AI Alignment, Mental Health, and Belief Systems

Whether it's depression, whether it's schizophrenia or psychosis or otherwise, you name it, there's always an autonomic dysfunction in it. From a psychiatric point of view, that the belief confirmation would mean that the patient is on the expected trajectory, what we call psychiatric baseline. Because you can have like depression and have automatic negative thoughts and core beliefs of I'm worthless. If the loosening of association starts, rumination starts, that could mean that the psychosis is going to happen because it's beginning or the cognitive distortions are shaped in an odd belief type of way.

It's really kind of this idea that AI can know us and we can be aligned through it. Understanding our belief system, how we'll answer questions, you know, that might in our system, you might automatically be able to create like three recommendations of follow up questions to ask to, you know, clarify the belief. The heuristic of what's the next best question to ask is one thing that we're thinking about a lot and the other thing is how well can we predict what the patient's going to say next.

You will have a conceptual map like a roadmap, like a gps, where the AI therapist or the actual therapist can just maneuver its way through those onion layers and really target, or pretend to target one thing, but they're really targeting this so that the therapeutic alignment is high fidelity as well.

Today I have on with me Jonathan, my co host and co founder for Epistemic Me. And we have an awesome guest, Bishoy. In this podcast we bring on researchers, scientists, founders, entrepreneurs, leaders, builders of the future, people doing big things with big stories to tell. And Bishoy has a really interesting background. We were just chatting about it. Jonathan and Bishoy know each other. Bishoy works in psychiatry. He's done a lot of research. And so I'll go ahead and let him introduce himself and his breadth and depth of portfolio of work. And then there's this intersection that we're going to explore today around epistemic me. What we're building models of self belief systems and then Bishoy's world with psychiatry. Bishoy, can you introduce yourself?

Yeah. Thank you, Robert. So I'm Bishoy. I'm an assistant professor in psychiatry, board certified in both child and adolescent and adult psychiatry. So I got to see the whole evolution of the psychopathology. I worked with Jonathan before, collaborated on previous projects and I'm excited to learn more about epistemic pain. Previously I done a lot of research in behavioral and cardiovascular medicine.

The connection that research you mentioned Bishoy around heart rate variability tied to mental disorders. I forget if you mentioned which one. That was really interesting. Could you maybe expound on that a little bit for the audience? Just because me and Jonathan are like health nuts. And that's also interesting, I think, for folks listening.

So just to bring you a little bit to my world of psychiatry. When we're writing a note, for example, and that's in all medicine, we call it SOAP notes, Subjective, objective assessment and plan. So subjective is the subjective experience of the patient. Oh, I have a chest pain. Objective is, you know, the examination. Patient has tenderness in the abdomen when palpated. And that's the objective part. Assessment is synthesizing the subjective. And the objective and plan is what are you going to do? So the objective part in other parts of medicine could be a CT scan, could be, you know, an ekg. In psychiatry, the objective part is very limited. So that sparked the idea of trying to concretize a lot of, and add a little bit of the objective meters. And one of them was the link between autonomic system and our psychological states. Because autonomic system is so intertwined with a lot of mental illness, whether it's depression, whether it's schizophrenia or psychosis or otherwise, you name it, there's always an autonomic dysfunction in it. And that autonomic heart rate variability have a lot of indices that are measurable and are trackable. And there is a lot of interventions that also change them to the positive, including paced breathing and other things. And it can be measured, as, you know, by wearables. So in that sense, we started to think about, okay, how can we track the correlates of certain states and therefore create an ability to create predictive patterns, for example, of a depressive episode or predictive patterns of psychosis or otherwise. So that was the hope grant. This was like many, many years ago, before the advent of AI and all the pattern, the pattern recognition that is, can be very helpful in figuring out this data. Like I remember, you know, assess snippet by snippet, get the rmssd, like all the indices by myself. And in the analysis of those HRV spectrums and, and looking visually at the, at the patterns, and there were some like, although limited methods of prediction. But, you know, in a nutshell, I think the more that you're bringing out the subjectivity and putting it into an objective world like that, okay, and making that O much more bigger than the S. So the S is the experience and O is that the objective part, Our objective in psychiatry is the mental status examination, which is a status like a snapshot of the mental, emotional and physical state that the patient is in at this moment. And I think this is a good segue to tie it to epistemic me because I am. One of my curiosity was that although that mental status examination is a snapshot at this moment, it's not updating. That's number one. Number two, it doesn't have the whole context of the belief systems and the core beliefs. Now we go into that by therapy and challenge it and do the cognitive restructuring, cognitive reframing. But in that snapshot we don't have access to this whole context by which this mental status examination is arising. Definitely you try to do history taking, but it's not as powerful as having that map of the self model. So in that sense that was my curiosity about epistemic me. And I would like to learn from both of you today a little bit more about that and then we can, you know, tie it back in.

Thanks so much for joy. So it that's actually crazy to visualize you manually getting all that info before the advancements of electronic and digital health and different systems to help out and now with AI on the precipice of well giving us a better sense making tool than we've ever had. There's a stat that I always love from super agency. I referenced this a lot. But 2.3 billion ebooks of data are created every 10 seconds. And so like that's in the world. And so no human. There are some humans don't even read a book a year. Right. So just mapping that to the analogy of like a doctor such as yourself trying to make sense of somebody else's brain. It's messy, right? Because the human experience is quite messy. It's interesting that you talked about how that SOAP is static. You would like the objective to be a bigger size than the subjective. And you'd also like this to be dynamic and tied to the belief systems. And so one question before maybe I hand it off to Jonathan and maybe some of his views on how we might be able to help that use case is I kind of perceive that your interest comes from hey, I could maybe help. This could maybe help me understand patients better. Like having a dynamic map of their inner psyche, not just that static one time SOAP or and update itself. Is that where your interest comes from?

Yeah, definitely. I'll use a concrete example just to go through it. For example, you know, if we're looking at an ontological hierarchy or the belief system all the way to it's much more abstract. So you. I want to think about it like this is the most abstract and this is the most concrete, you know, specific thought. The most abstract, let me use an example is I'm being persecuted. And then after that is the layer where, well, the government is spying on me. And then after that in a more concrete level is oh, that's why there is that fire alarm above. And that's why my neighbor looks at me funny every time I go out. It's part of it. And so it starts to be pulling down. But there is a origin, core belief origin of, of all these intermediate and super superficial quote unquote delusion all the way to the behavior it starts doing, right? Like now I'm going to self isolate. I'm not going to eat because somebody's poisoning that food. And I'm going to, you know, not quit work because they are all part of it. So when we are doing the mental status examination, we typically see this part in this visual. We see the very dexterity, we see the behavior, the paranoid delusional, persecutory delusions and the behavior that it led to. And we see the superficial delusion which is, oh, the government is spying on me, all of that. But we don't go sink deeply all the way to the ontological origin unless we're in therapy. You know, where you go into the core belief and you start targeting the thought or the cognitive distortion on multiple levels. So that's number one. The second part is that although we are tracking by. By visits, right? Like weekly therapy visits and monthly psychiatric visits, we're not seeing how is that updating as quickly as we can in a, in a different way. We're not seeing the slices, we're not seeing the rigidity and the malleability of the false fixed belief. If we're talking about delusion, which is in its definition, it's a false fixed belief. So in a way there is that disconnect that creates a frame lapse between the provider and the patient. And therefore, you know, having a higher update frame could be something very, very helpful in understanding the whole context and knowing how I can personalize the treatment to this very person based on their values, based on their belief systems and based on what I'm looking at the behavior and the micro behaviors they're after.

Very cool. I liked what you said around having more frequent frames. That was a really great explanation. Before I hand it off to Jonathan, one of the things that popped in my head was you'd like a monitor with A really high refresh rate on your patient's brain. Right. Like the cognitive distortions that might happen, the delusions that might happen, and then the core beliefs underlying them. Jonathan, do you want to talk a little bit about maybe later on some of your thoughts as to this conversation and add in how maybe epistemic me might help and how you thought about the design of epistemic me that might meet this use case?

Yeah, it's really so interesting. I love the language and the description of the patient relationship. And so the SOAP process that you described, the subjective objective assessment and plan, it's really interesting because in the functional health AI coaches that we're working on right now building these modules, and they basically kind of work in that cycle. And I think with the functional health, it actually starts with maybe the objective, because you upload some data, it might be your sleep and exercise data, or might be some biomarkers or something like that. But then we take you through a cycle where then the AI coach is asking questions of the user to then understand their habits and their beliefs. And of course, habits being very closely tied to beliefs that are put into practice. Right. So that objective and subjective kind of automatically come out. And then we want to get to a point where there's an assessment of where you're at and recommendations. And I think that's where epistemic meet is in the AI coach as a recommendation system that really meets the existing behaviors and beliefs of the user. And then so then the user gets to select their plan. And so you have these different modules that you can kind of stack together from a functional health perspective to be AI health coach. And it sounds just like the inquiry pattern sounds very familiar. It's cool to hear the language that is actually practiced from this. And so I think if you think about that aspect of subjective inquiry and really understanding the belief system of the patient, one thing we've done in the epistemology of epistemic me especially, I think we talked about predictive processing like Carl Friston's work in the past, and the free energy principle. So in that epistemology that we have available in the repo, we have different layers of beliefs. So you have a belief statement, just something that you say. You can have an evidence belief which you call a falsifiable belief. And then we have a predictive belief, a belief that's evidenced and also sort of predictive of the reality of, of the user or patient. And so one thing you could think about as like you're inquiring with a Patient is like, what experiences do they have that grounds this belief that they have, or is it? And like you talked about that spectrum from the abstract to the concrete that they might experience. Like we could model those. Or the idea is to model those, you know, in the ontology and the epistemology that you can see, like maybe the percentage of beliefs that are being expressed that are evidenced or, or like you say in the delusional case, this is just like a fixed dogma of the world that is not serving and we can represent the beliefs that way. And so I thought that just having an epistemology where you can explicitly express beliefs in this way is useful and can help us know each other better. And that's how it's kind of like how I reflect on what we're doing and what it looks like through the lens of what you describe.

And Jonathan, I'm curiosity and Roberts, of course, like, what is the process of extracting those belief systems then?

Yeah, yeah. So in some sense it's just a survey and asking incremental questions of the user. And so in our system we try to train it to be aware of what, like the next best question to ask is in the engagement. So that like, let's say that you get a provocative belief statement from a patient, you know, that might, in our system you might automatically be able to create like three recommendations of follow up questions to ask to like, you know, clarify the belief. This is a disambiguation process, right. So we're incrementally collecting evidence to disambiguate like how the user is going to see this or how the patient's going to see this in their real life. And so I think like the heuristic of what's the next best question to ask is one thing that we're thinking about a lot and the other thing is how well can we predict what the patient's going to say next. Do we really understand the beliefs to the point where we can ask sort of like a test question, they answer in a way that we understand. Okay, like now, now we have a coherent disambiguated sort of understanding of how their beliefs are expressed in this context. And I think that's how you can picture it.

Yeah, I was going to say maybe it might be useful to demo my little mind mate demo and just kind of show what I envision a little bit of a flow would look like for that disambiguation process and then that might get the juices flowing in the brain for how we could meet this Use case is really interesting. Would that be. Do you think that might be useful? Just like, real quick, one minute, show pishoy what we have. Like, I have it up. It'd be pretty easy to go through.

That would be helpful. Yeah.

Perfect. So I built this little therapy exploration companion in lovable, and it's not yet integrated with epistemic me, but just to give you an idea of that disambiguation process, and then you can kind of picture in your head the underlying belief systems and evidencing that we do with the SDK that we're building with epistemic me. So I just made this real quick as, like, there's lots of different therapy options out there. Skipping ahead, there's different therapy approaches like cbt, psychodynamic therapy, humanistic therapy, dialectical behavior therapy, et cetera, gestalt therapy. And one of the friction points in the mental health journey for a patient today is one, first of all, finding the right provider is kind of like dating. Second, it's because it's like dating. You have to find the right person. And maybe they only specialize in one, maybe two. And so you don't really ever get the opportunity to try all the different ones and see what might work for you. And so I built this to sort of just mirror that. And so just going into this, like, I just used the concept of the feelings wheel of, like, you know, how are you feeling? Let's just say, like, I'm feeling sad. Let's say I'm feeling lonely. And then let's say I want to talk about my relationships. Let's say I want to talk about family. And then let's see if this still works. Yeah, okay, cool. I thought, because this is a demo, usually demos break on site, but it's working. So this is general conversation. So I had just kind of plugged in. I'm feeling lonely. I'd like to talk about family. And then here Mindmate says some emotional validation. And then a disambiguating question, which is, would you like to share more about your family situation, the feelings that come up? I'm here to listen and support. Let's say I change this to acceptance and commitment therapy. And so now Mindmate puts on this role. The AI underlying this is given a prompt to act with this role. And so in this demo, this hasn't gone through rigorous evals and stuff like that to make sure it's guardrailed to that particular therapy. But let's say I go to that share process again. Let's say I'm sad, lonely, or rigorous evals and stuff like that to make sure it's guardrailed to that particular therapy. But let's say I go through that share process again. Let's say I'm sad, lonely. And then it should ask me first. I want to take a moment to just notice that feeling of loneliness rather than just trying to push it away. Can you acknowledge its presence? It's like a cloud passing through the sky. Sometimes it's dark and heavy, but also pass. What thoughts come to mind when you feel lonely? What physical sensations do you notice in the body? So let's, let's just type in something. Some thoughts are that I don't feel like I'm enough sometimes and I get a hollow ache in the chest. Let's see what this comes back with. So mind mate's kind of thinking. And so with this. Cool. Now it's saying let's take a moment to practice cognitive diffusion, which is about creating space between you and thoughts. So let's say this is hooked up to epistemic knee. What Jonathan was saying is what we do really well is with our implementation of predictive processing as an epistemology, we can ask the next best question to. In your example, get the most information from the patient to map to their core belief and try to get that core belief to a high fidelity to then give them the tool that hopefully helps them move past this and process through whatever they're going through. In this case, it's loneliness and it's related to family and related to these feelings. So this is not yet hooked up to epistemic me, but if we pulled up something on the left here, you even kind of see the conversation. We could export this and there's a log of what the message is and imagine that that log is tied to belief data and that is versioned over time. So you get that fidelity. It's time series. It's tied to the conversation itself with that context, hopefully that gives kind of an idea visualization of how epistemic me meets the mental health use case with AI. What do you think, Vishal?

Yeah, no, I, I really like the, the interface and how easy to use it is. I think that the value when it's tied to epistemic me is that it's going to go through. It's because I felt like it went to the action right away. Like, let's practice cognitive diffusion. Right, like, and, and it was in a way too, too early to start the intervention. But I think that the magic would be in the Layers of the disambiguation and understanding to a very high fidelity, as you mentioned, Robert, and the real connection between the core belief and the superficial automatic thoughts, for example, and the value system, the axiology, how it's tying into that. And in that angle you will have a conceptual map like a roadmap, like a gps, where the AI therapist or the actual therapist can just maneuver its way through those onion layers and really target, or pretend to target one thing, but they're really targeting this so that the therapeutic alignment is high fidelity as well. And I think creating that map as well is going to show the biases, the Bayesian errors, right? Like all the, you know, distortions, cognitive biases, and it will show its rigidity over time. So there will be a temporal map of tracking those beliefs. Yeah, I like really like in the, in the app you develop that a person can attempt to try different types of therapy because you know, the majority are of the therapists are using CBT techniques. But there are so many techniques that also work really, really well for different types of people depending on their approach and whether it's psychodynamics, psychoanalytical or acceptance commitment therapy or different types. So I think that's something that to give someone the opportunity to try different forms of therapy so they know which one works for them is very valuable.

Thanks for those thoughts. I'm curious to ask you a question now that you have this picture in your head. Um, so it sounds like what would be valuable is having, let's say a patient comes to you and you get a, you perform a therapy session and then you're like, hey, here's my app and here's mindmate, personalized to Bishoy. And then they use it day to day past your office visit and then you get over time a higher fidelity map of their belief systems and their value systems tied to these contextual message or chat sessions, which we could do with epistemic me to show their belief coherence along with that map so that over time as they use it more, hopefully it helps them. One, so they get self service treatment beyond the in office visit. And two, you can kind of plug in whenever and see, oh hey, like my patient is lonely, has hollow aches in his chest, like maybe I should intervene, give him a call or something like that. But also you have this growing map that gets higher fidelity over time. That's what I'm hearing that might be useful for you. Is that what you're kind of imagining?

Yeah. So one caveat, I don't do as much therapy currently as we said. I'm in the ED ncl, but I did a lot of therapy before and different types of provision of CBT and different types of other therapies. And I do see that the value is in the two things you mentioned, Roberts. One is creating that map of the ontological tree, the ontological hierarchy, the belief systems and how they are evolving over time. And secondly is to have that companion which all therapists provide some homework to. Most of the therapists at least provide a little bit of homework, like, oh, let's work on this. I would like you to read this. I would like you to journal this and bring back your journals. All of these elements to kind of. To try to tap in gradually to what's going on and to, you know, consolidate and concretize the realizations. So in a way, having that assistant, in a way, we're saying, okay, we're working on this cognitive bias and this relationship, and it creates different plan pathways, and the patient chooses a plan pathway and then progressively the map is being updated with what's really unfolding because something else could. Could appear. And you also, you know, not only due to billing, but just due to the finitude of time and this reality. No therapist can really track everything. So I would, you know, in an ideal world, you want a therapist who really, really knows you, right? Who really knows all the layers, who really knows, like, just like you want lab work that really knows what's going on within you. So in that sense, even the microbehaviors that could be predictive of. So let's use the example you mentioned. Insomnia can trigger manic episode, for example, because of the circadian rhythm disruptions, if that is tracked in the very beginning. And you put that in the O of the subjective objective assessment of plan. And you're tracking also the microbehaviors, like, oh, he's texting way more, he's writing way more, which is a sign of the overproductivity, the hypergraphia that you see sometimes with mania. So, I mean, social media posts are so many, so much like other things that can track and prevent creating because you're putting a lot in the O you can prevent way earlier. And it's almost like predictive psychiatry because we are now in the realm of reactive psychiatry. Like, oh, this patient now have a manic episode. He's in my ed. Let's start the mood stabilizers, et cetera. But we're, you know, predicting in snaps where, oh, the therapist is really concerned Right now because when the last discussion he was, the patient was elated, euphoric, not sleeping etc. And hearing voices. Now we'll send him to the ED, but maybe on a sub syndromal level if we can intervene earlier. And that was like ties us back to the whole heart rate variability discussion and mapping the belief inception change because there is other metrics like loosening of association, which is how tied are these thoughts together. If the loosening of association starts, rumination starts, that could mean that the psychosis is going to happen because it's beginning. Or the cognitive distortions are shaped in an odd belief type of way that is very discrepant from the linear type of logical linear belief system that we know this patient has or this person has. So I'm just saying that also in mapping that it can create a predictive capacity to it by understanding the oscillations and the evolution of the beliefs all the way to the core belief as well.

Fascinating. John, do you have anything to add to, to what Bishoy said?

Yeah, I mean there's a lot of things that come to mind. I think the last thing we're talking about the predictive capacity. I wanted to mention a term that you introduced me to was this idea of epistemic emotions. And in epistemic me we actually like to model the self model of say you in a therapist capacity at least, um, you know, you could think about epistemic emotions being used as like a way to predict the patterns that you're expecting. And then so we actually have these kind of sense making cycles that work in epistemic me. And if there is like say that you got a log back from a patient that you know, was confusing. Right. Or surprising, then the idea would be that, you know, based on, based on the self model that the therapist has, that these things are surprisingly confusing. You can come back to them, you can see there's annotations and we're actually using that as a guide to say what is the next best question to ask the user. It's like one of the data points that feeds into our question generation tool set. So I thought that would be kind of an interesting thing to see epistemic emotions in action and where they, they help both get to predictability and also help to ask the next best question in kind of a guided way.

Yeah, I really love that. I love the idea of using epistemic, you know, emotions, the curiosity and whether the belief confirmation, meaning this person is in the, in the expected paradigm or it's a Surprise where it could be, the surprise could be healing, right? Like the healing, the treatment is working so you're able to track it. Or the surprise point could be that there is, you know, loosening of association or there is a new false fixed belief that is being rising. So the surprise moments in the epistemic emotion could be like the magnifying glass, that there is a delta, there is a change. And by tracking that change, you're either tracking the effect of the treatment or you're tracking going to the words, the confusion pathway in the epistemic emotion and therefore like oh, to something. We need to, we need to disambiguate.

This further and just for the audience, epistemic emotions are emotions that are associated with knowledge acquisition or knowledge states. So just quick definition also time check. We probably have time for one more question topic. Jonathan, do you have any burning questions to ask?

Bishoy I was going to expand on the use of epistemic emotions.

Yeah, go ahead.

To picture a dialogue, we call it a dialectic, you know, where you have a self model of say a coach or a therapist and then you have a self model of a user or a patient. Then now we're modeling the epistemic emotions for both parties. And so you can imagine that if you get to the point where two people are talking to one another and nodding their head a lot, there's a lot of belief confirmations that this is coherence and this is alignment and so that we can say like we understand one another. If you have two people talking to one another and there's this confusion one after another, there isn't information being guided. And so I think epistemic emotions, just like you said, Robert, they're a language of education and learning. And if we have this kind of notion of how well have we learned to understand one another. And you can think of the epistemic emotions as kind of like a gauge. You could see that on an interface as you would maybe would look through a chat log of two people talking to one another. And I think it's a very additive construct. It's something that's represented sort of statistically in the free energy principle, but in the symbolic representations we use in epistemic me. I think the epistemic emotions are kind of like a perfect way to categorize how well does each interaction work. And I'm wondering if that visualization is something that you've thought about before. Shaury?

Yeah, no, I really like the way you explained it. And I also see it and from a Psychiatric point of view that the belief confirmation would mean that the patient is on the expected trajectory, what we call psychiatric baseline, because you can have, like, depression and have automatic negative thoughts and core beliefs of I'm worthless, like, and then, oh, people don't like me. And then the automatic negative thoughts is like, I don't like myself, etc. And this could be your baseline. And talking about them would be just a belief confirmation because it's not a surprise moment. Right. The surprise moment would be like, oh, I feel like I'm. I'm success. I'm a. So that's a treatment pathway. Or the surprise moment could be like, I'm a failure because the government is poisoning my food yesterday. And they're like, oh, this is a surprise moment. This is a breach from the psychiatric baseline. So in that sense, understanding also the trajectory, so you can decipher the baseline so that you are also able to know when to intervene and is a form of the. Either the belief confirmation or the junction that goes either to surprise and curiosity or. Or the confusion. So what I'm hearing you say, Jonathan and Robert, is that there is. You're creating a mechanism by which. By which you're creating perpetual curiosity into the next best question in order to disambiguate, you know, what's behind the skull. Right? Like the whole being, in a way.

Yeah, exactly. That's exactly what we're doing. And we're doing it in a tool like an SDK that any developer that is building any apps related to mental health, health longevity could take our tools and make their tools better by using our tools as a standard protocol and set of interfaces and an awesome data model that we put a lot of thought into and with a lot of multidisciplinary skills and domain expertise to really think through, how would we make this type of experience better for the patient and the provider? I know we're running out of time. One last thought. I'll additionally add. And I'll end it there. So the audience will have to come back for part two because this is so good. We'll have to have part two, probably part three and more. One of the thoughts that I had when I was listening to you talk about this was I use a meditation app every day, like an insight Timer is the app I use. I also have a whoop. And so there's all these apps. And you mentioned the social media app use case, which is, hey, can we get some predictive psychiatry around their usage of social media? Maybe they're texting more. Maybe that gives us some predictive psychiatry in understanding. Oh, okay. This is their baseline. Typically, this is the model. But hey, they're deviating here. It's related to maybe HRV or their sleep is off. And then we're seeing correlated behavior that is backed by literature of, hey, they might be on the trajectory of a manic episode or something. Now what I'm thinking is mental health is one use case, health is another use case. Holistic health is what I was thinking as you two were talking, where we could have all these different touch points that a patient might have even just on their phone outside of the doctor's visit, where we would be able to better help them and personalize care for them. Right. So I'll end it there because we're on time and me and Jonathan have a hard stop. Although I would love to continue the conversation and I bet the audience would too. So stay tuned, audience, for part two. I'll end it here. This has been ABCs for building the Future with me, Robert, my co founder Jonathan, and our awesome guest Bishoy, who we're obviously going to have back on. And by then we'll probably get more reps on this Mindmate app and then we could probably meet it with epistemic me and start tearing apart that use case. Thank you so much for hopping on the pod. Really appreciate your presence and your expertise on this awesome topic and we will catch you next time. Take care of y'. All.