[0:00:00] What a phenomenal start to the morning. That was super energizing. I enjoyed that. And hopefully I'll touch on a couple of the themes that Matthew brought up. So I'm going to talk a little bit about some new ways that we think maybe technology might be able to enhance patient care. But before I get into that, let me sort of give you a bit of background on DeepMind. It's a company that I started six years ago, a British artificial intelligence research company.
[0:00:26] And back in 2014, we were acquired by Google. We became Google DeepMind. And we're responsible for all of Google's general purpose AI efforts. And fundamentally, our mission is to create general artificial intelligence. So these are algorithms that are capable of learning directly from the raw data and can perform well across a very wide range of tasks, much like a human being learning from scratch.
[0:00:51] to do inference and prediction in unstructured environments. But the second part of our core mission is to use the technologies that we develop to make the world a better place and really try to tackle some of what I think are society's toughest social problems that we're facing. At our heart, we're a research organization.
[0:01:10] We're based in King's Cross in London. We have 300 or so staff coming from 40 different countries and 200 of the best machine learning and artificial intelligence research scientists in the world. So 200 postdocs and PhDs, the largest concentration of AI researchers anywhere in the world with 12 tenured professors full-time in the office.
[0:01:30] And we publish everything. We're very committed to being open and publishing 80 or so peer-reviewed research publications since we founded the company.
[0:01:40] But we have very much what we call a hybrid culture. We started the company as a response to what me and my co-founders saw as a kind of failure of conventional organizational structures and institutions to respond to the complexity of the tough social challenges that we're facing. So we like to think of ourselves as retaining the long-term focus of academia whilst dropping some of the somewhat more
[0:02:05] Painful challenges in academia where it's really difficult to scale things up with good infrastructure and engineering and at the same time sometimes getting trapped in the cycle of publications that aren't always critical to the long-term research effort that you're focused on.
[0:02:21] But secondly, combining that with the pace and the scale and the agility of a commercial environment where we really can afford to have a very large research engineering support infrastructure around our core researchers, all the while underpinned by the social values and the mission of the third sector. And I think in some sense that makes us reasonably unique in attempting to combine all of these three different ethoses in a single organization.
[0:02:50] And so let me briefly tell you how we attempt to go about building general purpose learning algorithms. Everything starts with an agent. You can think of an agent as a control system for a robotic arm or a self-driving car or a recommendation engine. And that agent has some goal that it's trying to optimize.
[0:03:09] Hand code that goal. It's the only thing that we give the agent. We say, these are the things that you should find rewarding in some environment. And the environment can also be very general. So it could be a simulator to train a self-driving car, or it could be YouTube, where you're really trying to recommend videos that people find entertaining and engaging.
[0:03:27] And the agent is able to take a set of actions in some environment. So it's able to experimentally interact independently and autonomously in the environment. And that environment then provides back a set of observations about how the state has changed in that environment as a result of the agent interacting with that environment. And of course, the environment passes back a reward, which the agent is able to learn from. So it's really learning through feedback or through the reinforcement learning process.
[0:03:56] And we train in the Atari testbed. So these are the 100 or so games from the 70s and 80s where the environment is actually the Atari simulator. And the raw pixels are provided to the agent in their rawest form. So really just think about an RGB, a red, green, blue.
[0:04:14] Thank you very much.
[0:04:36] Pre-programming. We don't tell the agent anything about the structure of the environment, only that its goal is to maximize score. And so there's a single agent that learns to play all of these games. And I think the kind of best intuition is imagine a robot that is able to control the joystick and the fire button at the arcade and is really just processing the raw pixels that are coming into its visual stream. So here's a little demo video of a classic game called Breakout.
[0:05:01] Where basically you're controlling the paddle at the bottom here. And this is only after 100 games, so the agent really has just been dropped into the world and it performs really badly. It's sort of randomly exploring the space, figuring out what happens if I move left or right when the ball seems to bounce down here.
[0:05:17] After these 300 or so games, the agent gets pretty good. And this is pretty much human performance. It's learned purely from the raw pixel inputs that any time it bounces the ball up and hits one of the blocks, it gets more score. But then unexpectedly, we left it training for about 500 or so games. And the really interesting thing is it learned this cool strategy of tunneling up the sides and sending the ball up the back to get as
[0:05:39] As much score as possible with as minimum effort that is required. And so this is an example of a very, very crude, early, small-scale system in a toy-like environment that has learned something that we hadn't pre-programmed into it that turns out to be a very effective strategy. And we were lucky enough to have got a Nature paper for this work back in 2015, and we were given the front cover, which, as you'll know, research scientists were very, very proud of.
[0:06:07] A few months back, we went on to tackle one of the grand milestones in artificial intelligence, the classic game of Go. It's 3,000 years old. There are 40 million players around the world. And the really interesting thing about the game is its complexity. So there are 10 to the power of 170 possible board positions on a 19 by 19 board. So that's a 10 with 170 zeros next to it.
[0:06:30] Possible board configurations on a 19x19 board. And the objective is to really place black stones or white stones to surround your opponent and gain territory. So it's a very unstructured game, unlike chess, for example, where all of the pieces are worth different amounts and you can use the value of those pieces to bias your training. You can put lots of energy into trying to protect your queen, for example, and sacrifice your pawns. But there's no such structure in Go.
[0:07:00] So it was widely recognized to be one of the really hard, outstanding challenges in artificial intelligence. And in fact, just to put it in context, it's been estimated that there is something on the order of 10 to the power of 70 atoms in the known universe. So there are, in every liquid, solid, and gas,
[0:07:19] Fewer atoms in the known universe than there are possible ball positions in the game of Go. So traditional techniques of writing rules or hand crafting features, as it's known, just doesn't work. You have to train an algorithm to learn its own representations from scratch. And we were really lucky to be able to beat the Go world champion, the first time a computer system's ever beaten a human. And in fact, we beat the best human in the world, which was really exciting in Korea three months back.
[0:07:47] 280 million people watched it live. It's one of the most popular games out there. And that's more viewers than there were in the Super Bowl, which is pretty cool. 35,000 press articles. And incredibly, 10 times more Go boards sold in two weeks. So they sort of ran out of Go boards, which was pretty cool.
[0:08:05] And we were, again, lucky enough to get a Nature paper for this, and we were given the front cover, which is the first time that a computer science lab or an AI lab got two Nature front covers in 12 months. So pretty cool.
[0:08:19] But really, what we're about is making these sorts of advances in a research and engineering environment and then trying to apply our expertise to make the world a better place. So why did we pick health? Well, about a year and a half ago, a bunch of my team got together and spent about three or four months trying to figure out where we should go next. I mean, we could pick all sorts of different application areas. And there were three core principles guiding what we were looking for. Clearly, we wanted rich, complex data environments.
[0:08:48] which are relatively unstructured that require us to make interesting predictions. As you can see, we're clearly motivated by social impacts and that's something that was very important to us to find an opportunity to take our technology into the real world and make a meaningful difference. And of course, we wanted to build a sustainable business model around this. But the kind of most remarkable thing about health compared to all of the other things that we looked at like synthetic energy,
[0:09:14] Developing sustainable food, water purification, nanomaterials, material science in general. We looked at all these different areas. The remarkable thing about health was that there's an incredible margin for improvement if we're successful in being able to deploy cutting edge and modern technology systems. I mean, there really is no other sector that I can think of in the world that is so far behind the cutting edge in terms of technology. And if we're successful,
[0:09:41] That represents a massive opportunity for us to have a beneficial impact. And many of you will obviously be familiar with this being the kind of default reality. So much of the valuable information that describes the current state of a patient's condition or the current efficiency of operation of a particular ward exists in this kind of data environment. And so when we look at the current health IT systems and research that was done recently at Imperial, talking mostly to junior doctors and nurses, people repeatedly say things like,
[0:10:11] It's clunky. It's not intuitive. Everything takes too many steps. And stuff doesn't mean what we all think it means. Or it took an hour to teach a doctor how to prescribe paracetamol. Or I don't think I'm stupid, but it couldn't be more complicated. If I can take a computer out of a box and understand how to use it straight away,
[0:10:31] Why is it so useless that I need six hours of e-learning to be able to order a blood test? And I think it's this really strange dissonance, this sort of frustration that in your lunch break you're using Facebook and Uber and you're on messaging platforms and they're at the very cutting edge of the technology that's available to you in the world. And then in the afternoon you go back and use a desktop-based system that struggles to boot and you may be using pages and fax machines to go about your daily work.
[0:10:57] And so we launched DeepMind Health in this context back in February of 2016.
[0:11:02] 400 or so clinicians at the Royal Society of Medicine. And in many ways, it was like an incredibly daunting prospect. I mean, in some sense, as many people have already pointed out, it's kind of a graveyard of failed technology efforts over the last 20 years. And so I think in that context, we really had to think about what are we going to bring that's going to be very different. Clearly, we have machine learning and artificial intelligence. But I think a lot of this is about the approach that we take to developing software and how you put both patients
[0:11:30] And clinicians at the very forefront of that. And so the approach that we take is to frame everything as starting with an observation of what does a user do on a day-to-day basis. We spend lots of time immersing ourselves in wards and with nurses in the mess rooms.
[0:11:46] Trying to observe what they do, define their challenge, gather as many insights as we can, and then immediately start to build something. As fast as possible, we want to show what a rough design might look like. Here are some wireframes. And then develop that a little bit further, test it.
[0:12:02] And then as we start to develop a solution to try and measure, build, and learn, and then just rinse and repeat, try and do that in very, very quick iterative cycles. And so within three weeks or so of meeting our first nurses, signing our agreements with the Royal Free back in September and October,
[0:12:20] We had a working prototype, obviously not connected to any data, but that nurses and doctors could actually point to and say, this button's in the wrong place, this color is difficult to read, this menu hierarchy is sort of in the wrong order. And so we can instantly get feedback and deliver pretty much what nurses and doctors tell us they want to see.
[0:12:40] And so this is kind of our mantra, ABC, always be clinician-led. And every single project that we will work on and the projects that we've worked on so far have been brought to us by a nurse or a doctor who has some idea, some insight into how they can change the behavior in their day-to-day operation and how a technology solution might actually work. And here you can see this is one of our whiteboards in the office, always reminding ourselves who our users are and what they need.
[0:13:07] And so how might patient care be better supported by technology? Well, I think obviously there's an enormous opportunity for improvement. One in ten patients experience some kind of harm in hospitals, and half of those are completely preventable or avoidable harm.
[0:13:23] And in many of those cases, 50% of those cases, detection of the patient deterioration in question has actually been delayed. And that's actually a communication and coordination issue. And largely, I think this is because of these current limitations. Most of the really valuable data sits on paper and on charts and isn't logged or tracked or recorded. There's no auditable log that you can verify of the pager messages that have been sent, the reminders that have been sent, those that have been missed. I mean, who knows, maybe that person's
[0:13:53] Not even at work at the moment, maybe they're on hold-in, and you don't know until they don't reply for two or three pages.
[0:13:59] So I think there's two core patient safety challenges that have framed everything that we do in DeepMind Health. The first is, how can we do a better job of identifying which patients are at risk of deterioration, largely in real time? And then the second is, once we've identified which patients are at risk, how do we actually intervene? I mean, we don't want this to end up just as a report that advises on some reorganization of facilities on a ward. We actually want to deploy technology in real time.
[0:14:28] And so on our patient safety challenge number one, how do we do better detection? We've looked at acute kidney injury over the last sort of 12 months or so. And this is a remarkably important problem. 25% of all admissions present with some kind of evidence of an AKI. And there are 40,000 or so deaths in England due to AKI alone. It's estimated that something like 20% of these are actually preventable.
[0:14:57] And then the cost could be as much as a billion and a half pounds. And so sort of a couple years ago in 2014, NHS England issued a patient safety alert to mandate that the acute kidney injury algorithm be deployed in hospitals.
[0:15:14] And so once again, the first thing that we did is try to observe users in their day-to-day settings. So we went into the role-free and we mapped out the pathway. What is the experience from a patient perspective today? And it turns out it's actually really, really complicated. There are lots of different stages to the path that a patient might go through. And what we noticed is that there are a whole series of life-threatening and
[0:15:38] complicated stages in that pathway, which actually seem to be where we're missing on the key bits of deterioration. And so what we wanted to do is take a step back and see how could we intervene earlier to do better risk assessments, more real time prevention and monitoring, and then hopefully try to redirect patients through the pathway towards potentially a full recovery and then a discharge. And once we've broken it down into these sorts of steps, we have a shared visualization between us and the clinicians on
[0:16:08] Where all of the key intervention opportunities actually sit. And so in response to this, we developed Streams, our AKI alerting platform for blood test results. So I'm going to show you a little demo of what this looks like. We start with our login screen. We're able to securely log in on mobile. And as soon as you've logged in, you're presented with a list of all of the patients
[0:16:34] That appear to be alerted on the acute kidney injury algorithm. And you can scrub down that list, maybe tap on Robert Jones here. You can see all of the blood test results. You can see that their creatinine has spiked over the last sort of few weeks of the baseline. And we can also compare that to any other blood test result and pinch and zoom, scroll through the others like CRP or potassium. And so that's the very simple intervention that we've built so far.
[0:17:03] Keeping it really focused on one very specific condition using the blood test results. But in the future, I think there's a real opportunity for us to go much, much further and extend this to a broader patient-centric collaboration platform. So here's a view of what things might look like in the future. We might be able to search through all of the patients, maybe by consultant, by ward, or by specialty. We can scrub down our list of patients and pick
[0:17:32] Or we can type in a particular patient that we're interested in because we're on that ward. Then we go through to Robert Jones. We can see their date of birth, their medical record number, the ward they're on. We can go through their overview, see their current medication, see their vital signs. Here we can see that the patient's temperature is pretty high in their blood test results. Their blood pressure is high. Maybe we want to look at their past diagnosis. We can see that this patient has pneumonia. We can see their previous procedures.
[0:18:02] We can also see their current medication that they're on. And if we swipe right, we can go across to the timeline and see that this patient is being treated for sepsis. They have an AKI2. We can see when the last blood test was ordered, that it's in the lab at the moment, that it's ready for review. We can also scroll down and say, look at the SATs, for example, or we can see that this patient had been previously transferred. We can obviously go across and take another look at the blood test results.
[0:18:31] And all of this information is in the palm of your hand, able to be reviewed whenever you need it. Here we have the x-ray on mobile. We can see that this patient has pneumonia, and we can also see the report from the radiologist. You might want to zoom in and then share it with a colleague. And so if we think about...
[0:18:51] What this does for us, this actually essentially puts in the palm of our hand the ability to detect in real-time patients that are a risk of deterioration. But that's only sort of one part of the challenge. The next key thing that we need to be able to do is escalate and better intervene. And this is where messaging and commenting become so important. So take, for example, with the x-ray that we looked at just now. Here we see that a registrar is able to...
[0:19:17] Make a comment on the x-ray, on that report, and then plus in a respiratory consultant to get an expert view. And that exchange can happen in an auditable way that allows us to verify retrospectively, if needed, what the senior clinician has actually said and what action is subsequently taken.
[0:19:39] Here we see an example of the blood test results where an F1 is able to call on the specialist nurse to come and see the patient as soon as possible and potentially organize a renal ultrasound. And so that gives you a bit of a sense of some of the things we're doing in clinical decision support and direct patient care, how we're trying to build apps that put this kind of information and this kind of communication and coordination in the palm of a clinician's hand.
[0:20:05] Quite separately to this, we're also embarking on a research program to see if our machine learning and AI technologies can actually help with some aspects of diagnosis. So let me tell you a little bit about one of the problems we've been focusing on. Sight loss and diabetic retinopathy, as well as AMD, is something that actually affects 625,000 people in the UK and 100 million people worldwide. And obviously diabetes is a very much growing problem. And the remarkable thing is that...
[0:20:32] If you do have diabetes, you're 25 times more likely to suffer some kind of sight loss. But most interestingly, the very most severe types of sight loss due to diabetic retinopathy can actually be prevented through earlier detection. So one of the things that we've been thinking about is how could we potentially help with better real-time classification of those radiology exams coming through to enable a more sensible triage of which patients require more immediate responses.
[0:21:02] And so the current reality is that in human performance, there's a great deal of backlog in reporting, which means that the results aren't available potentially in clinic for weeks. And there's also a lack of consistency between different graders. And sometimes the reporters will miss some of the sensitive changes in diabetic retinopathy and AMD.
[0:21:20] With machine learning, one of the things that we hope we might be able to do is to do much faster and near instant results, but also more consistent and more standardized performance. And I think this will also help us to understand, to adjust for some of the normal variations that we see, which will allow us to increase our specificity.
[0:21:38] So recently we've been working with the Moorfields Eye Hospital. And so I want to show you a very short video so that you can hear directly from our friends at Moorfields, Sir Peng and also Piers Keane, who we've been collaborating with on the project. My name is Piers Keane. I'm a consultant ophthalmologist at Moorfields Eye Hospital. I'm Professor Sir Peng Tee Koo. I'm a consultant eye surgeon here at Moorfields Eye Hospital.
[0:22:10] Wharf Hills is the oldest eye hospital in the world and we see over 600,000 patient visits a year which is more than twice any other hospital in either United States or Europe. It's a wide-ranging hospital where we see very common conditions and also very rare conditions. Every week we do many thousands of OCT scans. OCT gives us very high resolution images of the eye in a very non-invasive manner.
[0:22:38] Some people are waiting long periods of time to see specialists in the eye clinics and therefore have the potential to lose sight in that waiting period. This is where DeepMind is able to help us understand these huge data sets and then put it together so it benefits towards making a good diagnosis and achieving the best possible treatment for our patients. If you have an OCT scan done, a machine learning algorithm will be able to tell you if it's something urgent
[0:23:07] versus something that's not so urgent. I think that this has huge potential for patients. It will allow us to get much earlier detection of these blinding diseases and as a result of that I think we'll get much earlier intervention and much earlier treatment for these patients.
[0:23:27] So this is very much early work, but we're committed to publishing all of the results of our work, including our algorithms, our methodologies, and our technical implementations. And so hopefully, when we're ready, you'll hear more from us on the results of that research towards the back end of this year.
[0:23:44] So I think it's really important that we outline at least four of our key principles that define everything that we do with respect to deep mind health. I think the onus is very much on us to find a new way of working with the NHS. And we really do want to take this opportunity to invite everybody to participate with us in helping us to do things quite differently. And I think there's at least four key things. First of all, being patient centric, being clinician led.
[0:24:10] Committing to open and operable standards and the highest levels of transparency and accountability. So let me touch a little bit on each one of those. The first thing to say is that I think we really have to focus on improved patient safety outcomes and experience. And I think in order to do that, we actually need to be putting the patient first in everything that we do. And it's an incredibly energizing speech for Matthew just now to hear him talk about the patient portal, talk about the patient owning the data and having access to it.
[0:24:38] Thank you very much.
[0:24:46] Giving them care. At the moment, I think the experience is to always be on the receiving end of treatment and all these other things that are happening to you. I personally have a partner who has a chronic condition and we've been in and out of hospital for four years or so. And the number of times that we refill in the same form very patiently with our date of birth and address and those details, the number of times that we're asked, you know, what are the drugs that you've just taken this morning, which...
[0:25:09] Expert or which specialist did you see earlier can be very, very draining. And I think if there is a canonical record of this information, then it'd be much easier for us to play an active part in our own treatment.
[0:25:20] And so we're really pleased to announce that we're going to have our first patient engagement meeting on the 7th of September. So anybody is welcome to come and join that. Please email us at patients at DeepMind Health. We're looking for feedback. We're looking for people to help co-design with us. We're looking for ideas and, in general, any thoughts or suggestions that people have are most welcome.
[0:25:44] We also recognize that it's incredibly important and delivers enormous momentum for us to be designing directly for clinicians first and foremost. And so all of our clinician facing tools will be led by
[0:25:57] I actually met Piers Keane because he cold emailed me on LinkedIn and a couple days later we got together for a coffee and he pitched me this incredible idea and I was like this is really awesome we have to collaborate on this the same story with our work with the Royal Free and so I think this not only generates an enormous amount of momentum but allows us to really be led by the specific needs that nurses and doctors have and so once again
[0:26:23] Please email us if you have any ideas or you want to be involved in user testing or if you want to give us some feedback on something that we could be doing better. We can only be successful if we hear from people who are at the front line of service delivery.
[0:26:37] I want to say something really important about how we go about creating this open ecosystem to drive innovation. People have been talking about open standards and interoperability for 20 years. And it's actually not that difficult. The technical implementation can be painful, but we're actually converging on the standard, the FHIR API standard, and aggregating the data in the back end, despite the fact that it's often spread across...
[0:27:02] 100 plus databases of different schemas and in different standards in many hospitals is actually very tractable. This is not a technically challenging problem. This is not a research problem. And we've actually had some success in starting to think about how we might do that with the role-free. And I think one of the barriers to the question earlier this morning, one of the...
[0:27:26] The barriers to smaller companies, startups, to the doctor hackers that have been mentioned, to getting involved, is that it's all well and good designing a really nice app, but if you can't get hold of the data, then you're never going to be able to test it and trial it. And so there has to be a secure and controlled way of managing the release of that data to the appropriate experimental app. And I think that is what will align the ecosystem around a single backend canonical record that I think
[0:27:56] It's what we really need to drive much more competition in this market. There needs to be much more transparency and competition. And so I think this is where patients and clinicians really need tools that actually encourage the innovation process. And I think it's sort of a false psychotomy to sort of on the one hand say that the comprehensive top-down systems are all that we can hope for on one side versus the piecemeal and clinician-led innovation on the other. I actually think we need both and we can have both.
[0:28:25] And that's something that we'll be working towards. And so we've made a commitment before, but we'll restate it again, that all of our work will be built on open standards, non-proprietary standards that can be edited and updated by the community in the conventional way. There's actually a long history of doing this at Google. Finally, I want to touch on the importance of governance. This is something that is very dear to me and I feel very passionately about.
[0:28:54] Before we launched DeepMind Health in February, we announced that we would have an independent review panel of unpaid reviewers who meet four times a year, have a small budget for a secretariat to do audit, and can talk to any member of my team. They can one-on-one interview any member of my DeepMind Health team. They can look at any of our agreements, our product roadmap. They can diligence our technical infrastructure and all of our IG.
[0:29:21] They're not bound by any NDA or any contract, and so they're free to speak their mind. But we've appointed nine expert reviewers who I think will have the freedom to try to hold us accountable and build trust in the community so that they're able to sort of say what they think of how we're doing to give us guidance and direction, but also to have conversations with people outside the organization about what they see and about what we're achieving.
[0:29:52] With that, I want to say thank you very much for listening to me. Cheers.