AI Therapy and Going Viral

I started my academic life at Cambridge studying natural sciences. It's a bit like the American system — you kind of major in one area, and that area for me was neuroscience. I was fascinated by how brains work, how people think, what self-cognition is. I finished my undergrad knowing very little, but keen to learn more.

At the time I thought the solutions to these problems were going to lie in a mathematical approach. The days of putting on a white coat and cutting up brains — there's still very much a place for it, but the emerging field of computational neuroscience was looking quite interesting to me. So I came to UCL and did a postgraduate master's in mathematical modelling — upskilling, learning to code, learning to think from an engineering perspective about constructing mathematical models. Then I married those two things together, neuroscience and mathematical modelling, during my PhD in computational neuroscience, also at UCL.

There was no master plan here. I was following what I was interested in, in the moment, and I was very fortunate that I ended up doing something I really loved. When I finished my PhD I thought, right, what next? There were some postdoc opportunities I was keen to explore, but the thing I'd enjoyed most during my PhD was the immediate impact of my work. And I was beginning to think that academia was losing its monopoly on where all the cool research and breakthroughs could happen — some notable companies like DeepMind had come out of UCL. Actually, companies in the right space could be the ones at the leading edge.

So I threw my hat in the ring and started a company. I started it on Entrepreneur First — I don't know if you're familiar with the programme. For the benefit of your listeners, it's an accelerator that was based here in London and has now spread to different parts of the world — sort of the Y Combinator of the UK at the time. Matt Clifford and Alice Bentinck were great founders leading the way. They spoke to me, and I was really interested in the idea of being supported at idea stage to build a company, and being matched with a co-founder who'd share my passion. I joined the programme in 2017 and met my co-founder, Sebastiaan de Vries, a Dutch programmer. I met him in week one and we've never looked back — we've been working together daily for the last three and a bit years now. That was where Limbic was born, really, on the Entrepreneur First programme.

The story was that it was all story. It was a tongue-in-cheek project, definitely not a business. A really good friend of mine during my undergrad, Ed Moyse, lived next door to me, and we were always scheming on different things — we had a very bad music record label going at one point. We came to the end of our undergrad and didn't really want to just go and apply for a job. We didn't know what we wanted to do. The Deloittes and PwCs were coming and setting up exhibitions to attract applicants, and that wasn't really something either of us was interested in.

Around 2004 we came across the Million Dollar Homepage. Do you know of it? It was a guy, Alex Tew — who's actually now the co-founder of Calm. He bought a domain, hosted a webpage, divided it into a million pixels, and sold each pixel for $1. The idea was that you could advertise on his website. Now, it's worth nothing if nobody's visiting it — but the more people talk about it, the more of a story gets generated, and the more value each pixel has, because it's getting more visitors. Because it's getting more value and making more money, it becomes more of a story, and it upward-spirals.

I thought this was such an interesting systems approach to solving a problem: you've got nothing of value, and you turn it into something incredibly valuable just through an upward spiral and a story, making sure there's this positive feedback. So my friend and I took the exact same principle and said, well, how about this — we'll sell advertising space on our faces. We'll make it purposefully weird and wacky. We didn't know how to face paint, but we bought some face paint and gave it our best go, painting a logo onto each other's cheeks.

Each day was a different logo on our faces — like an online calendar. We went to local newspapers and said, "Hey, this is weird, isn't it? What do you think?" And they said, "Yeah, it's funny, we'll ask you some questions and run the story." That meant local businesses said, "I'll give you £5 for a day." We were slowly increasing the price of each day, so there was an incentive to get in early when it was only six, seven, eight, nine pounds. It worked — nowhere near as successful as the Million Dollar Homepage, it was a secondary attempt, but the goal was to pay off our student debt, and we managed to do that. We ended up getting sponsored by Ernst & Young and some big law firms.

The fun part was that the companies who understood what was happening realised that the more they contributed to our story, the more advertising they'd get. It's not just about visitors to the website — it's about the newspapers running the story and telling the world what we'd been up to. Very early on, companies like Ernst & Young clocked that if they bought the flat rate for the day but then also paid for us to go skiing, or sent us to a Michelin-star restaurant, they got way more advertising out of it. So when The Times did a big two-page spread on this ludicrous idea and asked "what have you been up to?", we'd say, "Well, we were on our way to France to go skiing, because Ernst & Young paid for that." Marketing isn't a core competency of mine — it's not something I'm educated in or particularly passionate about — but I did like the interesting way of solving a problem. You've got nothing of value and you need to pay off your student debt, so what do you do? It was a fun one.

Really good question. Not explicitly. Marketing is actually something Limbic isn't really doing much around — we're not shouting very loudly about what we're doing. Part of that comes from the fact that health tech is sensitive and it deserves respect. The data we work with is very sensitive, and the patients we're there to help are very vulnerable. I don't think it's appropriate to be pushing virality — it's not the same mindset we were in with buymyface.

What is interesting, though, when you think about virality, is how you make the product something people genuinely want to use. A lot of health tech companies fall short here. They theorise around what they're going to build and why it's going to be useful, then they deploy — but maybe they haven't had much patient or user input during the design process, and when they deploy, no one actually wants to use it. The clinicians say, "Nah, it's too complicated, too much of a mental hurdle." The patients say, "Nah, it's annoying, I don't like it." Suddenly all the value that was supposed to come from the product never gets realised, because nobody uses it. That happens a lot in health tech. So when it comes to virality, for us it's not about the marketing or PR — it's about how we instil positive feedback loops in the product. How can we make it something patients and clinicians build habit loops around? That's something we're very focused on.

I think clinical recommendations carry a lot of weight in healthcare. A peer-to-peer recommendation is worth a lot compared to Facebook advertising. So if you're going to grow and get the word out and have it received in the right way, you really want those clinical champions — clinicians telling their colleagues at conferences and by email that this tool is really good and they should take it seriously. That's how you spread and enter services.

Patient case studies are important, because ultimately patients are the end users and the whole system is set up to help them. But if you've got a value proposition that also helps and supports the clinicians, it can be very powerful, because then they talk to their colleagues about how useful this tool is for them personally. It's front of mind because it's bringing value to them. That's definitely something to build into a health tech product if you can.

Sure. Why don't I start high-level, and then you can nudge me with specific questions and we can go down different rabbit holes. At a high level, our mission statement, so to speak, is to revolutionise psychological therapy — and we want to do this with the latest state-of-the-art technology. So data science, machine learning, and just beautiful product design.

In mental health care there are so many products you can use, particularly B2C wellbeing apps — the app store is inundated with them. A lot of these would loosely fall under the category of a digital therapeutic: a digital tool designed to directly alleviate symptoms through some guided exercise or connected content. They're taking a specific route — digitising existing content like guided meditation or CBT exercises, and then trying to build an evidence base that their specific repertoire of content is helpful. Some are doing really well, and I don't want to diminish those efforts; there's definitely a place for it.

But at Limbic we believe the way to really help patients is to amplify the effectiveness of clinicians. Our tool does face patients, and it is there to alleviate symptoms and help in multiple different parts of the patient journey through the care pathway — but we're always linking back to how we can take the clinical expertise that exists in the system, increase it tenfold, spread it further, and give it better impact. That leads us to our software tool, which is an AI therapy assistant. From the moment you enter care to when you leave and beyond, Limbic is a little digital assistant that helps both patients and clinicians.

At the front end we've got Limbic Access, where our AI therapy assistant lives as a web chatbot. It's a conversational AI, and it helps with referrals into NHS services. Specifically, the services we're looking at right now are called Improving Access to Psychological Therapies, or IAPT — the NHS talk-therapy, primary and secondary care for mental illness.

At the moment you've got a huge influx of patients seeking help with common mental illness and requiring talk therapy, and every year that's been increasing. Eight years ago it was 800,000; today it's 1.7 million being referred in. But while you've seen that steady increase in patients, service capacity has remained largely the same. So fundamentally there's a supply-demand mismatch. Limbic Access — our flagship product, the first thing we built — sits at the front end and tries to improve service capacity by supporting the referral and the initial assessment.

At the moment, what would happen is somebody self-refers into the service, and then a clinician needs to call them and spend an hour on the phone collecting basic patient information and the outcomes to different clinical questionnaires — the PHQ-9, the GAD-7, these sorts of things. Limbic says, look, clinical expertise could be spent focusing on clearing the wait list and treating. So why don't you let us take the lion's share of that work — making life easier for the patient entering the service, GP lookup, little tools built in — and we'll ask for the patient data and the clinical outcome questionnaires and signpost to the correct care level.

This is a very simple thing, but it has big impact, because 25% of the IAPT budget is spent on that clinical assessment stage. By supporting it, you free up resources to tackle this growing wait list, which is a huge problem, and devote more time to the stuff that really can't be automated right now — like CBT treatment. Does that make sense?

We're working that out day by day, following where the huge pain points are in the pathway — what can we help with today, and what can we help with tomorrow. You've hit the nail on the head: there are aspects of cognitive behavioural therapy and psychological therapy which are fundamentally human. To go at this thinking "we're going to automate clinicians" is arrogant, naive and short-sighted. What we're focused on is amplifying the effectiveness of clinicians.

To give you another concrete example: Limbic Access helps with entering care, and then that AI therapy assistant moves to a mobile app. Patients entering the service and on the wait list have this mobile app, which keeps collecting patient information — mood journaling, negative thought diaries — and provides some validated CBT tools and techniques in the moment. So it's a little bit like a digital therapeutic in that sense.

But then, when the clinician is assigned, we're still there — and with this idea of amplifying the effectiveness of the therapy, we let the clinician coach Limbic on how to deal with their patient out in the real world between sessions. Limbic is a support tool that's always there for the patient. The patient goes in once a week for their one-hour session, and the clinician can say, "Look, when Jenny is feeling overwhelmed out in the real world, tell her to step out onto a balcony." The clinician has this very human relationship — Jenny has mentioned there's a train track outside her window, so the clinician can tell Limbic, "Remind Jenny to just watch the trains go by for a bit," because that's effectively a grounding exercise. But rather than a generic grounding exercise for everybody, this one is really personalised to Jenny — and it can be, because the clinician knows Jenny. By telling Limbic how they would respond to Jenny in the real world, the clinician takes that clinical expertise and personal relationship and spreads it, keeping it front of mind between those one-hour sessions. So again — we're not automating the therapist, we're trying to give them superpowers.

Good question. Fundraising is always hard, but relative to some areas, mental health is getting quite a lot of attention right now from health tech investors — so it would be unfair to say it's particularly hard to fundraise in mental health. Regulatory is an important part of innovation across the whole healthcare space, so again it's not really mental-health-specific; the problems we have to navigate are largely similar between different healthcare sectors.

Confidentiality, though — I think this is an interesting one. Whenever you're processing patient data, you need to overcome the barrier of: why would the patient agree to share this data with you? And quite right — there should always be a reason to share data. You share data in order to get a functionality you want and that's going to help you. That's the transaction, that's how it should work. Sharing data should only ever be to achieve that function, never just a blanket "yeah, you can have my data and good luck."

In the mental health space — maybe due to the historical stigma, which is breaking down but probably still exists — health data is always sensitive, but mental health data occupies a particular place in terms of people feeling guarded around it. So when communicating with patient users, you need to be crystal clear and give them a very good reason to share information with you at all. This was something we had to navigate during our design phase. We assembled a panel of patients to input while we were scoping out the product requirements, and this idea came up time and time again: "What are you going to do with my data? Where's it going to go? Who's going to see this? I'm going to share some really personal feelings here — I want to know."

What was helpful for us was that, because Limbic sits between a patient and a clinician, there's that clinician in the loop. We're able to demonstrate that we're part of the care pathway and here to facilitate it — and really, sharing the information is only ever something you'd share with your clinician anyway. That's a trusting relationship you've built, and we're going to piggyback on that trusting relationship. That's how we overcame it with Limbic, but sensitivity around this information is quite rightly very high, and we had to navigate it with care.

For sure. Your B2C play would be consumer wellbeing, usually. You're basically saying this isn't healthcare, this is wellbeing — we're going to make it available in the app store, anybody can download it. You're really targeting what are often called the "worried well." That space is really important, because the majority of people don't reach clinical thresholds of mental illness, but they're somewhere on a spectrum, and keeping those people healthy adds huge value. The problem is that the wellbeing space doesn't have healthcare-level regulation — you don't need validation in the same way, and anyone can create a wellbeing product and launch it. That's why there are about 10,000 in the app store. Consumers are drowning in a sea of options, and it's warfare trying to get your wellbeing product seen above everyone else's. Big companies like Calm and Headspace can throw big budgets at marketing and stay ahead of the competition — classic B2C market dynamics.

To be honest, working with the NHS is also B2B, but the corporate wellbeing setting is a really interesting one. A lot of companies have seen success there. It's usually wellbeing rather than validated healthcare, but they need a bit more evidence base and to tighten up, because the employer is going to ask, "What's the risk here? Is this validated?" The value prop is that, from an employer's perspective, for every pound spent on employee mental health they make a number of pounds back — it's very costly for employees to require treatment and miss work days, essentially lost productivity. There's been economic analysis showing it's very costly to an employer to have poor mental health in the workplace. Obviously there's a social component too, which you'd like to think is also front of mind, but even on pure economics there's a reason to assign budget. So you've got companies like Unmind — they're really good, doing a lot of great work — and BioBeats was another. And Calm and Headspace have both models: the consumer model, and the corporate wellbeing model where you make these tools available to employees as wellbeing perks.

The healthcare setting is quite different from those two, because now you've got healthcare regulation. Your users aren't the worried well — they're patients who've met clinical thresholds and require intervention. So you need to navigate the care pathway, understand how care is delivered, and fit in with it. You need to integrate with central NHS systems and widely-used third-party systems, which isn't really the case on the wellbeing side. Validated, evidence-based therapies are critical, and you need to be very careful about what you claim to be doing — and indeed what you're doing — because medical device regulation kicks in. You can release a product in a healthcare setting and not be CE-certified, but only if you're not altering the clinical pathway or impacting clinical decisions in a big way. So you'd very much need to navigate that MDR space.

Yeah, you're totally right. The NHS is a very daunting customer — this big beast, so hard to enter. There's not one front door; there are hundreds of front doors. Even people within this gigantic system don't understand the process. You ask them who the decision maker is and they don't know — they work there and they don't know. You're speaking to somebody very senior and they don't know what approvals they need internally. I love the NHS, but it's a very big system and hard to manoeuvre.

"The NHS is where startups go to die" is probably a fair comment, because a lot have. But if — and this is a big if — you can get in and gain that bit of traction, the NHS is a fantastic partner. Firstly, the reputation carries at a global level. If you have ambitions outside the UK, which Limbic certainly does, that NHS seal of approval carries a lot of weight in healthcare systems in different parts of the world. Secondly, the NHS is a very sticky customer, so that inertia — the difficulty of moving them — can act in your favour once you're in, because they don't switch providers regularly. To be honest, that's probably given rise to a lot of NHS providers being a little complacent, because the NHS is a reliable customer and there's less drive to stay on top of the product. So that's something to bear in mind. But as a startup it's definitely worth going after if you think you've got a way in. It's a risk, though — you're gambling.

With entrepreneurship, I personally didn't get on well with a lot of the books, because it's a really hard thing to distil down. There's no rule book. You often come away with a collection of anecdotal stories, but the author wasn't able to extract a common rule, and the reader won't be able to either — because there really is no common rule. So I found entrepreneurship books interesting, but not that informative.

In terms of staying up to date and being able to pick out important themes, I really like a newsletter by Matt Clifford, the co-founder of Entrepreneur First. He sends out Thoughts in Between once a week — a great way to stay abreast, in five minutes, of big changes happening in the startup space and interesting perspectives. They're usually quite balanced, so I'd definitely promote that.

On computational neuroscience and the academic side, Theoretical Neuroscience by Peter Dayan and Larry Abbott. Peter Dayan was my PhD supervisor, and his colleague Larry Abbott — they wrote Theoretical Neuroscience, which is the bible of comp neuro. Anybody interested in this space should have that book; I'm looking at it now on my bookshelf. And Kevin Murphy's book on probabilistic machine learning I really like. My school of thought with machine learning — my interests lie in probabilistic approaches: not just being able to output a prediction, but a probability distribution over predictions. I think that's really important, particularly in healthcare, where confidence in a prediction is a key part of the decision-making process. It's not enough to say "is this picture a cat or a dog?" — I want to know how likely my prediction is to be wrong. So probabilistic methods in machine learning, and Kevin Murphy's book on that, are fantastic.