#425 Is it unethical not to use AI in mental health care? (Part 2): Dr Ross Harper from Limbic

0:00Welcome to the Health Tech Podcast. Here we talk about everything healthcare and technology. and I'm your host James Summeru. Hey everybody welcome I have with me or back with me I should say Dr Ross Harper the CEO and founder of Limbic. I'm on the road at the moment hence I don't have my fancy usual podcast background and I've got my Yeti Nano that I entrust with my on the road activities so yeah I hope the sound the video quality is good enough for you anyway but recapping on part one with ross that we did a few weeks ago um limbic obviously the clinical ai used across the nhs uh since about 2020 they've supported close to half a million patients now used by around 45 percent of uk services in part one we talked about if you want to go back and listen to it we talked about what clinical ai really means regulated evidence-based integrated into pathways and not just this consumer wellness stuff with a chat window.

1:06Limbic being the only mental health chatbot with class two medical device status, which raises the bar on safety, efficacy, and oversight. And what I found super interesting that we talked about last time was how Limbic does diagnosis. Today, we're going to be talking about treatment, but to just remind you about diagnosis so ross explained why he never lets an llm make a clinical decision so at limbic they have a dedicated predictive system a clinical reasoning layer which ross described previously as similar to how you might think of older school machine learning remember that term everybody before llms and how that does the diagnosis support the llm is simply holding the conversation And that pairing is what has led to the success, both from a diagnostic perspective and from an company perspective of Limbic.

2:03Now, you might remember last time, if you listened to it, that there is also an equity piece here. They did a huge clinical study where people knew they were speaking to the AI and they still rated it as empathic. And crucially, this is the bit that really made me think about things minority groups were more likely to self-refer when it was AI greeting them at the front door rather than a human and that was one of the things that made me talk about the attention to detail that limbic have over most other people that I've talked about in this space they make really small what seem like small but very big choices neutral voices careful wording you know you might benefit from help with things like that all of these things matter and clinicians retaining final oversight of course and so Ross welcome back delighted to have you mate and really looking forward to getting getting into this because as you can tell it's a bit of a it's a bit of a passion project of mine to kind of get to a point of us actually using AI in the right ways with that clinical oversight from human beings so yeah welcome how are you I'm great thank you for having me James and thank you for the the introduction you uh I've got nothing to add you you said it better than I could invite me to the next investor meeting Ross what are you series b series c next it's probably a big one absolutely yeah it's great to be back good good so let's talk about treatment then so set the scene for me if you want to kind of round off anything that I said about the diagnostic side and the link to then treatment for me.

3:46So I guess, first of all, obviously limbic being known as in part diagnostic, and we've obviously been through that, it does diagnose you. How much treatment does it do? What treatment do you guys do? What are you comfortable with? And let's, I guess, start there. So it's diagnostic support on the front end. And we spent a lot of time in our last conversation talking about that intake, that triage, that clinical assessment. And we have this diagnostic model that is, as you mentioned, a class to a medical device, which helps behavioral health clinics assign patients to the correct treatment pathway.

4:27So now we're in. Now the patient is in. They've used AI at that first step of treatment. And now they're waiting. They're on a waiting list. one of the benefits of AI is it's 24 7 available so it really does create a world of abundance and in healthcare that is important so you've just entered and rather than waiting for a couple of months for your first treatment session with a clinician Limbic is there immediately and it's able to now start delivering validated tools from cognitive behavioral therapy and it delivers these through generative conversation okay so it's these large language models but they are delivering validated therapeutic exercises active listening psychoeducation you know everything from the literature um and this is important because cbt which is one of the most widely used mental health care sort of treatment modalities it's very skills-based it's all about adherence to the treatment protocol and that's where generative AI and large language models become critical because beyond just being this sort of cool sexy technology they also solve for one of the biggest problems that has plagued mental health apps up until now which is that no one uses them for very long okay an independent study for digital mental health solutions separate from limbic an independent study found that amongst the top performing mental health apps in the app store average two-week retention was four percent right that means one in 20 people are still using that app after two weeks you cannot have a therapeutic effect if no one uses the tool so large language models and generative AI, they buy you engagement.

6:33They make the solution adaptive, naturalistic. It gives it a memory so that people come back. And, you know, you can see the effect that that is having across all different industries. ChatGPT gained hundreds of millions of users in record time because the user interface is so adaptive and engaging, right? So you should think Think of generative AI and large language models as solving for one of the biggest problems in digital mental health, which is user engagement. That's why we need it. And so if we can deliver cognitive behavioral therapy tools and techniques through these large language models, you buy treatment engagement, which is the way it has a therapeutic effect because it's a very skills-based exercise.

7:20And that's what we saw at Limbic and why we worked hard to unlock large language models by making them safe, clinically accurate and explainable for use in a healthcare setting. We didn't need to be convinced that large language models would have a powerful effect on the patient and clinical outcomes. What we needed to solve for was how do we make them safe and how do we deploy these responsibly within a care pathway. That's where all our research efforts went. And so just to be absolutely clear, then it's the same thing that you talked about on the diagnostic side it's that the llm is holding the conversation but it's the reasoning layer that is actually making the decisions absolutely and as we discussed last time so so james you're quite right to pull this out okay in theory it's identical what's happening at intake triage and clinical assessment and what's happening in um cognitive behavioral therapy in terms of the technical stack is the same thing but the goal of the ai agent is different because it's at a different point in the care pathway that's the that's the only difference um and so yes we use the limbic layer this specialist clinical reasoning system which pulls from the evidence base the clinical literature and uses specialist machine learning models to evaluate all conversations against different clinical criteria but now we are looking at cbt tools and techniques rather than the initial clinical assessment so this there's an obvious question that this brings to my mind here now which is that and if i may be frank people are probably asking chat gpt just chat gpt without your reasoning layer the same sorts of questions from diagnosis but also probably from treatment hey chat can you just give me a bit of cbt to help me through this thing or maybe perhaps not as as gauche as that but you know what i'm saying right essentially people are doing this diagnostic and treatment activity with just chat gpt how do you feel about this you're right and just to add some color to what you're saying james i believe unless it's changed recently i believe well-being wellness and mental health conversation is the top use case for chat gpt it is we've put it into health tap pigeon a few times there you go it is the top use case okay Last week, there were 800 million unique users on ChatGPT.

10:10My goodness. So, last week. So, if that many people are using it, I know, right? It's in absolutely unfathomable scale. Last week, 800 million people used ChatGPT. And we also know that the primary use case is mental health-related conversations. okay so the horse is bolted oh yeah this is out there the question then is no longer should people use chatbots for mental health related issues because it's not should it's happening the question then becomes do we want to bring this inside the boundaries of health care or do we leave it out in the unregulated wellness space. And I think that's where we really hit upon the key point.

11:03Because healthcare is not the same as wellness. Healthcare is a regulated industry with standards that must be met, and, you know, an evidence base behind it. So that's the way I look at the situation, I am fine for wellness to persist, and it can be very helpful for a large number of people and that's fantastic i'm really pleased about that and in many ways chat gpt has democratized access to information um with with valuable wellness information being a key aspect of that but we must not conflate wellness with health care we must recognize the differences between these two things and when it comes to mental health care generic chatbots are woefully incapable of delivering rigorous evidence-based intervention.

11:58And we are already seeing with some tragic news headlines that I'm sure you've come across, James, we are already seeing what happens when you allow generic chatbots to overstep with vulnerable users suffering from diagnosable mental health conditions and that was actually going to be my next question what what what is the problem so overstepping clearly is one of them how else does this problem show up obviously because again it is being used to therapeutic conversations these are happening to your point and we are seeing that turn up in the news i think there's a there's a there's a job that we can do because even people listening will probably be using it you know edge cases and things like that or in the run-up or there's and certainly they'll there'll be clinicians listening that see patients there'll be gps listening that see patients that are probably going to use it in this way what can you what can you tell what can you arm the listeners of this podcast with to as as evidence for why they shouldn't be doing this how how might this show up in a negative way the more people that use chat gpt for those more health care conversations as it breaches wellness what what are the issues that people are going to see is it wrong diagnoses is it is it i mean is it is it going to shot i mean you must have used charge of bt for this stuff does it does it actually give you does it try and give you a diagnosis or does it direct you into the system it depends on how you prompt it quite frankly um it's just it's fallible in many ways i think it tends to be less about diagnosis and um which it it often tries to steer clear of because it's been told not to sure sure within the the literature you may come across sort of a distorted thinking patterns right cognitive distortions are a very well known um uh sort of aspect of cognitive behavioral therapy maladaptive behavioral patterns and what is a distorted thought pattern and what is a maladaptive behavioral pattern differs depending on the context chat GPT is not well versed in adjusting its language or behavior according to these sorts of clinical constructs.

14:13A concrete example that often comes up is because it has sycophanty tendencies and it seeks to please, it seeks to engage, it often tends to agree with what people are saying. And so you can very quickly create this um spiral where chat gpt may well promote somebody with an eating disorder going on a diet which would be um totally inappropriate given the wider clinical context of the situation but that's where that's that's where it tends to misstep we had a tragic uh case recently where somebody who had been uh preparing to end their life um you know chat gpt was promoting the decision, was supporting the conclusion that the user had made, and was even suggesting places nearby where they could purchase a rope.

15:07And so it's these sorts of scenarios where it becomes patently clear that it is not equipped to handle vulnerable users suffering from acute mental health issues and how could we expect it to be james because it's not a specialist system for those types of users it's meant to be serving people from multiple different backgrounds in different psychological mindsets and so if if the same model that is uh you know wrapping the recipe for a quiche in the style of M &M is also needing to deliver these very sensitive and nuanced conversations. It's a bit strange that we would expect it to be appropriate.

15:59And so that's why I think we need specialist systems with a clinical evidence base behind them, whose behaviors are constrained by the clinical literature, and where you have these ongoing clinical models like we have at limbic looking at cognitive distortions looking at maladaptive behavioral patterns cross-referencing everything it is saying with the clinical protocols love it love it and again you're showing me that you're showing me the detail again which i think is so so so important here so let's let's talk about outcomes then you talked about the improvement in engagement you're we've it's great that we've got llms because they're going to increase engagement is this a better system i'm going to be sort of bold in my question here is human plus ai better than just human or are we aiming for parity well look i i think it's pretty obvious at this point that ai plus clinician equals better care i think that's an obvious statement I think that is supported not by conjecture and speculation, but there is now a growing and consistent academic literature and clinical evidence base that shows that things are better.

17:20Exactly how and why does vary depending on what you're looking at. So at Limbit, we discussed last time, James. um there are access to mental health care is materially improved by having an ai at the front door handling intake and triage we have seen that across multiple sites and hundreds of thousands of patients we also see diversity equity and inclusion improve and the holy grail we saw clinical outcomes patient recovery rates improve due to better identification of the primary presenting issue and timely allocation to the correct treatment pathway okay so these things are just a given these are these are high impact peer-reviewed findings at this point we see something similar happening in treatment and again i can pull on limbic's data just because i know it's top of mind for me um but we are seeing a within treatment so the clinician continues to have their weekly or fortnightly sessions with a patient but because limbic is being used in between those sessions we are seeing better patient activation better engagement through the treatment protocol more likely to show up to their next session and we are critically seeing they recover faster in fewer sessions, okay?

18:49So by having this always on clinically rigorous and evidence-based AI, handling all the time in between those individual sessions, progress through a treatment protocol is much, much faster. And as a result, you tend to see outcomes improve. And again, we've got peer review findings that illustrate this point, I think, quite nicely. So in that case then Ross, what is the opportunity here? Because given where mental health service provision is in the UK worldwide, and we can actually talk about both here, I think it's probably important that we talk about both, that we've talked for so long about the volume issue.

19:37There aren't enough human beings to deliver one-to-one talking therapies. It seems that the conversation then turned a while ago to, well, perhaps tech can help us. Perhaps we can scale human delivery through tech, through telemedicine and things like that. Now, obviously, AI coming in, perhaps tech can do a good job. We're now, as you quite rightly point out, we're seeing a position where volume and quality can both now go up. Is the opportunity therefore now to, I don't want to say solve this problem, because in some ways, you know, there's always going to be issues. But do you see the next two, five, seven years of us making a serious dent in what is frequently called one of the biggest problems that the NHS has in delivering the amount of mental health services that we need?

20:42and secondly is that the case globally do you think and is that through companies like yours i'm obviously biased but i do spend a lot of my time thinking about this and i believe so the reason i'm so excited about the next few years is because for the first time we do really seem to have a credible path to scaling high quality care um and as you pointed out the digital revolution you know like 1.0 2.0 whichever one it was um teletherapy was great it was massively convenient and it helped you to connect clinical supply to patient need um more efficiently, more effectively, because now an individual didn't need to get to a real brick and mortar clinic.

21:38But what it didn't do is it didn't solve the underlying supply constraint. There was the same number, and there are the same number of trained mental health professionals alive on the planet. And teletherapy didn't change that. And we already know that there are insufficient number of clinicians to serve the massive patient need due to disorder prevalence in mental illness you know depression and anxiety are the leading cause of disability worldwide right if you're looking for quality of life um and you know adjusted years in that sense uh i i find it hard to find an area of health care that has even more broad impact and so with this intractable workforce supply issue.

22:27Teletherapy didn't solve that. But AI might. I really believe it will. Because what we are seeing is these clinically rigorous, evidence-based AI agents like Limbic are now essentially creating the final tier in the staffing pyramid where you had psychiatrists, psychologists, therapists, licensed clinical social workers, the entire gamut of different clinician roles. But AI now can be that infinitely scalable workforce. AI represents an autonomous member of the care team, which means that we now have a way to scale the clinical workforce. If we had psychiatrists, psychologists, CBT therapists, trained licensed clinical social workers, whatever the clinician role, limbic and AI agents represent the opportunity to have an infinitely scalable workforce of AI agents sitting beneath those trained clinicians.

23:43And that is a way to 100x or 1000x the human clinicians and their expertise in a way that teletherapy just didn't. You know, a video call didn't 1000x clinical expertise and clinical reach. AI agents that do autonomous delivery overseen by trained clinicians, that does have the potential to 1000x care. And if you do it in a way where, as we're seeing in our peer-reviewed papers at Limbic, If you do it in a way where you can observe improved access, reduced wait times, improved experience and critically improved patient outcomes, then you really are scaling quality as well as clinical reach. And that's the key thing.

24:31So I have very little doubt that AI represents the only credible path to scaling high quality care. And mental health and mental health care is such a great specialist area in which to sort of be the crucible for this new way of delivering treatment. because unlike many other healthcare domains, therapy, psychological therapies, they require hours of clinician time because the conversation is the treatment vehicle. And so compared to maybe a GP visit that's 10 minutes, a CBT session could be 60 minutes. And so every hour of clinician time becomes the atomic unit of care that cannot be scaled and ai really gives us a way where we can break this constraint and bend the cost curve on delivering quality outcomes does the specialty need to change in light of this new technology is there an adaption that has to work the other way around so obviously we're talking about treatment now and obviously this form of treatment with limbic in between the appointments and so many positives that are coming from the increase in signal and then processing it's more sticky so we're getting more of everything so everything leads to higher quality better care is there like i see this with the scribes because the scribes have to have to do a lot of research on this as to how does this change the consultation because now we shouldn't just be doing the same thing but digitized because that's not the optimum because that just assumes that we got it right first time it's obviously there's obviously going to be an adaption here do you feel it the more that you look at treatment and what limbic's doing that there's going to be some adaption in the way that the specialty evolves or specialties plural evolve i think there will i don't i don't think there needs to be we could hypothetically stick with the current treatment modalities and treatment protocols and just have ai amplify them that that That is a possible scenario.

26:52But there is an opportunity to go further. And that's what gets me very excited. Because one thing AI is very good at is integrating multiple different data sets and experiences that no human could ever have in one lifetime. And combine that in a statistically meaningful way to see patterns and to evolve. That's what AI is just great at. That's one of the things it's designed for. And so what I see as a huge opportunity over the next five to 10 years is AI agents like Limbic working as part of a care team, working with patients and also providing clinical decision support to clinicians and using this position to develop new modes of treatment, new models of care delivery based on the experiences that it is having at the point of care.

27:50and seeing who recovers faster and who doesn't. And it can then abstract that into brand new models that humans have never conceived of before. So I actually see the phasing as AI will initially scale our current models of care to reach everyone everywhere, which is already a huge win to society. And then it will go one step further and we'll begin devising new, highly effective ways of delivering treatment. You say that you've started thinking about this already. It's not surprising given the level of scale that you have. You don't necessarily have to answer this question, but are you starting to see those things across the level of data that you have?

28:33Are you seeing specific opportunity for some population level impact here or intervention here? We are. And it's an active area of research from our team of 10 PhDs in AI and psychiatry. It's not a quick one to answer, unfortunately. There's no single soundbite. We use things like reinforcement learning, where your AI seeks to update its own parameter weights according to some downstream reward signal that it seeks to optimize. So if you make that reward signal something like patient recovery, then it begins to adapt itself, but it does so in an abstract way, which is empirically effective, but doesn't necessarily provide a clear answer to what it's doing differently, if that makes sense.

29:29I think it's a very exciting time to be alive. I think it's a very exciting time to be in this field because I have no doubt that healthcare specialties like psychological therapies, but other domains like primary care and musculoskeletal care, sort of like physiotherapy, almost all areas of medicine will see AI develop on the traditional models of care. and they will develop in the direction of improved patient outcomes if we set the problem up as i described where the downstream reward signal is something like patient outcomes can we talk about economics ross so what's this this put me in the position of someone that buys limbic and what they end up feeling here because what what we're talking about what we talked about a lot in this conversation over two parts there are just so many it feels like there's just so many areas of real stepwise improvements in efficiency this even simply the stickiness of patients seeing the programs through end-to-end increasing because llms are used must then lead to obviously increased adherence better results and therefore they bounce back less to the system and all this sort of stuff and that's just one small component of everything that we've talked about here let alone the quality of the diagnosis the quality of the actual treatment and the factor optimizing for all of that with human plus ai so what does some what what does a service who brings limbic into it notice well in terms of um when we uh provide our ai agents to provide a services like nhs talking therapies services what we're solving for them is uh primarily workforce capacity issues they don't have enough staff to be able to deliver the the care that they are contracted to deliver um or if they do then there are downstream problems around the waiting list and and areas that they're looking to optimize for so what limbic does is we come in and we say hey look at intake, at triage, at assessment, and even at treatment delivery, we are going to do a lot of the heavy lift.

31:57And as a result, there will be more clinical hours in your workforce to allocate to other areas. That's the primary way that this would be positioned as we offer this to provider organizations. What's exciting, though, is that this is really just the foundational step. once we are in and we are we are supporting in this way then limbic as a clinical ai experts as you know i think is reasonable to describe ourselves as paired with clinical care delivery experts who are these nhs provider organizations these clinicians these services we work with in the united states we then have an opportunity to work together on true pathway transformation and that's where the models begin to evolve and we start to take the data that we are seeing and we start looking for the next low-hanging fruit so that we can really first 10x then 100x and then 1000x the clinical workforce without compromising on patient outcomes have you had to be quite adaptive with your business model globally because what works for the nhs certainly doesn't work for the US or vice versa so how how do you think about that with licensing and cost per well whichever unit we're going to choose here um on whichever side of the table how do you how do you even go about figuring that out I mean look this is non-trivial and I've long been of the opinion that it's not just a technical innovation that is going to define the winner.

33:35It is also business model innovation and truly understanding the financial incentives within the market that you're going into. Because at the end of the day, it is operating as a business. And we can't be blind to the flow of pounds or dollars. So you're absolutely right. The UK is different from the US. One thing that is worth pointing out about the UK is that what we call healthcare, the US describes as value-based care. And we tend not to talk in these terms because it's just healthcare. But what's very important and very helpful about the United Kingdom, and particularly the NHS, is that more often than not, the provider of care works with the payer of care, and those incentives are aligned.

34:30So the provider is incentivized to deliver quality outcomes at the lowest possible cost. And the payer also seeks to allocate their spend to good outcomes at low cost. Now, in the United States, things are a bit different. We have a system or a structure often referred to as fee for service. Now, the provider of care can be wholly independent from the payer of care, the health insurance company. And so you end up with the scenario where the provider of care is incentivized to do services because then they just send a bill to the payer. But the payer would like high quality care at the lowest possible cost.

35:13But maybe the provider organization, due to the way the system is set up, it's no one's fault. But the provider organization actually maybe wants to run a number of different services and they're going to get paid regardless. So there often isn't even a financial reward for getting a recovery in fewer sessions in the United States. So that alignment that you see in the NHS, which is different from many other health care markets, actually makes the NHS a phenomenal place to launch clinical AI agents. Because they are incentivized to deliver a solution that can act in the way we've described. whereas in the United States it's a little bit more complicated and we've managed to do it at Limbit we're now live serving patients in 13 US states so we are scaling out there but it's it's its own financial modeling challenge.

36:09So Ross just starting to wrap up here I want to make sure that we've captured what's most important on this treatment side I know that when we talked about diagnosis we went into so much detail around the limbic layer perhaps you want to start there and just what what is most important to consider when thinking about treatment for a an AI system in the mental health space how do you do this differently what is most important and for people looking at this as a potential solution for systems looking at this as a potential solution what makes the most difference for you as you build limbic here what are your main differentiators and what's most important to you in a phrase clinical rigor we always wanted limbic to be a clinically rigorous evidence-based solution for healthcare it's very easy to put a thin user interface on top of a chat gpt and call it a healthcare solution but it's very dangerous and healthcare organizations be it the insurer or the provider they are wise to these superficial solutions that belong more in a in the app store in a consumer app wellness space so for us we mentioned the limbic layer it's this clinical reasoning system that's where all our research efforts go and it's about making sure that much like we did with diagnostics on the treatment delivery the ai is pulling from a validated evidence base it has specific models designed to capture clinically relevant patterns in the user interaction and then surface this in a way that actually augments care but the tech stack aside that's how we do it but the most important thing is that at the end of it you have peer-reviewed clinical studies demonstrating outcomes you have regulatory oversight and accreditations on information governance, data protection, quality assurance.

38:20And you've got widespread proof within a care setting. It is not about taking a wellness solution with millions of users and saying that that is proof within a healthcare setting. These are different industries. And so as we move forward, both with the diagnostic support and the treatment support, it is all about making sure that we have the largest clinical evidence base, that we are the gold standard for regulatory approvals, and that every single thing we do is geared towards integrating within care flows and augmenting and amplifying existing care provision so that this solution really does embed within mental health care delivery itself.

39:06And yes, the limbic layer is how we do all of that from a technical perspective, but the proof comes from everything after that. Absolutely. Well, listen, Ross, it's been an absolute pleasure. I am not conflicted and therefore I can say this, but I've spoken to a lot of people that do this. I've been on this quest. I've done keynotes on this stuff. And Limbic does this differently. I can only say that to be perfectly honest. I've never heard the level of technical and clinical detail that, and you, and you do more than even what we've talked about on this podcast from times I've spoken to you previously as well.

39:45But I think it's very, very, very important that anyone that's looking to build AI into healthcare approaches it with the same level of rigor and respect from the clinical side, from the technical side in order to deliver what you can deliver. And the proof is in the pudding. that clinical study the fact that minorities are getting a better experience and outcomes this is all due to the level of thought and detail that's gone into the way that this has been approached and this it really is a call to action for people listening here that this is what is required to ross's point this is not about putting a thin layer over anything and expecting quick wins and quick results this is health care this is different and ross i really appreciate the level you've gone into um i think this will help a lot of people building this stuff i think it will help a lot of people critically appraising this stuff to bring it into their organizations and i think for those of us that are patients as well this gives us a huge amount of hope that these volume issues these complexity issues and these quality issues are being addressed with the human and the AI models that you're building.

41:00So I really appreciate it. Hey everyone, thanks for listening and making it all the way to the end of this episode. Remember to subscribe, rate us and leave a review. And you can head to the description of this episode to follow me on all of my social media so you don't miss out on any of the latest health tech content.