Being a Little Bit Obsessive

Thank you very much for having me. Good to start where I am now, I suppose. I'm a clinical research scientist in the Health AI group at Google — the artificial intelligence group — working on applying AI to healthcare, mostly focusing on radiology and ophthalmology, the two areas I've been working in. The nice thing about this job for me is that it brings together lots of different parts: clinical medicine, a love of technology, machine learning, research, building things, and a bit of business as well. So it really suits me.

I also work at St Thomas' as a senior clinical fellow on the neonatal intensive care unit, on top of the Google job, just to keep the clinical side going.

In terms of the journey, it's been a very wiggly route. As a child I was just obsessed with electrics and electronics — I spent most of my life buried in wires, making things like robots and a dishwasher for my mum, all this random stuff. And then I really wanted a computer. I actually got caught in the boys' toilets, age six, handing over my life savings in the form of a bag of 50p coins to another boy who had a Spectrum at his house. The headmaster was very unimpressed, and we both got told off. But that led to me getting a computer from my next-door neighbor, who worked for Xerox — I was about seven — and he showed me BASIC, this programming language, and it kicked off a love of coding.

From then on I was making lots of things on the computer. I had a fictitious software house called Kelly Software Productions, with the logos all around my bedroom. I made one called What's Up Doc, from a family medicine book we had at home — I had an algorithm for "should your chest pain go to hospital or not?" It asked you questions and you could say yes or no. A bit like some of the chatbots around today, actually.

Then, at school, I really wanted to make a website — the internet was kicking off. My English teacher gave us an essay to write in the Yorkshire dialect, writing the phonetics of the speech. So I made a program to do this: you'd type in a phrase and it would write it out in the phonetic way a Yorkshire person would say it, with all the Yorkshire vocab. And I thought, maybe this could be a good first website. So I did a few more — Scouse, Scottish, Brummie, Cockney, Irish, Geordie, Posh, and Ali G, an important one at the time — and put it all on a website called whoohoo.co.uk. That was hundreds and hundreds of hours of my life, age 14 or 15, going through dictionaries and poems and books just to get all the vocab.

And it became oddly popular. About half a million people a month were going to this website. It got into, I think, almost every national newspaper and most of the magazines. It was a really weird experience having all these people using the thing you'd made — a real adrenaline thing for me. Then I got contacted by these ad networks saying, can we put adverts on your website? So I put some adverts on, and before you knew it, it was making a couple of thousand pounds a month for me as a secondary-school student. It opened my mind to what the internet can do. At the time the internet was quite static — there wasn't much dynamic stuff — but it definitely led me to think: if you focus on doing what you're interested in, money will probably follow if you do it right. Even a stupid website that translates text into dialects.

When it came to university, I was really torn between engineering and medicine. I went to open days for both, and in the end I did work experience in medicine and thought, this is really good. So I applied to medicine, went to St John's College, Cambridge, and had a really good time.

But the website was still going, and I was still really interested in ads — how you get people to buy things online, or how you find the right people to buy the right things online. So in my first term of medicine, when I really should have been focusing on the medicine, I started my first startup: an ads company. It was inspired by the website. The website had originally been paid on cost per thousand views of the adverts, then it went to cost per click, and then to cost per sale — where you actually had to buy something. And, as you can imagine, if you're going to a dialect-translating website, you don't really want to buy anything. So the revenue dropped right off.

So I thought, well, how do I find people who do want to buy the things being advertised? At this point search engines were just starting to have advertising — Google Ads had launched two or three years earlier. So I thought, let's find people on search engines and send them to the right place to buy stuff. That's what the company did. We paid per click to send traffic to our advertisers, and they'd pay us if the customer made a sale. So the AA, for example, might pay us £30 if we got a customer to sign up. The system in between worked out the value of each keyword, each click. You'd know a click on "breakdown cover" cost 20p, and if it took ten people clicking before you got a sign-up, that sale cost you £2 — and you had a £28 profit.

So began this rollercoaster of trying to run a startup while doing the first year of university. It pretty much went bankrupt the day before my first-year exams, which was not good — I actually failed the vast majority of them. I was in a very stressed state. But ultimately it gained some momentum, and by the end of undergraduate I had about 25 customers, including Apple, Sky, Jessops and British Gas — lots of quite cool customers.

This was the beginning of a fascination with the internet and how the whole thing works. I became really interested in companies and finance and wanted to learn more. All my friends at uni were going off to work for banks, and I felt a bit like, ooh, maybe I should do this too. So I applied to the investment banks in London. My long-suffering director of studies encouraged me to do what I was interested in, and so I actually left medicine after my third year to go and be an equity analyst at UBS. We moved to London, and it was a really interesting time. I was the only medic they'd had on their program before. We spent most of the time learning how to value companies — profit and loss, cashflow, company modeling — and I ended up doing the pharmaceutical companies, so the medical background was useful.

But I discovered I didn't like getting to work for 6am, which was the requirement. And I was running the ads company in the evenings at the same time. Getting up really early for the banking and then going home and running the company at night was not sustainable. Ultimately I just wanted to build things myself — in banking you're basically advising or commenting on others, and I wanted to be in it myself. So I asked to go back to Cambridge, and my same director of studies very kindly welcomed me back.

About a year after that little adventure started, I went back to clinical school, which was really good. I had a bit of renewed vigor by this point. Everyone else was in their fourth year, getting a little tired of everything, and their friends had left — and I was there with renewed vigor, having seen how the real world works and really appreciating being a student again. I really enjoyed the clinical school experience, and the practical medicine as opposed to the academic side.

I kept running the company, and in my penultimate year I got an offer from someone to acquire it, which was quite exciting. By now we were a team of about five and it was going quite well. That kicked off three months of very stressful due diligence. I was trying to do a GP rotation at the time and I'd just had a knee operation, so it wasn't a great time — but it all went well in the end, and they acquired the company in 2008. I stayed involved for another year and a half or so.

From that, I learned that building things is fun — you can work on a little bit of it each day, and when you look back you've built something much bigger than you could achieve in one day. It's a really nice feeling. Same as research, actually. And because I was genuinely interested in ads and really passionate about it, it didn't really feel like work. So I thought: for the future, I need to be passionate about things. The only part I didn't like was the quarterly VAT returns — the most awful part of the whole thing. I actually still have the deadline in my calendar now, so I appreciate not doing it.

Through all this I'd caught the startup bug. Then some of my friends started getting married, and one asked me to build a website for their wedding. That led to the next silly venture — gettingmarried.co.uk, a website where you can create a website for your wedding: your gift list, RSVPs, a bit about you, photographs, a map, everything. That gained traction too. In the end we were a team of about five, about one in twelve people getting married in the UK used it, and it got another offer to buy it in 2016, from a competitor — a gift-list company. There was a great synergy: we had lots of people signing up for their weddings, and they wanted people to use them for their gift list. It was the obvious thing for them to buy.

The thing I learned there was that I wasn't really that passionate about weddings compared to ads. I realized it wasn't the thing I really loved, and I found it harder to be motivated about it. So I decided that, from then on, things had to involve health and technology, and I'd try to restrict myself to those two, because that's where I really wanted to work — and try not to get distracted by ideas that come up along the way.

A side effect was that, because I'd got involved in internet stuff, whenever a friend had another friend doing something on the internet, they'd say, "oh, you should meet Chris, he likes the internet." So I ended up meeting other people doing interesting companies, and that led to the idea of angel investing, which I hadn't really heard of before. It led to almost 20 angel investments over the next 10 or 15 years, which spawned a whole set of fascinating journeys — living vicariously through others. Many haven't gone so well; some have gone quite well. It highlights that the startup world is really hard, but also really exciting. If it goes well, it's a really exciting ride.

Then, ultimately, back to medicine. I specialized in pediatrics, got an academic clinical fellow job — which gives you nine months of research — and ended up at the Evelina at St Thomas', and this amazing biomedical engineering department at King's College London, working on baby brain MRI. It's a great place to work: a small number of doctors surrounded by really bright engineers, computer scientists and physicists working to make MRI work better for newborn babies and fetuses. I really enjoyed it — it felt like the dots were joining up. A bit of engineering, some medicine, some research, all coming together. I ended up doing a PhD there in medical imaging, looking at how the brains of babies with congenital heart disease develop. Three really enjoyable years with a great team at King's.

That led to a fascination with machine learning. Around 2015 I was getting quite excited by what it could do. I met this amazing person, Tony Young — the national clinical director for innovation in the NHS — and he invited a bunch of us to a conference in San Diego called Exponential Medicine, in 2015. We met some really interesting people, and it opened my mind to medtech. TensorFlow was actually launched by Google during that conference, and everyone was very excited. When we got home, I started trying to train some models to predict things from the different images I was using in my PhD, and it got all very exciting — you're amazed at how well it works. This technology has a lot of potential.

From then, I remember waking up in the middle of the night in a bit of a sweat that I wasn't working on AI and health. It became the new obsession. I was hunting around King's for expertise in this area, and, oddly, it was still quite early — there wasn't much. I spoke to one of the co-founders of DeepMind after a talk I went to, and that led to a recruitment process. I ended up joining in 2018, which was pretty lucky. For me it was a bit of a dream — everybody I'd followed on Twitter seemed to work there. Suddenly you're going to lots of amazing talks, there's a great environment and culture, this new era of computer science linked up with medicine. I got involved in lots of interesting work: ophthalmology with Moorfields and Pearse Keane, who you've had on the podcast before, and Joe Ledsam; and radiology work, like our project to detect breast cancer in mammography images, which I've been working on for the last few years now. So that's where I am now.

I was thinking about what things were most useful along that journey. I don't know if you've watched Steve Jobs' 2005 commencement address at Stanford. He tells some stories, but one that really struck a chord with me was about the dots of life joining up later. You can't join the dots looking forwards; you can only join them looking backwards. If you think about all these slightly random things I ended up doing, that talk gave me the confidence to make these slightly random jumps — the confidence that one day it would come together. And I think it sort of has, a little bit, in the current job at Google.

I think the first thing is luck — a huge amount of luck. Being in the right place at the right time. The ads were just taking off, the internet was just taking off for the first one. That's a large part of it, to be honest.

The other thing is just being a little bit obsessive. For the first website at school, I emailed every editor of every newspaper telling them why they should write a little article about it. With hindsight I'm not sure I'd do that now. Or the ads company — it was thousands of hours of work and lots of late nights. And it wasn't all good. Like I said, it almost went bankrupt the day before my anatomy exams. I only had one customer at the beginning, and they went bankrupt, leaving me with this huge debt. I thought, gosh, things had gone so well up until this point, and now I'm down a monstrous amount of money — tens of thousands of pounds, which for me as a first-year student was huge. So it was, okay, I've got to start again. And from then on you learn something each time. That time it was: I'm not going to work with small companies now, I'm going to work with big ones that won't go bankrupt. Aim for big companies I've heard of, rather than small ones I haven't. And just keep chipping away at that debt.

Ultimately, though, it's about really enjoying what you're doing. Trawling through those databases of dialects trying to find Yorkshire words — that's not something you'd have done if you weren't really interested in it. And it wasn't something you were doing for money. So maybe it's just about doing something you love.

That's a really good observation, and I feel the same, ironically. I think it's one of the challenges of health in the UK specifically — I don't think it applies to other countries. In the US, if you're doing a healthcare company, people are fine with it. It's just that we've got used to this NHS system, which is fantastic, and we're all very grateful for it — but it has led to people being very suspicious of anyone trying to make a business in healthcare.

I don't really know what the solution is. But if we want innovation in healthcare, we need people building companies in healthcare as well. Obviously there can be innovation within the NHS, and that's brilliant — we need to encourage it. But it's really hard, because the NHS is often one of the hardest places to change. In my first year I tried to change the drawers on the cannulation trolley, just to make them a bit more sequential so you could pick things out as you wanted them rather than hunting around the stock room — and the opposition to that was immense. To the point where you think, this is really insignificant; how would I ever manage to do something more meaningful if this is so hard?

Companies are a good way of being really nimble — you can just do what you want, move fast, and you aren't held back by legacy. I'd love to see a really vibrant ecosystem of healthcare companies in the UK, and I hope we get there. Actually, I think the culture has changed a little. People do admire those trying to make a difference now. I've got lots of friends building healthcare companies in the UK, and I get the feeling people admire them much more than they might have done ten years ago.

It does sound stupid, doesn't it? The thing that was really cool was that never before had there been a system that could filter people according to what they wanted. If you imagine a newspaper: you open the page and there's an advert for a car, but I already have a car, I don't need one — so that's a wasted impression. Whereas on a search engine, you type in "breakdown cover," and the subset of people who type that in probably are interested in breakdown cover. If they then type in "the AA," they're pretty much trying to get to the AA website to buy some breakdown cover. Never before in the history of the world — apart from the Yellow Pages, I suppose — had that been possible at such scale.

I found that fascinating. You could go into the psychology of it: what keywords are people using to look for information, versus which keywords mean they're interested in buying? How do the combinations fit together — the buying funnel? What do you type when you're interested in digital cameras? Maybe first you're after reviews. Then it's "Sony digital cameras" — you want a Sony one. Then it's "Sony A7-something-something," and at that point you really want to buy. This whole field was completely new — a really interesting place to just experiment.

I don't know if this is 100% true, but I came across it somewhere: apparently randomized controlled trials came from newspapers and marketers. Have you heard that?

I heard that in the 1870s newspapers started experimenting with different ads. If you had two styles of ad for the same thing, you'd run half your newspapers with one style and half with the other, then see which performed better. And that's supposedly where RCTs came from. I don't know whether it's true, but there does seem to be a lot of interesting innovation in marketing and ads that crosses over into other fields. A/B testing is a core part of web design and ads.

A silly example: one of the first things I was doing in that first year was mobile phone contracts. I took out adverts in national papers — made them myself on some desktop-publishing thing and paid a thousand or two thousand pounds to place them — and if someone signed up using my phone number, I got paid for each contract. I tried all sorts of designs. One I thought looked really classy — Helvetica everywhere, beautiful, really nice — got about two sales, a huge loss. Then I tried another one with that Microsoft Word Impact font — really in-your-face, "summer scorchers" at the top — and that was one of the best ads. It's so funny how it's not necessarily the most beautiful ad that's the best one. It's something that catches your eye and makes you dig in a bit more. It's not intuitive, so you need to A/B test to find out. If you'd never tried it, you'd never know.

When people look at where to focus, they look for big problems with quite standardized inputs. Screening programs are a great place to start, because they tend to take standardized images or collect standardized information, they're done at population scale, and they have the potential to impact large numbers of people. AI technologies aren't like humans — they can't necessarily deal with random, unexpected things, the very difficult, complex case. It's all about pattern recognition, and you need plenty of training data to learn what to do in each situation. So screening is ideal: large numbers of people to learn from, conditions that are quite well defined and follow guidelines.

Breast cancer screening is a good example — there are strict quality-control measures applied to every image, which is very good for machine learning. Or diabetic retinopathy screening: very good quality guidelines. These make good projects because you can do quite robust training and testing. Whereas if you think about someone walking into a GP practice with a problem, it's so much harder to capture. A lot of the information is in how the person walks in and sits down — what's their affect like, do they make eye contact, what do they say, is there something they're not saying that they really want to say but need an in for, which you have to provide by asking the right question. These things are really hard, and the era of machine learning doing that is quite a long way off. Standardized image capture is the best place to start, and the least likely place to cause damage within an established screening setup.

I totally agree. The first exploration of medical AI has definitely felt like, "wow, we have this great tool, how can we use it? Here's a good area, here's another, here's another." And then you publish — there are thousands of medical-AI publications, and I don't suppose many of them have actually been used in patient care, because when you come down to it, it just hasn't been designed with the clinician-centered approach it needed.

So I suspect the next round will be much more clinician-driven. You talked to Pearse Keane about this a little — trying to find a cohort of clinicians who are really interested in machine learning, who understand the basics and the concepts, so they can drive the projects and say, "I've got this problem, I think it's really amenable to machine learning. Let's map it out carefully, think about what it would have to do, where it could go wrong, what unexpected things we'll come up against," and then pull in the computer scientists to help. Where 95% of the project is thinking about the problem: what the data look like, how you get the right dataset, how you make sure it's fair and representative and isn't going to cause harm through something you haven't thought about, and then testing it in a way that actually makes sense for the healthcare system. If it's a screening thing — what does the screening program actually want? How do they measure their existing clinicians? What metrics do they use? Really thinking about all of that, and then running the project. Hopefully that work will be much more translatable into clinical practice than some of the papers, which can be, "oh, we've just solved this problem" — but when you dig in, you haven't really; you've made some shortcut that isn't possible in the real world and is crucial to the translation.

I think we're all becoming more aware of that now, and you can see the quality of publications improving. That's the most exciting thing. We're still at a very early stage of machine learning and health, but I can't imagine how, in 30 years' time, it won't be quite ubiquitous — where we're not talking about it so much anymore, and it's just a thing that has improved healthcare. We've got an interesting decade or two to come in between.

Yes, definitely. I get that all the time. Most of my friends are about to be consultants now, and I see everyone else plowing ahead in their medical careers. But when you really think about it, I've enjoyed the journey so far, and I think it is about the journey. You can rush to becoming the consultant and not really enjoy it along the way. I've really enjoyed it, and I continue to. I don't think it really matters where you end up, so I've tried not to focus on that too much.

I actually find the medical side the most difficult, because we've planned our lives around being doctors from quite an early age — starting medical school and so on. It's quite hard to let go of that. Medicine has a very strong pull once you've trained, and I feel that quite strongly. But, a bit like public health, you think: maybe if you can do a really good job of the technology side, you can have a really large impact on patients even if you're not directly seeing them day to day. That's maybe the thing that keeps me going.

I think everyone should try to learn to code if they can — not just medics, everybody. It's a valuable skill and it does open up opportunities. If you can say, "actually, I could probably just do that," suddenly you get the chance to do something you wouldn't otherwise get to do.

And you should never underestimate the power of a prototype, even if it's not particularly good. A lot of the companies in the past, you'd write the first pass at something that just about works, and then someone much better would come and write something actually production-ready. If you're working on a project and just want to get something started, what a great skill to have. I know it's hard to get started, but there are some great ways to learn, and if you try, you'll probably quite enjoy it — it is quite fun.

The other thing is the value of being able to speak the language of people who do this full-time. When you're working with engineers or scientists — even in the hospital, if you're a consultant working with the lab on a project — if you can do a bit of analysis yourself, it changes the environment and lets you work much better with them as colleagues. You don't suggest things that are quite so silly, because you know that won't work. Even if you can't do it yourself, you understand the ways of doing things. So I'd encourage everyone to have a go, and if you really don't like it after trying, then fine — but it's worth trying.

I'd just be very aware of everything going on around you and hunting for what interests you — then taking the time to learn more about it, thinking, "how might this change medicine in the future? How could this be improved? How could I be part of this journey?" And really latching on to someone already working in that field and learning from them. That's what I'd try to do most, because there are loads of things you could work on. The list is much like it was ten years ago — genomics, machine learning, personalized care — but I think it's just about hunting for what interests you and being willing to take risks. Ask for an out-of-program experience. People often don't, because they think, "oh, I won't get it, it won't look very good." Just ask people for help — you'll be surprised how well they often respond.

I think we've spoken about some already. It's about being quite an interested person in lots of different things, going quite deep into something, and getting maybe a little bit obsessive about things at times. And enjoying taking calculated risks — not being too scared to change track, and having the belief that things will be okay. That's quite important. Also not being in a huge rush to get anywhere — trying to enjoy the journey as you go along.

And one that's quite funny, but: not being afraid to email people. If you admire someone, send them a message saying, "I think that's really good, I really appreciate it" — not necessarily asking for anything. Everyone loves receiving an email like that, and if you genuinely think it, you should probably just say it. I've always been surprised by the response you get. People you'd think would never reply often do, and sometimes it leads to a really nice friendship. And just trying to be a nice person, a good person to work with. Everyone tries, but it does make a big difference.

Yes. One recommendation, again from Pearse — he's come up a few times in this podcast — is Creativity, Inc. by Ed Catmull, one of the co-founders of Pixar. It's a really interesting insight into Pixar from the early days, and how you create a culture of creativity that leads people to do good work — they made these groundbreaking movies. I got a lot of read-across into DeepMind and Google and the ways they set up their culture. I've often thought about that book over the last few years.

There's another one called Drive by Daniel Pink — one of the few books that's really challenged my underlying assumptions about what motivates people. It talks about the difference between intrinsic and extrinsic motivation. I remember reading it on holiday and telling my wife, page by page, what was going on in this book, much to her surprise. A really good book — I recommend it to anyone.

And one thing that really helped me was Andrew Ng's course on machine learning on Coursera. It's quite old now, but it covers the basics really well — the basics you'd have learned doing computer science. I often think the things I learned in that course are useful in general life. But if you're interested in this kind of work, it's a good starting point for a clinician without a formal background who has an interest in computers.