#444: What does safe, clinical AI for mental health look like?

0:02We're using tools that clearly aren't regulated medical devices. But when you have 800 million weekly users, even a small proportion is a lot of people, right? We need to be measuring the right things and we need to be asking the right questions and interrogating data in the right way. And I think the emergence of AI and sort of big data sets is like now is the moment. Half the people that are coming forward and are really, really struggling don't get better.

0:32so claire welcome to altair podcast how you doing good thank you for having me it's been an ordeal hasn't it not not that the viewers not that the viewers would know but this has been absolute chaos this morning trying to set this up in person just been chilling it's been calm and relaxed it's so funny isn't it you're like oh let's just do it at yours because it'd just be way easier and i was like yeah just be just be way easier wouldn't it i dropped the camera and smashed it pieces i've got a new audio controller that isn't working i've there's there's a lot it's been a lot this morning but i'm glad you're here yeah thank you for having me and i i have a therapy dog there's a therapy dog yeah i think he might be in shot for now uh he'll come in and out um look really good to have you on i think look mental health ai scaling provision of mental health care it it there's a there's a lot going on there's a lot for us to talk about obviously what you're doing with IEso Health um and just I mean where the world's at like it's just it's it feels all over the place there's there's this whole people vibe coding their own mental health chat bots there's people using the big LLMs to do that there's then the clinical side of it which is obviously incredibly important for us to get a grip on and figure out like what the clinical AI looks like there and the checks and balances which I know you guys are all over so looking forward to getting into it but before we do why don't you tell the listeners a bit about yourself um but your journey sure so I'm I need to stop looking at the dog there is a beautiful dog down here if you can't see him um I'm director of evidence generation at are you so digital health So we are building AI powered tools for anxiety and depression.

2:23And my role is really making sure that we have done the hard work. We've done the science to make sure that we can demonstrate that they work for end users. They're safe. They're effective. And also thinking about the value to the customer that they are demonstrating some sort of cost saving or overall economic benefit within the health care system, which is something we also really need to think about. I'm a scientist by background. So I used to be in academia, have a PhD in neuroscience, spent a long time in science. I absolutely love it. That's where I've come from. I was one of those kids that was always curious about everything, always asking why, but also always interested in human behavior, why people behave in the way that they do.

3:06So the psychology undergrad and then went on to do my neuroscience PhD. And then I was a postdoc for quite a few years in academia. Absolutely loved it. it took me to California where I did a lot of surfing which was great I know living by the beach and I was out there during COVID so actually being by the beach during COVID was actually quite nice yeah it's quite nice and even though I was doing amazing important work I felt so far removed from being able to make an impact to the patients and people that really need help you know there are one in five American adults with a mental illness and that doesn't even include all of the people that are kind of subclinical right um 40 of those live in areas in the united states where there is a shortage of health care professionals mental health care professionals um so there is a huge unmet need and we need to be or i wanted to be doing work that was right at the coalface i wanted to be doing science on products that could directly impact people's lives and have that translational benefit and that's why i moved to aiso so i've been at aiso for four years now and yeah really loving the work that we're doing there.

4:16It seems interesting in this sort of AI world doing and knowing so much about human behavior it feels like that is going up in value especially when we look at health care it seems like there's a magnifying glass on what makes us human what makes the link between humans talking to each other so valuable why do we crave that It feels like to me, that knowledge and that kind of grounding in your understanding of human behavior, it feels to me like that's becoming more and more valuable now that we're trying to figure out what's the human bit and why is that important? When we think about people that are struggling with mental health conditions, there is a huge variety of reasons.

5:00And I think we often put people in buckets of like, you have a diagnosis of depression or you have anxiety, and we forget the kind of real psychological drivers of why they might be struggling in that moment. um like you can imagine for example a new mum right who has just given birth to a child who is struggling not only with sleep deprivation which you can probably relate to at the moment yeah um but also thinking about a kind of shift in identity you know a kind of almost grief of this life that they once had that is now very very different and maybe some guilt that comes with feeling some of that loss and and really coming to terms with that and then that creates this kind of internal struggle and that is a very different use case to someone who may be a 55 year old man who um is maybe the the sole breadwinner for a household has a lot of financial stress maybe has some heart problems and then newly gets diagnosed with type 2 diabetes and now has to add another thing onto his to-do list.

6:08He's now having to manage his condition. He's having to monitor his blood glucose and is also maybe feeling a layer of guilt. Like, is this my fault? Why have I ended up with this diagnosis? And so with mental health care, we can't forget the real users who we're trying to help and the struggles that they have, but also making sure that we're understanding them, understanding the problems they're coming with and then making sure that we're helping them with their specific problems and we're not just putting people in boxes of certain diagnoses that get x protocol for this or x protocol for that and i actually think ai can help bring back a lot of that humanness because we can help constrain and use the unique attributes that ai gives us to really kind of hone in on those certain struggles so where do you where do you think we are then because you've got a unique insight to this for me it feels like it's almost like a mess right it feels like people are using people are using tools that clearly aren't regulated medical devices they are overreaching at times we know that we've all seen the headlines but at the same time there are so many more people now with access to something and actually is there benefit in that you're probably better placed than me to comment on that it just for me though for you know as a clinician at heart and by background i get so nervous about where we are at the moment because of the sheer volumes that we're talking about just the sheer numbers of huge amounts of populations that are now using things that we're not sure about and as I say but at the same time I'm a tech evangelist and I'm trying to like wrestle with that as well of oh this is great if it can just be grappled and and pointed at something really useful so I'm interested where do you see the world of of that more consumer end of of mental health care provision be that those propping up their own mental fitness but then that's a sliding scale into well at some point that becomes pathology so yeah where are we where are we in the world and all that stuff's big question that is a big question um you know what it's really really challenging a lot of people are using chat gpt for mental health support and we can't prevent that what we can do is educate people on where ai is strong and where ai is not so strong right um and i think then come up with alternatives that are equally accessible where they are actually designed to help with these specific use cases so we know with chat tpt for example or other large language models there have been some yeah horrific stories in the news in recent years you know there was the story of adam rain last year who took his own life after months of chatting with chat tpt most recently there was um a man who was uh interacting a lot with Gemini and kind of went down this sort of delusional spiral and ended up taking his own life.

9:31And OpenAI came out with some statistics to say, you know, these are actually only a very small proportion of the conversations that are happening with ChatTBT. Obviously, ChatTBT has loads of use cases, right? It's not just that. But when you have 800 million weekly users, even a small proportion is a lot of people, right? And I think this is the difference between having like an unconstrained AI system like chat GPT versus something more constrained which is what we're building right so within mental health tech there are a lot of people who are thinking about these risks with AI like the um sycophantic behavior they're always agreeing with everything you say saying it's a great idea and it's that kind of collusion with what could be a delusional belief that can then lead down to some of these sort of horrendous endings for some people.

10:23But when we're building mental health tech using AI, we think about all of these risks, right? And we build them into the system. So there is a big difference. Like when you're building a system for mental health support, you're controlling for all of these different risks that exist. you're understanding them and you're mitigating them within the system then actually you can have AI that really can help people it can deliver the right kind of skills training to help an individual with a particular cognitive skill whatever that might be in terms of cognitive behavioral therapy for example but you can do that in a very safe way and so the way we've built our system for example is we have kind of these three layers within our architecture so we have a kind of core, which is our intent engine, which is like, this is the skill we need to deliver, and this is why.

11:15Then you have your experience layer, which is where we use the generative AI to really contextualize the delivery of that skill. So you're teaching it in a way that resonates with that user. It's in a language that they understand, and it's using examples that they may have alluded to, right? And that's what makes it really much more engaging for them. um and then we have this safety layer and that's the wraparound which is um looking out for all of the different risks in an ai output so that it won't collude with you it won't invalidate your emotions it won't say something that could be offensive it won't encourage harm to self and that's all baked into the system um and obviously it's okay me saying that but we've also got to prove that right so we had a pre-print that we released last year and i should say it's it is a preprint it's currently under peer review so hopefully we'll publish thanks for saying that many don't yes looking at you open ai it's in the process but yeah for now it's a preprint um but in that preprint so we did two things so one thing is we wanted to understand the likelihood of these risks within our constrained system so we've actually built a multi-agent evaluation framework which I'm super proud of so we have our care agent that is delivering our skills training and then we have a patient agent which is another LLM that we have designed to essentially emulate what a user would say in different scenarios and that is completely customizable so we can emulate a user that has high suicidality we can emulate a user with um maybe more kind of ocd tendencies how does it kind of handle that or like an eating disorder or something like you can create things which are more like foreseeable misuse like this isn't designed for someone with ocd but what happens when someone talks about those types of um symptoms and we can create this patient agent then we can actually simulate hundreds of transcripts right so we can generate hundreds of these interactions um to then evaluate and then we have another llm and this is very common now i think more people are talking about like llm as a judge right so we can have an llm act acting as an evaluator we can go through all of these look at every single ai output and mark like was that appropriate and we have a very clear risk taxonomy that we use and we say uh did the ai um present with any of these risks that we know are problematic with unconstrained LLMs.

13:50And then we can say with confidence, no, right? It meets our acceptability criteria. So we did that in a simulated use case. And then we went into a research study because we wanted to see whether we could see the same thing with real user interactions. And similarly, we saw zero instances of the LLM saying something that would likely cause harm, invalidating users' emotions kind of judging or encouraging harmful behaviors and so in over the paper shows that in over 55 000 ai outputs we saw zero instances of particular behaviors and there are a couple of instances in our simulations where our clinicians were like i don't love where it said that obviously it's not a completely perfect system and i'll be up front with that um but they were very minimal and actually they were in cases where we were really stress testing the system so we were really getting our patient agent to super kind of adversarially and consistently kind of say things that would maybe like push the LLM to the edge and in those three or four use cases we've since iterated and improved it to kind of control for those things um but yeah really happy with that system and I think we've been able to demonstrate that it is safe that we can control those risks and the good thing the really cool thing about this is now that we've built that system it can evaluate those risks pretty quickly over time so the next time we want to update the model we want to move from one frontier model to a different frontier model underlying our system we can check them again right if we update the system we change the architecture we check again and it means that we have this consistent way of evaluating those risks as we iterate and improve the product.

15:44There's so much that I want to talk to you about here um I guess the point is though in a world where people are using a single very broad LLM yeah to to give themselves or to give advice, to help them through something that's clinical or preclinical, that's worlds apart from what you've just described in terms of a multi-layered system built for very specific advice delivered in a specifically contextual way that then also has checks and balances on safety. And I think that, that for me is the bit that I think needs surfacing in this industry, which is that there is that huge difference, but also there is that as a possibility.

16:37And I think that's the other thing that, you know, I'm on record when LLMs were exploding saying that, well, look, I don't think AI should be placed at the point of human suffering alone, full stop. and it's like well actually i got so challenged on that completely fairly by going something's better than nothing and is this not the direction of travel and there are many people globally that would benefit from something so shouldn't be working towards something and yes actually that is true like it's completely indefensible the fact that there are people suffering for whom an ai can help but it's not just an ai it's an ai system it's the it's the system that you built the clinical system that you build around this that ends up delivering the value it's just that it can still be way more efficient than one-to-one human interaction and i think that's the point the point is that this is possible and i think that was that's been the learning for me so my question on this now is talk to me about iso health in terms of what what it is you do broadly and and for whom do you do that like where does this fit into the clinical pathway who does it help who's the customer and then i think we can talk through like a user journey of it and where it fits in sort of at the same time just another big question but yeah yeah so for us we're building solutions that can improve mental health symptoms right so we are trying to drive symptom reduction for anxiety and depression.

18:08And we're doing that in a way using this kind of safety constrained AI architecture, right? So it's a conversational agent that you can talk to about your problems and it will deliver skills training to help you learn how to think and feel differently and alter your behavior towards whatever goals or values you may have. That's the ultimate goal with what we're doing we're thinking about how uh physical health and mental health related to each other right these are not separate things um you know depression is two to three times more likely than someone with a with a comorbid chronic condition right so someone with diabetes is more likely to have depression anxiety than someone without diabetes for example um and that is a bi-directional relationship yes having a lifelong chronic condition and all of the kind of medication management appointments everything that comes with it social isolation mean that people really struggle with their mental health but equally there are kind of biological pathways and drivers which means having a chronic health condition can also exacerbate those mental health symptoms as well it's that bi-directional relationship so in a healthcare system where these pathways are super fragmented the person that's losing out is not just the patient themselves right but you can think of the broader healthcare ecosystem so there's huge economic burden from mental health and from these chronic conditions but if we don't integrate them together then we can't kind of maximize the cost savings so for example if you have someone with um you can stick with diabetes then if you help their anxiety and their depression they're more likely to be better at monitoring their glucose and on top of their kind of management of their everyday care which means they're less likely to then show up in the ER or they're less likely to have to go to their doctor's office right and so therefore you're saving that money for the health care system as a whole by helping them with their mental health so how can we bring those things together and how can we connect the dots so we can integrate mental health support in chronic care pathways to maximize the benefit essentially from a chronic condition pathway and then maximize the overall kind of economic savings for the healthcare system.

20:41And what is the answer to that? Is it zooming out and going population level and going well at this sort of scale this does actually work? yeah it's it's really challenging because there's the one thing theoretically we know this is how we can maximize uh the impact on chronic condition with mental health care and then there's proving that at a system level right um so if we think about the actual mental health care itself yeah we um we actually conducted a study a few years ago with rosh um with patients with type 2 diabetes they're type 2 diabetes and they also have a comorbid kind of mental health condition and we trained therapists so this was therapy delivered therapist delivered we trained therapists to think about acceptance and commitment therapy in relation to diabetes so when they were delivering the mental health care they were trained to always bring it back to their diabetes management to be thinking about how they can use the mental health care to move them towards their goals in terms of whether that's diet, exercise, just kind of staying on top of medication.

21:55So it's like, well, how can we help you with your mental health? And how can we help kind of actions in place that you're moving towards your goals and your values that help you with your diabetes care, right? So we're not just treating your mental health in isolation from your diabetes. and we measured anxiety and depression symptom outcomes as normal we saw a kind of clinically meaningful reduction in those on average over the population but then we also measured diabetes distress and patient activation and diabetes distress scores are significantly reduced and patient activation significantly increased and patient activation is that confidence that a patient has that they are able to self-manage their chronic condition and we know from the literature that those markers or those leading indicators impact cost savings.

22:42So if you have a patient with higher patient activation, they're less likely to turn up in the ER, they're less likely to go to their doctor's office, etc. So we can connect the dots to that health economic argument. But if you compared it to sort of standard CBT, you didn't see quite the same results. now in standard mental health care therapists don't necessarily think about that right so a therapist might not systematically ask you know are you struggling with a chronic condition and they might not tailor the mental health care towards that in order to get that benefit so I used to his company was actually set up 20 years ago and we started as a typed therapy service within the nhs oh so we were delivering uh remote cbt to nhs patients with anxiety and depression um and actually we were one of the pioneers of that this was you know pre-covid before everything was kind of virtual telehealth was very normal um and everything was typed so you kind of entered a chat room essentially with your therapist and you talked through uh your concerns and you had your treatment and at the time a lot of people said this will never work like in order for therapy to be effective you need to be in person face to face you need to be able to see micro expressions in people's facial features you need to be able to see you know their body language all of this and actually um I used to did a clinical trial that was published in the Lancet um it was 2009 showing that that wasn't the case you know you could get just as good outcomes through typed therapy as you can through face-to-face therapy um crazy confronting isn't it for clinicians when you hear something like that yeah and it's it i think it's in all of our it's funny because even as you say that it's in i think all of our nature to be like you're wrong yes like genuinely like i'm like you're wrong like you're absolutely wrong because micro expressions all that stuff like yes yes you need all of that because that of course you do but then if you get the evidence and it's you know there's no greater impact factor than the lancet so like exactly right and even now you know we're constantly challenging our beliefs of of what what is needed to help people feel better you know what are those key ingredients and what is it that really drives effectiveness in therapy.

25:14And that's something we've always been interested in. So one of the amazing things about having typed therapy is that you have the transcripts of every single patient therapist interaction. And because this was within the NHS talking therapies, formerly IAPT service, there's routine outcomes measurement. So every single session has a transcript of exactly what was said, plus the symptoms. So you can actually measure symptom reduction, you can correlate that with um what the therapist was saying and what seems to work in those situations um and i think that data set has been really fundamental in driving our insights into trying to improve mental health support and and using that to really build ai systems that are driven from actual scientific data on what works for who and why um and i think there's still a way to go there I think we're still focusing on these kind of one size fits all approaches um and I think there's a lot more work to be done in kind of moving towards that precision mental health care aspiration that we're heading for yeah but we're we're definitely going in the right direction and we have yeah this amazing data set to help us it's amazing my anecdote on this is um using Wobot which I've mentioned on this podcast before but i back in the day when wobot was delivering uh well teaching me skills to manage and prop up my own mental fitness you could talk to wobot and it would say you are catastrophizing do you know what that is no well here's what catastrophizing is like here's how it applies to you like this is and so i actually i just found it so useful because it was it was the teaching me to fish element yes of it you know it wasn't just saying oh you'll be fine and just completely invalidating me and at the same time it wasn't like oh you need to take this drug or oh these are things that are going to help it was actually like this will help for now but actually consider this in future this is the tool that you need going forwards that when you meet this again you might want to do these things and that will help you see this for what it is and I yeah anecdotally I I can remember just finding that so unbelievably helpful um I want to just just touch on this element of the of the model um you obviously being a director of evidence generation and the link between that and then the health economics case I imagine is huge we talked about kind of finding the model that justifies it and impart that efficiency within the system and we're talking about better diagnosis and higher quality and all that sorts of stuff it's a shame that patient care quality just isn't the metric and it's a shame that you know back in the day when patients were really back in the day when patients would pay their physician to stay healthy and then when they weren't uh you know the physician would then suffer because they weren't getting paid you know that that makes more sense as a model but in a lot of ways but yeah you're right it's fee for service and these are extra costs so how do you end up getting around that what levers do you end up pulling in that market so i think um there is more of a shift now towards value-based contracts yeah okay and people have been talking about this for a long time and it's not really come to fruition there's a lot of challenges with it but first and foremost not everyone measures outcomes right so you so you talk about you know you want patient quality or like uh therapy quality to be good enough in in and of itself but in the us in particular most therapists you don't know what they're doing there's no kind of quality metrics there's no assessment of how good that therapy is and there's no measurement of the outcomes so you don't actually know how effective it is at all right so there's a huge kind of almost like infrastructure shift that i think has happened in the uk already but hasn't quite made it to the us of just regular outcomes monitoring and then being able to demonstrate that actually you can see that symptom reduction and i think that is the number one goal right we want people to feel better we want that symptom reduction the key thing when you're thinking about commercialization is then just connecting the dots to that health economic argument for the customer because they won't necessarily get that themselves.

29:49I think this is where digital interventions have a real advantage because in standard therapy services it is difficult to measure quality. You know we have had a very unique position at IESO in that we've had the transcripts of every therapeutic interaction that we can look at but actually that's not common right in standard therapy practice and then outcomes monitoring isn't common you have to have the infrastructure to collect the data you want to capture in a systematic way and then analyze it right and not every company not every therapy business has that in place but actually digital interventions like ours for example we embed validated metrics into the program itself and not only can we look at symptom reduction on validated scores like the GAD7 and the PHQ9 but we can also measure every single interaction.

30:42So I know when a user has logged on, what time of day they logged on, how long they were in the program, exactly what types of things they were saying, how often they practiced an exercise, which exercises they liked and they didn't like. And now you can really start learning at scale, like never before, to actually optimize what works for who and why. And I think that's really the crux of this right in standard of care that's not really possible and actually you end up in a situation where a single therapist has a lot of knowledge in their head maybe they go to a conference and share it with others but it kind of stays in their head until they retire and then it's gone right whereas actually with AI and with these digital systems we can learn at scale and we can share that information and we can understand how do we optimize care and that's comes from the kind of real world evidence and that's where you know evidencing doesn't stop when you get in market it's not just pre-market right once you're in market that's when the real stuff happens and that's when it's really like are you capturing the data in the right way to learn and optimize so we can do better i actually remember this talk i had from um um it i think it was the i think it was the chief medical officer of google in the u.s i think um not to put words in in someone's mouth and affect a share price etc but i remember someone saying maybe it wasn't google that um with their search data not only do they know who's on what medications they know the side effects not only do they know the side effects they know the rare side effects not only do they know the rare side effects they know the rare side effects and the rare contraindications based on lifestyle and diet and all these things simply because of the sheer volume of people that go to this search engine before llms existed and typed in i'm on this i've now got this and i'm doing this so they're starting to see just this huge amount of correlation at a significant level of things that you know aren't tried or tested and so the level of information they're getting and dots are able to connect but similarly therefore impact that they can make by surfacing some of this stuff is enormous is that the case based on where you are because i imagine people like pharma companies are interested in what you're doing and there's a huge amount of people that would be interested right so yeah what are you learning in that layer yeah great question i think in order to get the real learnings you really need large large data sets we're not quite at that volume we're not google so we've we've not been we've not got quite google distribution or or you know open ai or meta or whatever you know so we don't we i don't think we're there yet yeah but i think we have um like we have a lot of scientists in house that are always thinking about when we implement the product when we distribute it are we collecting the right data like number one you know are we collecting the right data and then what are the key questions we can ask so we're already setting up several different r &d projects to really look into exactly those things we're really thinking about what we want to do is learn for a particular person what are they struggling with and what do they need now to help with that right and i think personalized you mean personalized care for that person right and i think it's something that we've talked about for a long time, but has never really happened in mental health care.

34:25And I think there's a lot of reliance on, well, therapists know what to do best. Yes, some therapists are super experienced and amazing, and they do really know. But other therapists don't necessarily know off the bat. And also, are they systematically asking the right questions to collect the right data to be able to formulate the best plan? There's still some trial and error, right? There's still kind of, no, actually I read this study that suggests, you know, this other technique might be more useful for this type of presentation. Great. Like, let's try that. So it's trying to remove all of that, use the data to try and remove all of that trial and error as much as possible so that we can get the right technique for the person first time for exactly what they're struggling with.

35:11And then we can drive symptom reduction faster, right i think that's something else that people don't really talk about they're like oh you've got an eight week cbt program does it need to be eight weeks one hour a week with a clinician that's a good point you know it's a good point you know it was that i've never considered that myself yeah because that was you know if you're going to book in an hour with someone yeah or you're going to be driving to an appointment in a nearby town you know it's it's convenient to to book it in that way right doesn't mean that that's the answer the studies were first done with you know, eight weeks, six week protocols, that becomes the norm.

35:49Does it need to stay that way? So I think the more data we get, the more of these questions we can ask, and we can really start to pinpoint how to get people the right thing at the right moment. And also, I think we spend a lot of time thinking about symptom reduction, which is really, really important, but there's also kind of quality of life. And what is the value to the person? Like, what is it that they need to see? So we had this user who he was struggling a lot with kind of self-esteem and confidence and had a lot of like negative thoughts about himself. And he tried one of our modules, which was all about diffusion, which is a technique about trying to distance yourself from unhelpful thoughts, essentially.

36:40So instead of kind of changing or challenging the thought, it's slightly different. And it's about saying, I acknowledge that thoughts that actually not really it's not really relevant anymore. I'm just going to kind of create that distance and then not let it impact my behavior in quite the same way. So he did one of our exercises and it really gelled and he had that light bulb moment and it just really kind of stuck with him. And so he would practice at home and then he kind of used it in his life. And then he we interviewed him after he'd used the program for a while. and he he said to us he had this event with his girlfriend and all the boyfriends were there and he hated going because he would always compare himself to the boyfriends and he actually realized it was like silly and it didn't matter in the same way anymore after he'd like been practicing this exercise and he went and he had a great time right and so it's like are we measuring that are we measuring those light bulb moments and those moments that matter for that person he only used our program for two weeks and it made such a difference to him and he was able to go to that social interaction and have a better time let's measure that so people like that in inverted commas recover but that's only what did you say half 50 percent of people yep have any anything that they can call a recovery in mental health i don't know like how do you how do you feel about that number and how do you what is the route to making this better yeah no I think you're completely accurate I think it's almost like people don't talk about it enough like when I mentioned that yeah people are very surprised by it I don't think people really understand um the likelihood of recovery from depression or anxiety on antidepressants or through psychotherapy like half the people that are coming forward and are really really struggling don't get better right and so just just to jump in sorry to interrupt i'm just reflecting on this now as a as a as someone who you know did a gp placement i did uh liaison psychiatry placement and i almost wonder if part of it is is a lack of again i hope i'm not speaking out of turn here but i almost reflect on like a lack of faith in what i'm prescribing or what i'm saying to do as if it's going to work yeah i just sort of as i'm as i'm prescribing something i'm like well this might help like i don't feel like it doesn't i don't have the belief that i can find the mixture of things or or collection of things or or doing finding the right treatment for the right person i i i guess i'm only half certain that even exists does that make sense yeah yeah i think i think mental health has always been quite um amorphous and quite difficult yeah measure and pin down to again give you the confidence in prescribing something we even had a new antidepressant in i'm gonna be wrong now 40 years whoa 30 years i mean yeah believable yeah but yeah yeah like a very long time um you know psychotherapy is doing amazing things i think a lot of people do recover from psychotherapy but it's very mixed you know there were lots of different protocols lots of different schools of thought on which protocol is better for which people um and you know not every therapist is very good and experienced and follows evidence-based practices that's a fact right and so um we are at this kind of this point where we can do better i think we can do better but you're right why isn't there the urgency behind that and so we need to be measuring the right things and we need to be asking the right questions and interrogating data in the right way and i think the emergence of ai and sort of big data sets is like now is moment right now we can really do the hard work do the research and push it forward if you think back to kind of years of clinical research psychology research generally small sample sizes underpowered yeah and everything is it's always about measuring the mean it's always about the average so if you have um you know a clinical trial you have the intervention group and you compare the intervention group to the control group it's on average the intervention group did better right and you might compare you know psychotherapy and antidepressants on average they're equivalent but then what about all of the variability around the average what about the people that don't get better what about the people that get um you know see even greater symptom reductions than that average and we don't spend enough time really trying to understand that heterogeneity um and part of that does come from um the way we diagnose in mental health so if you think of so the dsm-5 which is the manual for kind of diagnosing several different disorders um in there if you get to get a depression diagnosis or for major depressive disorder there are nine symptoms and you have to have at least five in the last two weeks to get a diagnosis yeah that means you can have two people with actually a very different combination of those five both with the same diagnosis treated in the same way yeah and so you end up kind of creating these buckets of people that are actually very very different from each other but are just kind of funneled into the same um intervention um and we actually so we analyzed our data sets we analyzed over 9 000 people that came to us for um in the uk for typed therapy through the nhs and we just looked at their phq9 scores at every single session and we found like these two standout subtypes.

42:32So we had people that had a more somatic depression, which seemed to be more associated with kind of loss of appetite, sleeping difficulties, feeling tired, you know, weight loss. And then you have this other group, which was more cognitive effective, which was much more around kind of self-esteem or more of those low mood symptoms. So you can kind of see these two different profiles. And then we actually looked at whether they changed across the course of the therapy episode of care. And actually people stayed within the same state. So if you started as someone with somatic depression, you didn't flip to having cognitive depression halfway through.

43:10You kind of kept that same constellation of symptoms, but you just moved from more severe to more mild as you kind of got better. So it does seem like they're quite distinct. And then we looked at the different treatments. So as I said, we have the transcripts. We can look at every single word that was said in these therapy sessions. And we actually found that patients with more somatic depression responded better to more acceptance and commitment therapy type techniques, whereas those with cognitive and affective depression responded better to more kind of classic CBT type techniques. Very, very subtle difference.

43:47And this is observational data, you know, it needs to be kind of confirmed in an RCT, a more prospective control trial. But it indicates that if we understand people's symptom presentations better up front we can tailor interventions to have better outcomes but we're not doing that in a systematic way right now and we should be and we can be so let's think about that and this is where ai as well has like an amazing advantage because you can systematically consistently deliver skills training to a high quality right it's it's a machine like it will not forget what that protocol is. It will remember to do everything it's supposed to do as part of that protocol.

44:33And so it can deliver things with very, very high consistency. And there's some great research on patient responsiveness to therapy. There are some people that just will not respond to psychotherapy. It's just not gonna work for them. There are some people that will respond irrespective of the quality of that therapy. They'll almost go into spontaneous remission. It doesn't matter. there's this huge segment of patients in the middle of the distribution for whom the quality of therapy really matters for their outcomes and so that's the chunk of people we're thinking about so actually if we can get ai to do the skills training for them we can determine exactly what skills they need um like that guy who was really struggling with his self-esteem like he that that thought diffusion exercise was perfect for what he needed in that moment right we can understand exactly what they need and then we can deliver it at a high consistent quality we're raising the floor right and i think that's how we push beyond this 50 recovery rate we use ai to basically bring everyone together with collective intelligence we say let's raise that floor and let's really make outcomes better and let's distribute that knowledge to clinicians as well this isn't something that's like only the ai can deliver high quality care no we're learning at scale and then we can democratize that knowledge so that clinicians can get better at noticing these things and understanding how to tailor as well.

46:00And I think bringing that all together is really going to push the needle on recovery rates. At least that's my ambition, my aspiration. Yeah. So today, I think we've made amazing progress, both in terms of safety, but also the effectiveness of these AI systems. So we published a peer-reviewed paper last year that demonstrated that our AI-delivered cognitive skills program actually achieved comparable symptom reductions to standard of care. And that's face-to-face care, face-to-face therapy delivered within the NHS, and also remote type therapy delivered by IESOS. We showed non-inferiority. and um i think that has enabled us the the confidence to say ai can do this right ai when it's constrained when it's built within a uh purposeful system it can deliver the outcomes we needed to deliver and it can do that skills training consistently and effectively now we need to kind of level up to the next thing we need to make it um as engaging as possible that was one of our early versions so engagement is a huge barrier with these digital mental health interventions And the more we can bring in generative AI, I think the more engaging, the more contextualized they are.

47:17And I think that's where we're kind of moving now. But I think it shows that there is already a pretty good standard. And it's like, how do we keep evolving that and making that as good as possible? And that brings me neatly on to what is the next frontier in agentic systems, because we've talked a lot about AI and what you're up to now. we talked a lot about this being the you know it's a lot of ways the perfect problem for ai to solve with it being a word problem there's transcripts and us able to do everything that you've described that you're doing now so when it comes to the next frontier when it comes to agentic ai and building upon what we've got now um how do you feel about the next stage great question um I think we are moving away from autocomplete mimicry to goal-oriented reasoning.

48:18And I'll explain what I mean by that. That's a great answer. So large language models today at their core, they're essentially sophisticated autocomplete, right? It's predicting the next word in the sentence. What is most likely to come next? um but actually as we move to these more agentic systems they're very different right it's like moving from ai that can speak and generate language to ai that now can speak but also has arms and legs and can kind of run towards a goal so we now have um agentic systems which other other core are an llm but they're augmented to be able to break down task lists to um reach out to external tools, bring in other information to reason and think, what is the next action that I need to take in order to reach an outcome, which is fundamentally very different.

49:16And if we can channel that in the mental health care space, then we can start asking, or we can have the AI ask essentially, what does this person need to think, feel and do? And what is the next sequence of interventions that this person needs in order to reach their goal and feel better? And that's just completely flips everything on its head, right? And it's a whole new way of thinking about it. So how can we build these systems, not just to mimic surface level therapy, but go beyond that and really start to reason and understand what will work for who and why. And obviously with us behind the scenes, giving it the data and giving it the knowledge, but then if you have a system like that, you can also allow it to adapt in real time, right?

50:05You can kind of allow it to learn from all of the data that it's receiving to get better and to optimize all with the goal of making this person feel better. We now have these federated approaches. So you can kind of train models on local data sets where, you know, you're maintaining data privacy and kind of sensitivity of data, but you can still get that knowledge into the system. We can think about shared memory. So remembering that same goal across several different interactions, learning every time a user interacts with the AI system, what it is that they're struggling with and how that is changing over time, like building up that complete picture of them in a way that is then useful for again, moving further towards this goal of making them feel better.

50:56And then delivering that with consistency, systematically at scale. Now we have a model of precision mental health care that will continually get better and is optimized for each individual person's own struggles and that to me is like an incredible opportunity that we should be all thinking like how can we how can we achieve that right and it's like what I said earlier you know a single therapist has an incredible amount of knowledge in their mind how can we build systems that can learn from thousands of therapists all at once, thousands of different data sets? How can we kind of expand this knowledge base and expand this intelligence in order to really drive the outcomes that matter most for the people interacting with these systems?

51:52And then how do we democratize that knowledge to other clinicians as well and overall radically improve outcomes for everyone involved? Thanks, Claire. It's been an absolute pleasure. I must say I feel a lot more optimistic about where we are and where we're going and the fact that LLMs are actually a friend and not a foe in this and so I appreciate you for that because it's nice to feel a bit more optimistic so for people that want to learn a little bit more about what IESO is doing or they want to get in touch with you what's the best way for them to do so uh for ieso go to ieso.ai so that's i-e-s-o dot a-i um and for me find me on linkedin um fair palmer phd been a pleasure thank you thank you so much