you can guide them in the right, towards the right treatments that, that have been studied and that are known to work better, you know, exposure and remodel, prevention therapy, other forms of psychotherapies and then medications as well. Oftentimes we are using the same kind of, you know, serotonin based, you know, medications and other ones, but the dosages required are often different. Different augmentation strategies are often different. So, you know, and so you're probably familiar. There's a, there's a lot of, you know, like, you know, anti psychiatry discussions in the mental health realm. And some people just say that, oh, we abandon all diagnosis. And I think it just. How can we talk accurately about these things without diagnostic language? And so I feel strongly that we need more sophisticated ways of talking about these phenomena. Now, the challenge here that happens is that we can reify that category. We can think, oh, obsessive compulsive disorder. It means that it's a thing in the brain that explains this category. There's some kind of a singular brain dysfunction that explains all of these symptoms. And so that's where I think people go wrong if they start thinking of categories in that manner. And so ocd, as far as we know, it's not one thing. In fact, even from a symptom standpoint, it has fuzzy boundaries. So its boundaries overlap with other anxiety disorders, overlap with depression, with other kind of things. And, and biologically speaking, there is no one single thing that makes ocd ocd, but rather it's a variety of different variants. So we have to respect that heterogeneity, but it offers us a certain descriptive potential. We can talk about it, we can identify it, we can use it for guiding treatment. And so this would be a good example of something that's very helpful pragmatically. It's an entity that exists from a pragmatic standpoint of view, but not from an essentialist, holistic point of view.
- [24:03]
Dan Shipper
Well, and, but the, the interesting implication of the pragmatic idea is that some, for some people, it's not helpful. And so part of, part of being a good clinician is knowing when it is and is not helpful. And I mean, you're going to mess it up sometimes. But, like, that's a really interesting thing, right? Because I think the entire this, you've, you've just led me into my AI, into, into AI land. So thank you for that. I think, like, part of the, there's, there's a lot of pressure to be able to reduce these things down into a checklist, right? That even if you don't have any clinical experience, you can just know, I checked it, I checked off the box, he has OCD or whatever. And I think what one of the things you're saying is like, yeah, you can do the checklist, but really you need to have like a good clinician has a little bit of like a smell for it, and they're just like, I, I just, I know. And I can kind of explain in this case why, but like, I can't fully. Fully, you know, but I think there's one there's sort of scientific academic pressure to have a really clear cut theory that you can just write down exact, this is exactly what it is. And then there's also a lot of like cost considerations and just scale considerations of if we want to, if we want to make treatment available to a lot of people, then we need to be able to reduce it down into, here's the, here are the rules for how to, how to diagnose it and how to treat it. And we make basically manualized therapy which has some, works to some degree, but is not the same thing as a really, really skilled clinician. So to get to my AI point, these questions are exactly the kinds of questions that early machine learning researchers faced, which is like, if you want to recognize a letter, you want to do character recognition. Do we, like, how do we define what an A is? Do we like create a bunch of rules for it? And if so like, there's a lot of different situations in which something might meet, might be an A, but it like doesn't look at all like an A. It's very contextual. So some people went down the route of like, let's define all the rules, which did not work. But what, what ended up working is deep learning where basically what you can do is feed in a lot of examples into a neural network network. And the neural network creates distributed representations of whatever the examples are such that no individual neuron knows like, hey, this is like A, this is an A. But altogether basically, when you feed something into a neural network, what the network is doing more or less is testing out a bunch of different hypotheses for what it, what the letter could be and whether it could be A all at once in a non rule like way. And then outputting. Yeah, I think this is, this is basically an A without any ner, any individual neuron knowing the rules. But at a high level you may be able to say like, yeah, generally if the A has a little thing at the end, then it's going to say it's an A. And so that's the basic kind of parallel I see to machine learning stuff. I'll stop there. But there's more stuff going on there. Does that, is that something that you've run into before or thought about? Hey there, Dan here. I wanted to take a one minute break from the episode to tell you about our latest sponsor. All right, let's play a game. What powerhouse productivity tool is also free for individuals? Nope, not that one. Try again. You may not expect this, but it's Microsoft Teams. Yep. The same teams that big enterprises swear by also has a free plan for individuals. Whether you're jamming on a side project or bootstrapping a startup or building a community, Teams has all of the features that other platforms Nickel and Dyneouf are using. You can get unlimited chat, 60 minute video meetings, file sharing and collaborative workspaces, all for free. And the real magic is that everything is integrated into one seamless collaborative workspace. That means there's no more hopping between different applications for messages, meetings and file sharing. Teams puts it all at your fingertips to save you time and money. So ditch the app overload and the subscription fatigue and use teams to experience effortless collaboration. Today, are you ready to streamline your workflow? Head to aka ms.every to use teams for free. Your productivity will thank you and so will your wallet. And now back to the episode.
- [28:36]
Psychiatrist Guest
Yeah, yeah, I mean, I think, I think there's this, you know, there's this recognition that the, you know, what, what we call the space of psychopathology, so this kind of, you know, this abstract conceptual space of the way symptoms are distributed in, you know, in the mental health realm, what are, you know, what are the ways in which this, you know, the space of psychopathology can be car. The original kind of idea, the hope for a lot of people was that we can find single big causes or we can find essences and that'll provide us a way to carving this space. We can carve the nature at its joint kind of thing. And then now we recognize that, oh, it's just this big fluid fuzzy mess of interacting causes and mechanisms and you can just really map it in many different ways. So we have used different strategies. Basically what the DSM, ICD, et cetera have been doing over the past 50, 60 years is that they have been relying on clinician observation. So they started with clinician observation and kind of refined it through operationalization, et cetera. There is kind of an alternative effort called high top or hierarchical taxonomy of psychopathology and that uses more statistical techniques like factor analysis and other kind of PCA analysis is to see patterns of symptom covariation. If you change one symptom, what symptom changes with it? And it has come up with an alternative dimensional mapping of these symptoms. But the space of how to classify them is very flexible and there are more innovative ways you can do that. And this is where probably AI and kind of machine learning, deep learning come in. It can help us identify new patterns and new ways of talking about the space of psychopathology. That we might not have thought of ourselves. And I think people are looking into it and kind of approaching this to my understanding. I don't think it has produced any actionable results yet. I don't think someone has come up with a new mapping that seems very useful. But I think in theory that's possible. And, and it would be, I think it would be exciting, you know, to see.
- [31:05]
Dan Shipper
Yeah, well, I think the interesting thing to me is, so one is talking about like observing clinicians and then reducing how they operate to diagnostic rules. That's exactly how machine learning researchers started in AI. And it's exactly what didn't work. It does work now once you have a deep learning foundation. It does work now. But at the very beginning, when you're just trying to make it like a checklist of rules instead of like a sort of fuzzy pattern matcher, it does not work very well. But I think the interesting thing to me is about, you know, like factor analysis or high tops or whatever is you're still trying to say, like, why? Okay, so like if there's like, you know, OCD is caused by some constellation of different factors, but like, we can at least say what the factors are. And I think one of the interesting things to me about, or one of the other interesting parallels to me about AI stuff with, with psychology and psychiatry is I think the development, for example of OCD or whether or not you're going to have OCD or whether or not it's going to work. That question that, that predictive question is a little bit like the way, I don't know how, how familiar you are with the way that language models work, but it's a little bit like the way that language models work to predict the next word. So the way that a language model predicts the next word is it looks at the context of all the words that came before it. And it's been trained that all the words that came before it each change the probability of the next word in the sequence and each affect each other. So all the words get a chance to talk to each other and then all the words get a chance to basically figure out what comes next next. And there's no like one set of rules and there's. And it's, and you can't really. I mean, it does learn the rules of grammar overall. You can look at it and be like, it knows grammar. Yeah, but, but if you're trying to understand, for example, when it might use the word ocd, it's, that's very hard because it's so multifaceted. It's like millions and millions and millions and millions of parameters all interacting with each other to like say OCD comes next next. And I think that actual OCD probably works a lot like that or can be, can, can be worked with a lot like that. Let's just say that. And the reason that we haven't done that is because in science we prioritize explanatory power over predictive power a lot. Because in order to get your paper accepted by a journal, you have to like, have a theory. And deep learning doesn't have theories. It just says, says we know it works. And so my, the, the my like, I don't know what it like my stump speech is basically like, we should just like throw out explanations in, in a lot of these cases and build predictive models with, with large, large, large data sets. Like instead of throwing 16 undergrads into a brain scanner and being like, I think of 15 minutes of meditation, like slightly affects OCD on average in these 15 undergrads, we should just like gather tons and tons and tons of, of, of you know, both contextual data and biological data and like whatever, like chat logs, whatever, and just throw it into a, into a deep learning model to predict when you're going to get OCD or what interventions might work or whatever, even if we don't have a theory. And I think science is generally blind to that idea.
- [34:39]
Psychiatrist Guest
That's a great, you know, question and you know, and suggestion and you know, to my understanding is that researchers have been trying that kind of thing in psychiatry and psychopathology for a while. And so, for example, we have had kind of people look at risk of suicide, for example, in large data sets and trying to predict what kind of factors lead to suicide and to what kind of accuracy. We can predict the risk of suicide over a certain timeframe. And there were publications that reported really, really high accuracy or really high predictive power, you know, 90 or more in some samples. What has happened so far is that that accuracy hasn't translated very well outside the samples on which the program was trained. So a program might do very well in a large data set, but then you take a wildly different data set and then sort of like the kind of predictive power goes down. Now that may change in the future as the training gets better, but that's what has happened so far is that you get exceptionally good at the, you know, at the algorithm, predict what happens in one, one group, but then it fails outside. And so they have tried that with, you know, things like disease progression. You know, if you take people at risk of psychosis, can you predict, you know, which one would transition to psychosis? They have, they have used development of bipolar disorder. If you take people with depression, can you predict, you know, who will go on to have many episodes? So I think that's where I have seen most kind of, you know, like machine, machine learning, machine learning, deep learning kind of stuff being used. Other stuff this is kind of happening is in kind of understanding patterns of brain circuit activation or patterns of kind of neuroimaging finding. So there were, for example, kind of, you know, a series of papers kind of last year that looked at, you know, brain circuit activation patterns in depression and anxiety and identified various kind of subtypes that correlated with certain clinical features, et cetera. So that's another kind of situation where people use. But I think we're still waiting to, you know, for this kind of methodology to deliver really, kind of, you know, results that are really actionable or that work well across many, many contexts.
- [37:06]
Dan Shipper
Yeah, I think my, my response to that, I think that's a really important, like, let's say, concern. My response to that is the same thing was true of, of earlier versions of text prediction or image generation. And the thing that worked was more data and more parameters and more compute. And so I'm not familiar with all of the studies that have done this, but my guess would be that one, what counts as a large data set in that world is actually comparatively not large, large to like, like machine learning researchers who are like dealing with like terabytes of data. And that's even that is like probably pretty small. So one is, is, is compile like the, the task becomes how do we get an actually gigantic data set? And I think a lot of the big tech companies have these data sets if, if they wanted to donate them to science, I think they should. And the second thing is, I think a lot of these studies use more simplistic, underpowered statistical models that are more understandable but end up overfitting and aren't able to represent the like, complex nonlinear interactions that end up actually predicting things like suicide or depression or whatever.
- [38:38]
Psychiatrist Guest
Yeah, you know that. Yeah, you're probably right. I think, you know, I think, you know, we're still early in this space and there's so much innovation happening in the neural learning space. And I think it probably takes some time before these kind of methods start trickling into the clinical research space. We probably haven't optimized what can be achieved using that. So I agree with that. So I think there's reason to be optimistic and there's reason not to give up on that, but rather continue exploring know these methods. One thing I'll say is that I think, you know, there are, I think there are things we can practically, you know, start doing even, you know, sort of like, you know, start exploring, you know, even now. So for example, one area where AI is, you know, you know, kind of like getting some practical application in medicine is in note writing where there are now softwares that have been developed that kind of, you know, like, you know, listen in on the, on the clinical encounter or the interview and generate and kind of, you know, a template of a medical note based on what the patient and the doctor discussed. Right. So that's because it's a relatively, the clinical encounter is relatively formulaic and follows a certain template. And I think it could have a lot of utility in clinical interview as well. So as you probably recognize from your clinical experience, the average clinician is not super skilled at picking up the nuances of the presentation. Right. But, you know, you could easily develop kind of an AI version of a clinical interview that, you know, exceeds the average clinician. Someone like you, for example, let's say, could have a clinical interview that is facilitated by, you know, a large language model. And you know, the language is going to have the resources and the model is going to have the time and the resources to go into the nitty gritty of the symptoms and, and inquire about, you know, clarifying, you know, what you're experienc than, you know, a rushed clinician who just has 10 minutes to talk to you in an appointment. So, you know, I think, you know, that's where I think, you know, we don't even have to wait for, you know, like, you know, new scientific discoveries to talk about clinical application. If someone develops a good, you know, clinical interview AI thing, you know, it would be usable straight away.
- [41:12]
Dan Shipper
Yeah, I agree. I mean, A, there are definitely a lot of people who are working on stuff like that, but B, I think like ChatGPT is already sort of that, to be honest.
- [41:20]
Psychiatrist Guest
Honest.
- [41:22]
Dan Shipper
And it is kind of this interesting thing where, I mean, I go to therapy every week and I'm a big fan of therapy and I have, I have a real human therapist and a real human psychiatrist. So, you know, but. And I think the dream is obviously for everyone to be able to have that and everyone to be able to have a skilled clinician. But the reality of the, of how expensive is it is to do that means that if you want to do that, that the clinician has to probably be under trained, understaffed, have 10 minutes to talk to you, have a, have a checkl and is going to get things wrong more. And I think while ChatGPT is not being billed as, you know, a therapist, I would bet you if you look at the way people are using it, a lot of them are using it for things that are sort of therapy like and that it's, that's actually probably a good thing because it like basically democratizing access to like the most basic level of mental and emotional support means that some people won't develop problems that might have otherwise developed them. But it also becomes this sort of funnel so that by the time that, by the time you get to a therapist, maybe it's told you hey, like here are some things you should talk to your therapist about or whatever that help the human clinician make progress more quickly.
- [42:35]
Psychiatrist Guest
Yeah, I agree. I think, I think that there are many of my colleagues, especially in the psychotherapy world are kind of wary of these developments and they're, they're, they're looking at a bit suspicion. I am more optimistic in my orientation about these things and I generally see this as a positive development. I think it's good and I think there, there are problem that are more amenable to this kind of self directed therapy through ChatGPT, et cetera. And then there are other more entrenched problems that might require relationship with a human. But to the extent that there are problems that can be addressed in a somewhat satisfactory manner by AI, then why not, why not utilize that? And I think a lot of people are drawn to, on to the promise of that and they like having more control and they like having to, you know, I think a lot of times clinicians can, you know, you know, they have all of the, you know, quirks of humanity. They can be arrogant, they can be dismissive, they can be rushed. Right. Versus chatgpt is patient. It's all always there. It's non judgmental.
- [43:46]
Dan Shipper
I think a lot of people don't want to face that.
- [43:48]
Psychiatrist Guest
Right? No, it's not just not judgmental. Right. So you can use it in ways that you can manipulated in ways that you cannot manipulate a human clinician.
- [43:59]
Dan Shipper
I'm curious if you've used it at all for this kind of thing because like for me, like, I don't know, I've been having these like stress dreams the last couple nights and so I just record a voice note of me talking about my dream and then, and throw it into Chat gbg and I have a whole like log of a bunch of dreams in the last few nights. And it's just like talking to me about different like little patterns in my psyche that I think are, are for real. Not in a necessarily like they're predicting the future kind of way, just like in here's, here's like the emotional landscape that you're in right now that might be useful for you to like have in mind as you go through your day, which it definitely has been. Or like I often like, I record all my meetings and like throw them into, throw them into ChatGPT and I'm like, I'm having a problem with this person or whatever. Like can you walk me through like what to do about it versus maybe I would go to therapy, but like that's once a week and it relies on me having to be like, well he said this and she said this and then I said and it's just, it's just different, you know.
- [44:58]
Psychiatrist Guest
Yeah, I, you know, I haven't used kind of ChatGPT or Claude or anything in that, you know, from that angle. I have mostly used it for kind of brainstorming ideas or doing kind of, you know, superficial research on some topic that I'm not very familiar with or you know, used it for editing. For example, if I wanted to edit a piece of writing and wanted input on how it could be worded differently or improved. So I've used it for those purposes. I have not you personally use it for psychological exploration, although this is something I should probably try as well. Well, one thing I guess since we are talking about this stuff, so if we are talking about AI applications in the classification realm, for example. Right. So to the extent that the process is going to be guided, you know, by human decision making, right. It's going to be about teaching the algorithms to kind of detect, you know, the diagnostic categories as clinicians do. But if we wanted to think about new forms of classification, you know, how might an AI kind of, you know, training algorithm approach that kind of thing? If we wanted to, if we wanted it to detect class or you know, come up with classification proposals that are different from the ones we offer already have.
- [46:20]
Dan Shipper
I think it's a great question. I think the, the, the thing to think about is one is they work on examples. So if you, if you give it a bunch of examples and ask it to, you can, you can give it a pre existing categorization scheme and say this example is like this or this example is like this that, but you can also just Say for example, predict whether or not this person will be helped by search relien.
- [46:56]
Psychiatrist Guest
Okay.
- [46:58]
Dan Shipper
And it will do its own internal classification of you know, searchline or no searchline that goes across diagnostic categories and that we probably don't. It's, it's super high dimensional. Like we like probably very, very hard to understand. We probably could get some understanding of how it does it, but it's doing its own internal sorting. So, so I think the best way to get it to come up with its own categorization is to set it a task that it would require it to categorize it in an, in a novel way and then it will find a categorization that works for that task. Sometimes like in, like by default that will be an internal one. But you can also get it to write out its categorization scheme or write out the reasons for why it does things in the way it does it. It takes a little bit more work to do that, but it's. I think it's possible.
- [47:56]
Psychiatrist Guest
Yeah. Okay. Yeah, that's helpful. And one of the, you know, interesting things is that some, sometimes cultures do that. You know, you know, we have different, you know, there's cultural variations and sometimes different cultures talk about a psychological phenomena in ways that are kind of radically different. So for, you know, so we get these kind of, you know, culture bound syndro, for example in the mental health realm. But a while back I came across this cultural description of this thing called egoria which essentially translates into self leakage. It was kind of like this idea that a person's boundary essentially has been violated in some way and a person is. Elements of their psyche are leaking out in certain ways. And this is a way of thinking that transcends for example traditional psychiatric classifications because you know, it can cross into things like, you know, it could be, you know, delusional experiences, could be hallucinatory, but could be kind of, you know, other form even, even some forms of, you know, obsessive compulsive, you know, phenomena sometimes cross cultural comparisons of mental health descriptions, you know, can, can really be surprising. And there's like, oh wow, this is a really, you know, this is a really different take on, on, on a familiar psychiatric concept.
- [49:22]
Dan Shipper
Yeah, I think that makes sense. The way that I think about that generally is for things like psychiatric diagnosis, diagnosis or classification, it's emergent. So it's super, we're talking about something that's super high dimensional. So any classification system is only going to, and I think you agree with this is only going to give you like A slice or slice it in a particular way, but it's not going to be the slice. And it's a lot like musical genres, you know, like what, what really counts as rock. That's might seem obvious, but if you really think about it, there's a lot of stuff there there. It's like they're, they're on the border, you know, like, what are, I don't know, what are the Beastie Boys like? There's some rock influences there, but they're kind of rap. They're like, there's. That kind of thing is. I think is, is the same question. And you can come up with your own categorization schemes, but the question is for what purpose? Like in order to reduce it down into, into a dimension that we can understand, you have to, you have to understand the purpose for which you're, you're doing the ranking. Otherwise the way that, the way that language models do this is they just create, they just map things into a high dimensional space so they put like things near each other and unlike things that are unlike things further away but in, in across many, many, many, many, many dimensions. And they don't have a, like a huge dictionary of each thing that each, you know, or they don't have a, they don't have a full map of each thing. What they do is they, you give it a thing and then it generates its location in that space. And once you have its location then you can kind of say like, well, it's near this or it's not near this. And that's how you get away with not having a like universal understandable categorization scheme. You just map it in, in a space that allows for thousands and thousands of different dimensions.
- [51:10]
Psychiatrist Guest
Yeah, it's, it's, it's all really fascinating.
- [51:12]
Dan Shipper
Yeah. Anyway, this is, this is fantastic. If people are interested in. Read more about your reading more of your work, where can they find you?
- [51:23]
Psychiatrist Guest
The place where I'm most active these days is my substack newsletter, Psychiatry at the Margins. So that's where I'm kind of talking about various kinds of controversies in the field and various debates and other developments that are taking place. I'm also present on X Twitter on Blue sky. And as you mentioned in the beginning, recently there was a volume published by Oxford University Press called Conversations in Critical Psychiatry, which is an edit collection of interviews that I had done for Psychiatric Times. So I encourage people to check that out too if you're interested in various debates in the field. So yeah, so I think substack and other social media platforms. Probably the best.
- [52:06]
Dan Shipper
Awesome. Thank you so much. Keep doing what you're doing. It was really great to talk to you.
- [52:10]
Psychiatrist Guest
Thanks for having me.
- [52:19]
Podcast Host/Announcer
Oh my gosh folks, you absolutely positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure, unadulterated knowledge bombs about chat GPT. Every episode is a roller coaster of emotions, insights and laughter that will leave you on the edge of your seat craving for more. It's not just a show, it's a journey into the future with Dan Shipper as the captain of the spaceship. So do yourself a favor. Hit like smash, subscribe and strap in for the ride of your life. And now, without any further ado, let me just say, Dan, I'm absolutely, hopelessly in love with you.