[0:00:05] It's now my pleasure to introduce our next keynote speaker. And the story, as far as you're concerned, begins in 2010 when three young entrepreneurs founded a company which turned out to be at the forefront of something that is going to affect all of our futures. Artificial intelligence, the algorithms that enable and develop machine learning. DeepMind was bought by Google four years later. And now...
[0:00:32] DeepMind Health is building clinician-led technology in Britain's National Health Service. It has led to headlines such as Google wants your health records. But if we are serious about harnessing the power of an era of big data, this is technology that we all need to understand. So please welcome the co-founder of and head of Applied Artificial Intelligence at DeepMind, Mustafa Suleiman. Thank you very much. Good afternoon. Welcome, everybody.
[0:01:15] So first and foremost, I wanted to extend our thanks and gratitude for an incredible conference so far to Her Highness. This has been a really exceptional 48 hours for me to hear about such a diverse range of opportunities and challenges that we see in the system today. Very excited to be here, so welcome and thank you to Her Highness. Over the last couple of days, we've heard an interesting range of challenges at the conference. Of course, we're all very familiar.
[0:01:49] With the challenges of our aging population, with the fact that all of our patients now are actually on multiple comorbidity pathways simultaneously. And all of these treatments are increasingly complex and increasingly expensive. And of course, the patient, let's not forget, is at the heart of all of this. She's demanding more and more. She expects more and more. As we heard on the panel just now, we have the most cutting-edge, beautiful applications to share photos, to play games, to read up about the news.
[0:02:19] And yet, when our clinicians are saving lives on the front line day in, day out, they're struggling with technologies that are, in some cases, decades old. In some cases, not used in any other environment apart from healthcare, if you consider fax machines and pages, as we do in the NHS. And of course, another one of the challenges that's been raised is the opportunity to really rethink the way that we do ethics and governance. Her Highness introduced this question in the opening plenary, and so did the professor.
[0:02:48] Lord Darzi, of course. This is something that's very close to our heart, a deep mind, and something we think very deeply about. Many of the strategies that have been raised are very interesting and very novel, been taken from other sectors, from other areas of academia, and from other applications. Of course, our focus is not just on hospitals or on direct care and primary care, but on healthy populations as a whole, on behavioral insights. What can all of these things contribute to healthcare?
[0:03:17] We're all very familiar with the opportunities presented by much more accountable care organizations and of course the long-term promise that big data and precision medicine offers us. Of course we'd like to see much more radical investment in the sector and much more targeted return on investment too. Our health professionals need much more education and that's something that we've heard about a great deal.
[0:03:42] But I want to argue to you today that, in fact, all of these opportunities actually require us to rethink the way that we capture the data that describes the complexity in our healthcare systems. And in doing so, aim to capture both the right piece of information at the right time, but also with the right underlying serving infrastructure. And that technical question sits at the heart of, I think, all of the future innovations that we'll be able to make over the next 10 years.
[0:04:10] But before I go on to talk about that, let me start by telling you a little bit about DeepMind Technologies, the company that I started with my friends about six years ago. Our mission is to solve intelligence. Sounds pretty grand, sounds pretty wild in some ways, but in fact, there's a very simple intuition that sits behind this ambition. Everything that you see around us, our culture, our music, our politics, our family structures,
[0:04:39] Everything is the product of this intelligent brain that sits on my head, the wisdom that arises, our ability to make good predictions, our ability to simulate the world. Everything is a product of our human intelligence. And so if we could somehow distill some components that make up this intelligence, if we could take the essence of our intelligence and distill it into an algorithmic construct, one that can benefit from incredible access to large data,
[0:05:08] Paralyzation from compute. Perfect memory storage that we get in silicone. Perhaps, perhaps, we could use some of these insights to try and make the world a better place. And that's what motivates my team at DeepMind Technologies. How can we use cutting edge machine learning technologies to try and tackle some of the world's most complex problems? And this is what really sits at the heart of our mission. General purpose learning algorithms.
[0:05:35] These, we believe, have the potential to learn directly from raw experience by interacting with the environment, not with brittle rules that have been designed by intelligent human beings that are very strict and rigid and can't adapt over time. And we'd like to use these methods to try to address some of the incredible array of pressing social challenges which I think are overwhelming us today. We're running up against hard limits, whether it's in climate,
[0:06:04] or energy, transportation systems, or the production of sustainable foods that feed the world, or of course our healthcare systems. Groups of the smartest experts in the world today are struggling to tell us what we should do to improve our system. And I think it's incumbent on us to try to use some of these methods directly first and foremost to solve some of these tough social problems.
[0:06:27] At our heart, we're very much a research organization, but we've tried to build over the years a hybrid organization that takes the very best from research with its focus in academia as we do on very long-term hard problems and combine that with the principles that you get in a company which make companies often so effective. The pace, the scale, the agility, the commercial focus, but all the while underpin that with a social mission.
[0:06:54] With the values that you often find in the public sector or in a charity. And this is so important to us, being able to hire some of the most talented researchers in the world. The best people want to work on the most meaningful and socially impactful problems. And that's why we've managed to bring together a group of 250 machine learning and artificial intelligence experts, PhDs, postdocs, tenured professors, now the largest AI research lab in the world.
[0:07:23] So what does it look like to build general purpose learning systems? At a very high level, our agents learn knowledge through interaction with the world. So think of an agent as a recommendation engine or as a braking tool in a self-driving car, and the agent gets to operate in some environment with respect to some goal. It's able to take a set of actions, and then it can observe how the environment has consequentially changed
[0:07:50] As a result of its interaction, it observes how the state has changed and it looks for reward in that environment. And in doing so, just as we do through behavioral learning or reinforcement learning, it learns to build a model of what is rewarding or effective in the world simply by associating the past set of behaviors that it took in the run-up to some rewarding action with the subsequent action that was rewarding in the first place.
[0:08:16] So we trained our agents a couple years ago in what's called the Atari testbed. So these are the old school 1970s and 80s Atari games that many of you might remember. A hundred or so different games.
[0:08:29] And all the agent is given is access to the very raw environment. So these are just the pixel inputs, and it's wired up to the action buttons, but it's not told anything about what they do. So all of the knowledge and intelligence is learned purely from scratch with zero pre-programming. So we don't tell the agent anything about these environments. And the goal is simply to maximize the score. So the best way to think about this is literally a robot standing at an arcade.
[0:08:55] Looking at the screen, watching the frames as they come in, 24 frames per second, and trying to associate the pixels that it's seen in the last few seconds with the rewarding action that it stumbled across. So let me show you a video of what this looks like. This was one of our early agents in playing the game of Breakout. The agent controls the paddle at the bottom of the screen that you'll see, bouncing the ball up and down to knock the bricks out. And this is after 100 or so games when the agent still
[0:09:23] Pretty poor, hasn't really learned much about the game. But after about 300 or so games, the agent has learned to move the paddle to hit the ball in order to get score, and it does it pretty well. This is basically as good as a human. I mean, this is pretty much better than me. But the interesting thing is, after about 500 or so games, we thought we'd leave it training, and it learned a really interesting strategy, which was to make a tunnel and funnel the ball up the back to get maximum points with minimum effort.
[0:09:51] And this was really interesting to us because it was the first time that we had trained a system to teach itself what was rewarding. So in some sense, the system had discovered new knowledge that we hadn't programmed into the system. It was teaching us that this was a rewarding strategy. And that was very exciting to us.
[0:10:12] We were lucky enough to be awarded with a Nature paper for this, and we made the front cover, which many of you will know in the research world is the highest accolade that all of our teams are motivated to try and achieve. Earlier this year, we extended the same work to the classic game of Go. You might be familiar with the 19x19 board with black and white stones, one of the most complicated games that man has ever invented, far more complicated than chess.
[0:10:39] We trained a system earlier this year to play against the world champion, and again, we were lucky enough not only to beat Lisa Doll, the world champion in Korea, in a live match watched by 280 million live viewers, which is bigger than the Super Bowl, but we were also lucky enough to get our second Nature front cover. So we've demonstrated that these sorts of methods are beginning to show signs of promise.
[0:11:03] And obviously, we're very motivated to try and find opportunities to apply these things in the real world. And that was why we started DeepMind Health. It's quite incredible. We looked at many different application areas, but none seemed more sort of struggling with technology than healthcare. None seemed to have a margin for improvement that was as striking as the opportunities that we see in healthcare. And so let me tell you a little bit about why I think digitalization is absolutely critical when it comes to delivering safe and high-quality care.
[0:11:33] This is an academic overview, a diagram, a theoretical view of what it looks like for a patient to go through a single pathway. In this case, the patient has been admitted with an upper quadrant pain and is actually at risk of appendicitis. And these are all of the steps that need to happen in the hospital in order for that patient to be discharged in a healthy way.
[0:11:56] But of course, the challenge is, as we've just heard on the panel, that patient also has all sorts of other underlying comorbidities which lead to other complications, which mean that this patient's actually on four or five different pathways. In this case, they've had an allergic reaction to the antibiotics. They've also contracted an acute kidney injury, and there's been a wound infection. And so this doctor now has to manage this patient with so many different pathways, but of course, she's also responsible for eight other patients.
[0:12:24] And each of those patients has scores of different pathways. And so our doctor now has to think about the clinical best practice for each one of those pathways and coordinate the intervention accordingly. So all of the steps need to be taken in sequence. And that's incredibly challenging for a single human being to be able to remember. But of course, she's not acting alone. She needs to coordinate with the porter.
[0:12:48] or the nurse or the consultant or the manager. And each one of these members of staff needs to be called on at the right time to participate in this workflow. And so trying to streamline that process is incredibly complicated. If you think about it in theory, there is in fact an optimal way to carry out this sequence of activity, be it initiating the blood test at the right time or carrying out a chest exam.
[0:13:14] Or changing that cannula just at the right time and not letting it be delayed. And so the numbers on the diagram that you can see here actually reflect the theoretical optimal pathway that this patient or these number of patients should experience all of this care. And how to allocate the tasks both to digital systems but also to humans that carry out those tasks. So in some sense this gives you an intuition for the complexity of the system.
[0:13:39] Each one of these members of staff is operating in a fairly self-organizing way. And in fact, that just represents what's happening within the hospital ecosystem. In the broader ecosystem, the patient is of course going to the pharmacy, going to the emergency department, visiting her general practitioner, and so on and so forth. And so when you actually map the entire system complexity, it's much easier to see how overwhelming it is. And I think this is the role that artificial intelligence has the potential.
[0:14:07] But of course, if you can't get access to the data and if you can't collect and store the data in exactly the right format, it becomes impossible to intervene in the system in a coherent way. And so the primary thing that we're working on at the moment is to expose all of the data in an open and interoperable standard. And that should mean that all of the databases that exist behind the system speak the same language.
[0:14:36] We think that's what will enable the transformative clinical applications that the panel have talked about, front-end apps that are under 30s, small-medium businesses are developing. But of course, we also think that it will enable us to deliver actionable analytics and derive novel insights. And this, we think, will be powered by advanced AI research. So let me tell you about some of these things. First of all, how do we build an open innovation ecosystem? Well, in the hospital today,
[0:15:03] At least in the NHS, there are on average 160 different database languages, many of which don't speak to one another. And so it's really difficult to share data across the entire system. It's very siloed, and that means that each clinician has to go and log on to a very painful, slow, desktop-based system in order to access the different piece of information they're looking for at any given time.
[0:15:26] And so what we've been building is a unified health database that is exposed or re-presented back to the various different applications that we might like to build. So you can imagine through an open application programming interface that we can deliver cutting-edge clinical systems, a patient portal, and in fact all sorts of real-time analytics, both for research and for direct quality service improvement.
[0:15:53] So one of the clinical systems that we've been developing using this data infrastructure is called Streams. Let me tell you a little bit about Streams by introducing you to our patient, Robert. Robert's just been admitted yesterday with right upper quadrant pain. And here we see our nurse, Jamie, on his mobile phone taking Robert's temperature. He's now able to enter Robert's temperature automatically on his mobile device.
[0:16:18] And immediately that generates a raised news score, an early warning score that sends an alert to our junior doctor, in this case, Sarah, who while she's on her way to visit Robert, knowing that this patient is in need of urgent attention because he's at risk of deterioration, she's able to look him up across the system. And our system now searches across any consultant or ward or specialty. And all she has to do is type in Robert's name at the top.
[0:16:45] And she's presented with an overview, the first stream, all of the urgent clinical information that she needs to digest before she sees Robert. So in this case, we see that Robert has a number of allergies. What has he been admitted under for the current episode? What are his past diagnoses or past procedures? And what kind of medication is he on and what have the past observations have been? So now, Sarah is able to quickly glance through this information and get a better sense for Robert's condition.
[0:17:13] If we swipe across to a new stream of data, we can now see in chronological order every unit of clinical activity that has happened to Robert during his admission. We can see that in the last couple of minutes, Robert has just been activated for a sepsis alert. It looks like he's at risk of an infection, and we now have a 25-minute timer counting down to alert the right clinical intervention. If she scrubs down, she can see that his bloods have been ordered and they're just about in the clinic.
[0:17:41] If she swipes across, she can now see another stream of data, a different perspective. Here we have all of the blood results, the ureas, electrolytes. Here she's tapping on the creatinine value because she knows that's important for acute kidney injury. And she can see that in the last 24 hours, he's actually seen a spike over the 24-hour baseline.
[0:18:00] So she's been able to review all of this information, including the radiology images where she's able to pinch and zoom and share it with a colleague. And all of this information has been digested before she actually gets to Robert's bedside. And now she can spend much more time in direct care examining him face to face. Now that she's had an opportunity to do that, she can start to think about escalating his care. It looks like he has an appendicitis. So she immediately texts her consultant surgeon and alerts, in this case, Maddy,
[0:18:28] Here you can see at the top.
[0:18:44] Maddy is able to tap on, prepare this patient for surgery, and immediately a whole series of tasks are disseminated across the organization, whether it's to the porter, or to the healthcare assistant, or to the nurse, or the junior doctor. We know that we have to make this patient nil by mouth, we have to get intravenous fluids access, we need to administer antibiotics, and we need to get the consent of the patient, etc., etc. And now after the patient's operation, Maddy is able to continue to monitor
[0:19:14] But of course, all of this is happening at the micro level. A whole series of alerts and notifications are sent to individual clinicians so that they act at the right time, ideally preventatively. But of course, what we really want to be able to do is better manage that pathway. How do we program that pathway so that we can send a whole series of alerts to the different members of staff at the right time?
[0:19:41] Here we see a desktop-based view that a senior clinician would use to set up a series of automatic blood tests, automatic reminders to trigger imaging, for example, to allocate and assign responsibility to various different members of staff, so the right person takes responsibility. We know that the vast majority of patient error actually happens because of poor coordination.
[0:20:04] When a patient slips through the net, someone's forgotten about them. And so what we really want to do is set in place a safety net that sits underneath that patient pathway so that the right person is able to catch the patient at the right time in advance. And of course, we're able to provide the appropriate guidelines or clinical best practice as prompts and reminders to the members of staff when they're carrying out a particular procedure.
[0:20:29] And of course, this represents a great opportunity for machine learning systems, some of which we're beginning to start work on for the future. But underneath all of this sits the conventional analytics dashboard. So here we have a view of all of the wards in the hospital. We can look at all sorts of useful information, like the number of beds we have available, the number of staff we have on call, how many early warning alerts have been triggered, how many tasks are open, how many cannulas need to be replaced.
[0:20:57] We can take a different stream of data, not just looking at the ward. We can look at a specific condition. In this case, we're going to look at appendicitis. Which of patients do I have in admission today who do have appendicitis?
[0:21:08] Do they need their lines changing? Have they been taken care of at the right time? Which prescriptions have been allocated? So now we have a meta view by condition or by ward or by specialty or by member of staff. This is the data that we need to do quality service improvement. And of course, it's exactly the kind of information that's going to be interesting to the hospital CEO. Am I hitting my targets? How many beds do I have available? How do I redirect my ambulances to my colleagues?
[0:21:33] And this, from a minister perspective, I think is exactly the sort of thing that we need to build in order to deliver an integrated care model, where we can actually see how all of our hospitals are performing in aggregate, where our ambulances are, what's happening at primary care, and so on and so forth. Artificial intelligence is obviously why we founded the company, and that's what we're really motivated by. So let me tell you very briefly about some of the things that we're working on to apply machine learning today.
[0:22:02] One of the big challenges that we face in the UK is a lack of radiologists. And so we've been working very closely with the Moorfields Eye Hospital to see if we can help to make sense of some of their three-dimensional eye scans. In this case, what you're seeing here is an OCT scan, optimal coherence tomography. It creates a very rich picture of the pathologies that you may have in your eye. And we're particularly interested in macular degeneration and diabetes.
[0:22:28] One of the challenges, as many of you will be aware, is that something like 30% or so in the UK and in the US of our population are actually obese. And if you are obese, then you are massively more at risk of diabetes. And if you do have diabetes, in fact, you're 25 times more likely to suffer some kind of sight loss.
[0:22:49] And yet the remarkable thing with this sort of sight loss is that 98% of the most severe forms of sight loss are in fact entirely preventable if you can get in there as early as possible. So we're training machine learning systems to accurately identify where in the image the particular diagnosis actually sits. And in doing so, we can actually do that in seconds rather than hours or in some cases days. And we can send an alert to the right clinician to contact the patient and provide sight-safe.
[0:23:19] They're saving preventative treatment in the form of injections and other forms of early intervention. Another example that we're working on, this time with the University College London Hospital, is that of CT scans. It takes a super qualified intelligent human being on the order of four to five hours to prepare a patient for radiotherapy treatment. In this case, they'll actually sit at a computer with a pen and label up each one of the pixels or voxels where the radiotherapy needs to be targeted.
[0:23:48] This is a very painful process. It takes a very long time. And in fact, in the UK, there's actually a backlog because we don't have enough clinicians. We don't have enough radiologists. There's approximately a 10x shortage of radiologists in the system. So if we can do this in mere seconds, then we can provide a prompt or an aid to the clinician to then validate.
[0:24:11] The labelled areas that we provide to that radiographer and therefore speed up the process in preparation for radiotherapy planning. So in summary, I sort of wanted to leave you with the idea that we're very much committed to working on some of the toughest global problems. We build transformational machine learning technologies and we've made incredible progress across the field over the last six years or so. But we really recognise the obligation and responsibility that we have on us to steward the technologies
[0:24:43] And we invite everybody to contribute to that challenge and support us, hold us accountable, and help us to do a better job of effectively improving the systems that we all care about. We recognize also that we'll need to be as innovative and as speedy with our governance as we have been so far with our technologies. And we're doing a number of things that you can see on our website.
[0:25:05] that are trying to push the envelope on this, if you like, and trying to invent novel forms of governance that can respond quickly to the kinds of technologies that we're developing. Thank you very much for your time. I appreciate it very much. Thank you, Mustafa. May I ask you one thing before you go? It is such an exciting future that you're talking about, and yet, in many hospitals, the notes are still on paper rather than electronic. It feels a world away.
[0:25:35] Yeah, I think that's a really good question, particularly in the UK. I think much of the incredible work that our frontline nurses and doctors do is using pen and paper. And in fact, pen and paper is what's responsible for patients slipping through the cracks, care not being coordinated on time, notes being lost. And the key issue is that we aren't able to act preventatively when we're using paper, because only the individuals...
[0:25:59] Who go and access that paper can actually do something about it. What we want to do is to digitize the system so that we can act proactively and get in there as early as possible. But you'd have to do the building blocks of all the individual systems before you can really...
[0:26:12] You know, achieve the full potential of it. Absolutely. There are enormous number of really innovative developers out there, and we want to do our best to make the data available to as many other developers as possible so that the hospital commissioners can actually create a competitive, a genuinely competitive ecosystem around the development of these new technologies. Okay. It is very exciting. Thank you so much. Good luck with your work. Mustafa Suleiman. Thank you.