Billion $ Health Ideas: Generative AI, DTC Weight Loss + Cross Border Care

So, first idea: multi-level summarisation of research using AI. Research is being published at a ridiculous rate — there are millions of papers every year, and it's super hard for anyone to keep up to date with everything, including active researchers in the field. And AI is just getting better and better. We've got ChatGPT now, which a lot of people have become aware of recently, and it's actually very good — if you give it some text and ask it to summarise, it does a pretty good job. There are caveats: sometimes it just bullshits and makes stuff up, so you have to be a bit wary. But ultimately we're definitely moving in the direction of being able to summarise things well, and a very obvious application area is academic papers — pulling the actual insights out of a massive amount of content.

The second level on top of that is the multi-level aspect: summarising things at different levels of complexity, because people come at this with different levels of understanding. A researcher might want a lot of detail; someone else just wants the high-level takeaway. So if you could train an AI to give tailored levels to different people — that's the initial high-level idea.

In terms of who it's useful for, I think there are a few groups. The first is actual researchers who want to keep up to date with their field and struggle because there's so much stuff — they just want to save time. The second group is people who are interested in science-based things: "what's the evidence behind my diet?" There's cool stuff coming out on diets, or on longevity — some people just want to know what they can do now to live longer, and they don't want to go through the research papers themselves. They've got other jobs, they're busy, but they love the quick takeaways. That's a little bit like Blinkist, the thing that summarises books.

And, actually — you said the Andrew Huberman podcast, which you're probably aware of. That's become super popular in a short space of time, and I think it shows the appetite for this kind of summarisation of the science in a condensed format, for a great audience of people who are just interested in general. Then there's probably a third group who just want the real nugget takeaways in a quick, digestible format — kind of like how you check the news. You're relying less on the media to go through and find these things and present them to you; you're using AI to do that.

So, those three broad groups. The second question is the tricky one: who would pay for it?

Researchers, maybe — although researchers generally aren't super keen on paying for stuff. Maybe you could get an institution to pay, the same way institutions get licences for academic papers. Could you get an institution to buy some kind of licence for the summaries too? Potentially, yes, if you could show it improves research output. Or maybe a Blinkist-type model, where people pay a subscription to get summaries of academia rather than books. It's a similar sort of product.

Those are your initial thoughts — what do you think? Who could you see paying for this?

No, I think it definitely needs to focus on a specific end output — and then who is interested in that output, and would they pay for it. That's the framing I'd take. So maybe the end output is some sort of summary of research trends, and maybe VCs are interested in that — that's one potential angle.

But you'd ultimately have to validate with these different groups whether that output is interesting to them. And there's always the question of what the advantage is of using AI to generate it — because something like the State of AI report is, I assume, mostly if not all human-generated, and it's popular. So what's the advantage of using a generative model to make the summaries?

But are we heading in that direction? I think so. There are some technical improvements that would be natural next steps, but ChatGPT has opened my eyes a little to how good things could get. Maybe you could fine-tune a GPT-like model into something more oriented around summarisation.

On using ChatGPT for summarising scientific research — you know that quote, that a human and a banana share 90% of their DNA? I don't actually know if it's true, but it sounds legit. A lot of the output I've seen is like that: it's 90% there, but that last 10% is almost the most important 10%. It's the difference between a really crappy writer and a really excellent writer, and that's where all the quality difference is. The one thing that might be useful with these language models is just the speed at which they operate. I have a lot of friends in consulting, law and VC, and they spend a lot of time parsing through reports and research, trying to find things to present to other people.

I wonder if almost a Grammarly-type add-on in your browser, that quickly summarised a 30-page PDF and gave you a nice multi-level summary — exactly what you're describing — would be more useful.

Because the second thing is, if you're using this in any serious capacity — healthcare, research, commerce — you can't just trust the summaries; you need to validate them yourself. So I'm less keen on things that produce a complete output, versus things that just make it easier to get where you need to go.

There's actually a whole other area — the applications of this kind of standardisation in healthcare, outside the research context. Should we go into that?

There's a bunch of stuff we do as doctors — similar to law and other fields — that's basically parsing text and extracting key points. In healthcare there's a bit of an acute issue in that our digital systems often aren't great. You have loads of letters, referral letters, discharge summaries — things you have to find or write that are quite time-consuming and could be done in a better way.

To some extent, some hospitals are already doing a better job of automatically populating discharge summaries — I worked at a hospital in London that had that. But you could leapfrog the need to set up that infrastructure and make it more of an AI-driven summarisation that works with quite diverse inputs. Because you could imagine —

Exactly. And discharge summaries are often done by quite junior members of the team, they're time-consuming, they eat up a lot of your time — but they're really important, because that's essentially the API between the hospital and the GP and future services. That's the summary of your stay, and patients use it as a reference too.

It makes complete sense: if you're recording all the information during the stay — which in medicine you have to, because by definition everything that happens should be recorded — then the information is already there. It just needs to be put into the right format. And that doesn't really need a human in the loop. So this could be one of those cases where you pass in all the information collected during the stay, create a summary, and then a human just checks it and maybe tweaks it.

The other example that's been talked about for at least a decade is having an Alexa-type device sitting in on clinics. You have a 30-minute appointment, the device records and transcribes and then automatically summarises everything that was said. Because right now a doctor will have a clinic, spend 20–30 minutes with a patient, and then dictate or write up a letter back to the primary care doctor summarising what happened and any actions. That's time-consuming, especially for older doctors who are less savvy with typing. So that's another one — but I don't really know why it hasn't been done. I hear it talked about a lot.

As a patient, I'd love it — then there's a record. You see this with the telehealth apps that do remote consultations: they offer you a recording of the consult that you can play again. As a patient I'd quite enjoy that. As a doctor, I really wouldn't like it. I'd compare it to security guards, bouncers or police having body cams on all the time, recording everything they do. I'd hate that. So I think there'd be more pushback from the medical side than the patient side.

One of my limitations with the automated summaries — and I don't know a lot about this — but when I was using ChatGPT, I asked it some questions. One was, "how do you get rich as a doctor?" It actually gave a pretty good summary: basically, you're unlikely to get rich as a doctor, you need to start a business. It was pretty good. But when I looked at its answers in general, they'd be maybe 500 words long with, I'd say, about 100 words of actual content, and a lot of fluff — like when you're writing an essay at school and have to hit the word count. So I wonder whether the earlier versions of these automated summarisers will actually create more of a headache for the doctor receiving the communication, where it's AI-generated, long and unclear. It's a bit of a human skill, at least right now, to write two or three lines that are really meaningful and summarise everything. I don't think we're there yet.

So there's a bit of a spectrum. At one end you have a full transcript of the entire conversation, which for a 10-minute appointment isn't going to be that helpful or passable to a doctor. Somewhere in the middle you have these ChatGPT-style summaries — the points are there, but with a lot of fluff around them. And then the gold standard is a nicely structured thing with just the key salient points, in that format of history of presenting complaint, past medical history, progression of symptoms, the relevant information. We're definitely heading in this direction. I'm not aware of anything fully doing it to that extent yet, but it feels like broadly the technical capability is there now — we'd just need to fine-tune it for a medical context. There are tools like Amazon Comprehend Medical, which sells itself as something to pick out salient things from text — you give it a long paragraph and it picks out that this is a symptom that's been getting worse for three weeks, and this is a past medical diagnosis. But I've not seen anything yet at that far end of the really high-yield output that would work clinically.

There's definitely some truth to that. One angle is to offer some concrete service of which this is a part. So let's take referral letters — a letter that says, "this is my patient, I'd love you to see them, this is why," requesting an appointment. Maybe you create a tool that generates that, someone uses it at the point of making the referral, and that becomes a standard thing you want incorporated by the existing software providers — so it's an acquisition play.

But I also wonder if you could be the wider software provider that services the whole pathway, of which your tool is just one component — you're actually doing the whole thing. If you can offer something concrete that covers part of the pathway, of which this tool is one component, that's a potential way it could become bigger.

I don't have direct experience interfacing with them from an API point of view — I have experience getting data out of them for research. But one interesting approach, which I spoke to someone about recently, is that you can get away without necessarily needing the API interface. My general impression is that they're not super easy to interface with. But you can have macros at the computer level that operate over the system without the system knowing you're querying it. You create a program that moves your mouse, clicks on the patient record, goes into the file and extracts it. This guy I was talking to — his company basically does this for some GP services, automating their processes. They run virtual computers with a login for EMIS or SystmOne, the two main GP systems, and it literally just clicks, gets the information, and automates processes around that — maybe even fires off a letter.

There are pros and cons. The pro is you're not dependent on them having an API endpoint or officially collaborating with you. But you're also vulnerable to changes in their software — if they tweak the flow, you have to rewrite the whole macro, because the previous one just doesn't work anymore. So it's a hacky way around it.

In general those kinds of businesses can be strong, because you're building shovels rather than trying to dig any particular hole. When everybody's trying to access health data from different places and build different services, if you can build the shovel everybody wants to use, you're in a very good position. And interestingly — there's Vital, but I've seen at least two other companies out of YC doing pretty much exactly the same thing. So a few people are having a similar idea at the same time, which makes sense — it's representative of the times: people are all thinking about this aggregation and combination of health data.

I'm generally quite bullish on these approaches — building genuinely needed infrastructure that enables other people to do stuff. Maybe it's a bit of a tangent, but I'm always thinking about what directions to go from a startup perspective, and when I was looking at what I could build to support machine learning in healthcare — one angle is a specific use case: say CT scans, where you build AI that analyses the CT and triages, or decides whether someone needs thrombolysis for a stroke. There are many, many specific applications like that. The other angle is: is there something I could do that enables more ML algorithms to be built? You could argue the enabler probably has a larger impact. Imagine there are 20 radiology AI startups with cool ideas that all make sense — but they all have one blocker. If you build that blocker away, you've enabled 20 companies to come to fruition, versus building one of the 20. So broadly I'm pretty bullish on enabler plays.

You need both, right? So it's good.

No, I think we've probably covered it.

I think I have seen it, but it's a very small percentage of cases. So many people try, and then if you look at the long term there won't be much change. There's a small minority who have a very big impact — and they're the ones in the adverts for a different diet, "look at this person, they did this" — but they're real outliers. I haven't seen the data, but I imagine it's less than a few percent who have some kind of transformative impact.

There are multiple facets here. So much of it is behavioural; there are physiological elements; there's an educational element of knowing what a good diet actually is — which to some extent still isn't defined, though there's been very interesting progress as a research field in the last five years. It's hard for many reasons. It's very personalised to each individual. And so much of it is behavioural — not superficial-behavioural in the way that, say, CBT makes superficial adjustments —

Yeah — CBT is typically a six-week program focused on, for example, helping with depression or anxiety, but at the level of "what are your thought patterns, what behaviours are you making," rather than going deep into why you see it this way. It's very effective for various conditions. But in weight loss, it feels like that's maybe not enough — you need to go deeper and more personalised on the behavioural side.

I wasn't aware of these trials, to be honest. It sounds good — if it's consistent across different people, yeah. Did they give any idea of the mechanism of action? I know you said they're both glutide-type things, which if I'm correct is something to do with one of the receptors — like glucose transporters or something.

Okay. If it works, it makes sense. So what's the value-add of a direct-to-consumer clinic versus going to a GP? I assume if it's cutting-edge it's not widely available from GPs — but what happens when GPs catch up and it becomes a commonly prescribed thing, assuming the evidence bears out?

I don't, but I imagine it's long. Ten weeks?

A few months?

It makes sense that there are lots of conditions affecting lots of people, so there's a big central market, and by going direct-to-consumer you go straight to the people willing to pay. One thing to be careful with in a healthcare context is the risk of overdiagnosing or being too profit-driven. With ADHD, there's concern that maybe too many people are being diagnosed — and it might be that diagnosing more people and putting them on more medication isn't actually in their best interest in the majority of cases. So there's a risk that you're building a business where, okay, you're helping people get their ADHD diagnosis and medication, but you have to be careful not to be too profit-driven, because you can end up with too many people on medications that might not be the right thing for them, even if they think they should be taking it.

One thing that's quite nice about the UK, from my experience, is that GPs are generally quite good at saying no to things the patient wants but that probably aren't the right thing for them. Part of that is because they don't have the same profit-incentive structure that, say, the US does. But as long as you can navigate that — if the waiting list is seven years, people want to get assessed, and you're assessing properly and not being unduly influenced by profit — then yeah, you're having a good impact.

Ultimately it's probably not impossible to overcome, but it's always a risk if your business is centred on that being the main profit-generating mechanism. Maybe you could make a different funding structure — people pay for the consultation, but there's no payment tied to the prescription. Basically decouple the financial incentive from the end output of the consultation, so your metrics are more about how many people you're seeing and whether you're assessing them fully.

No, you might be right.

Yeah. Which one do you want to do?

No, I think we're good. Let's do that one.

It's a super interesting problem, and it capitalises on a lot of interesting trends — an increasingly international world, cultural needs around caring for relatives, and the need to standardise healthcare in parts of the developing world, which is much less standardised than in the developed world. So I share the view that these are very exciting areas to play in, and I think people should build companies here.

There are some challenges that don't fully kill the idea. One is customer acquisition. There are a lot of Pakistanis living abroad, but if you go to any particular location abroad, the majority of people aren't overseas Pakistanis — so you're trying to find a needle in a haystack. A lot of the traditional, geography-based ways of reaching people struggle. The internet changes that a bit, because you can do more targeted approaches. But even then, there's a sub-segment: in somewhere like Bradford in the UK there'll be a lot of Pakistanis, but they also need to have an elderly relative back in Pakistan, and that relative also needs the services you're offering.

Startups usually focus on a few specific things and expand later. So say you pick one specific problem — you want GPs to check in on people with chronic conditions. You're going niche within a niche within a niche: people from Pakistan, living abroad, with a relative back in Pakistan, who also has a chronic condition. You're serving a small niche, which makes it harder for the VC-backed, fast-scale type companies. So I even wonder whether, if you were doing this, you'd not necessarily go the VC-backed route from the beginning — just build it from the ground up. If you can get good traction with that community and a way to engage the people who want the service, I think you could build a pretty nice business. And Pakistan is just one example — there are many countries with this diaspora and this cultural expectation to help look after relatives. Our generation has been described as the sandwich generation: we're looking after our kids and also our ageing relatives, so there are increasing pressures, and a lot of people work long hours and don't have as much time to check in on their parents as they'd want.

A few reasons. One is that it's harder to go the VC route, and I was very much looking at building a scale-fast, VC-backed company. There probably is a way to do it, but it feels less like one of those ideas and more like a ground-up business.

A few other elements. There's the customer-acquisition side, which is non-trivial. Another is the existing healthcare provision on the ground in Pakistan. We felt the place you'd add the most value — from the people we spoke to — was providing an almost GP, primary-care-type check-in with elderly relatives. With the clinic-referral stuff, systems already exist to a large extent — people will just pick up the phone and call the clinics — so there's not a strong pain point there. But getting someone to come and visit you at home, check up on your health and create a structured report makes a lot of sense.

The issue is that GP, as a specialty, isn't really a thing in Pakistan. There aren't trained GPs waiting for a job, or to do remote work or home visits. There's variation in the level of training different people have, and there isn't a clear-cut divide between doctors and non-doctors the way there is in the UK — "I'm a doctor" versus "I'm not." There are a lot of people with some general medical experience who might offer these services but aren't medically qualified. And likewise, you have people who are medically qualified but have only worked in hospitals and don't have GP experience.

When you look at other platforms offering this kind of thing, the main critique customers have is variability in the quality of the doctors and the ability to trust them. So if you're building here, you really need to create that trust — and that's key. If you're living abroad, arranging a doctor to look after your relative, you have to trust that doctor completely: you can't see them face-to-face, you're letting them into your family's home to examine your parents, and you're putting your name on that as the son or daughter arranging the care. But it's very hard to build a scalable company where you have that full trust, particularly if the pool of doctors you're tapping into isn't set up to provide this kind of care. So there's a transformation required — training people more around general practice, or a doctor-acquisition play to find the ones with the right experience, skill set and availability to deliver these visits.

So it's this challenge of matching things — and you add the complexity that, on the supply side, you also have geographical restrictions. You need someone in the UK with a relative who needs this care, but that relative might be in Lahore, and someone else's might be in Karachi — and you need someone on the ground to deliver those visits in each place, who also meets the requirements of being an appropriate doctor. So there are a lot of logistical challenges around providing this service and differentiating yourself on the level of trust and reputation of the doctors — because the company would live or die on that reputation, and it'd be very hard to really ensure you have it. Those are some of the main reasons we moved away from it. There's a lot of good in it, and maybe someone will come in and build something great — but there's a lot of weeds to get through around how you establish trust, verify the doctors, maybe even train them by partnering them with doctors abroad who have more GP experience, and then retain them.

Yeah, I agree. Nice.

Sure. The best place is my personal website, chrislovejoy.me. I talk about the different products I'm working on and link out to my socials there.