1 00:00:07,600 --> 00:00:11,360 An incredible revolution is happening in the world 2 00:00:11,360 --> 00:00:12,720 of health care 3 00:00:12,720 --> 00:00:15,880 that some claim could be the cure that doctors and patients 4 00:00:15,880 --> 00:00:17,520 desperately need. 5 00:00:17,520 --> 00:00:20,160 Artificial intelligence, or AI. 6 00:00:20,160 --> 00:00:22,800 Basically, making computers think. 7 00:00:24,200 --> 00:00:27,280 Artificial intelligence could really revolutionise health care. 8 00:00:27,280 --> 00:00:30,240 Allow us to understand our patients better, 9 00:00:30,240 --> 00:00:34,360 to develop new diagnostic tools, even new treatments. 10 00:00:34,360 --> 00:00:37,800 All of us could benefit from this change, which is driven 11 00:00:37,800 --> 00:00:40,400 by the private tech industry. 12 00:00:40,400 --> 00:00:43,160 They're saying that we're going to do with health care what Google 13 00:00:43,160 --> 00:00:44,720 did with information. 14 00:00:45,760 --> 00:00:50,720 One grand ambition is to get AI computers to diagnose diseases 15 00:00:51,720 --> 00:00:55,120 as well as a human doctor, or better. 16 00:00:55,120 --> 00:00:58,480 This could relieve pressure from the NHS and help bring 17 00:00:58,480 --> 00:01:02,880 good health care to parts of the world where doctors are scarce. 18 00:01:02,880 --> 00:01:04,480 So what's not to like? 19 00:01:06,080 --> 00:01:09,440 But the race to get ahead in the market involves 20 00:01:09,440 --> 00:01:11,560 a major culture clash. 21 00:01:11,560 --> 00:01:14,600 You have the technology world, the Silicon Valley, 22 00:01:14,600 --> 00:01:18,440 move fast, break stuff, and then you have the diligent, 23 00:01:18,440 --> 00:01:21,800 evidence-based, do no harm health care system. 24 00:01:23,040 --> 00:01:28,040 Some even see AI's roll-out as potentially dangerous. 25 00:01:30,840 --> 00:01:33,920 how to regulate it, we don't know how to prove it's effective. 26 00:01:33,920 --> 00:01:38,360 Just because we can do something doesn't mean we should. 27 00:01:38,360 --> 00:01:40,920 The stakes could not be higher, and this is a revolution 28 00:01:40,920 --> 00:01:43,040 that is happening right now. 29 00:01:43,040 --> 00:01:46,480 A seismic disruption for the NHS is on the cards 30 00:01:46,480 --> 00:01:51,160 and, in a blind trial, we will see AI being put to the test. 31 00:01:51,160 --> 00:01:54,480 So can the machines beat the doctors, and should we trust 32 00:01:54,480 --> 00:01:56,200 them if they can? 33 00:02:09,280 --> 00:02:12,960 The British health system, with its easily lost handwritten 34 00:02:12,960 --> 00:02:16,080 notes, isn't exactly renowned for its IT. 35 00:02:18,360 --> 00:02:21,160 I haven't seen one of these for a long time! 36 00:02:21,160 --> 00:02:26,200 But then, I don't work for the NHS, which is the biggest purchaser 37 00:02:26,280 --> 00:02:29,400 of new fax machines in the entire world. 38 00:02:29,400 --> 00:02:34,400 And sometimes our beloved NHS really feels every bit 70 years old. 39 00:02:35,840 --> 00:02:38,680 But maybe not for much longer. 40 00:02:38,680 --> 00:02:42,320 And that is thanks to artificial intelligence, 41 00:02:42,320 --> 00:02:46,240 using algorithms to get computers and machines to automate some 42 00:02:46,240 --> 00:02:49,240 of the brain work usually done by doctors. 43 00:02:52,280 --> 00:02:56,440 There is, in fact, a health tech gold rush going on. 44 00:02:56,440 --> 00:03:01,440 Loads of new smart apps promise to help us keep fit and well, 45 00:03:01,440 --> 00:03:04,280 and the corporate giants are pouring vast sums 46 00:03:04,280 --> 00:03:05,440 of money into AI. 47 00:03:07,560 --> 00:03:12,400 Apple wants to use your phone sensors to monitor your health. 48 00:03:12,400 --> 00:03:16,680 Google's Deep Mind uses AI to spot serious eye disease. 49 00:03:18,280 --> 00:03:21,480 IBM is working on cancer diagnosis. 50 00:03:23,640 --> 00:03:27,600 The profits are potentially eye-watering, but some people 51 00:03:27,600 --> 00:03:29,520 are seriously worried. 52 00:03:29,520 --> 00:03:32,920 In science, you don't believe anything until it's proven 53 00:03:32,920 --> 00:03:36,200 and, in medicine, you're especially cautious. 54 00:03:36,200 --> 00:03:40,840 This isn't some online shopping suggestions or AI recommending 55 00:03:40,840 --> 00:03:45,720 you a film you might like to watch - this can be life or death. 56 00:03:45,720 --> 00:03:49,400 But in the very competitive tech industry, it's quite normal 57 00:03:49,400 --> 00:03:52,600 for companies to race to get their products out. 58 00:03:57,080 --> 00:04:00,960 To witness this phenomenon from the inside, we've had unique 59 00:04:00,960 --> 00:04:04,160 access to one tech company in particular - 60 00:04:04,160 --> 00:04:05,280 Babylon Health. 61 00:04:07,560 --> 00:04:11,720 Based in London, Babylon have hired an international team of top 62 00:04:11,720 --> 00:04:16,440 computer scientists and doctors with a mission to use AI and mobile 63 00:04:16,440 --> 00:04:20,840 phones to transform the way we access GPs. 64 00:04:22,840 --> 00:04:27,240 British entrepreneur Dr Ali Parsa is their CEO. 65 00:04:28,960 --> 00:04:31,200 We started Babylon to see if you can make 66 00:04:31,200 --> 00:04:34,640 health care accessible, affordable and put it in the hands of every 67 00:04:34,640 --> 00:04:36,560 human being on Earth. 68 00:04:36,560 --> 00:04:39,960 To take most of the health care most people need and deliver 69 00:04:39,960 --> 00:04:42,200 it to them on devices they already have. 70 00:04:45,640 --> 00:04:49,920 2018 is set to be a big year for Babylon, who want to prove 71 00:04:49,920 --> 00:04:53,840 that their artificial intelligence, programmed to understand symptoms 72 00:04:53,840 --> 00:04:57,800 and offer health advice, is a match for human doctors. 73 00:04:59,240 --> 00:05:01,720 We're saying that we are going to create a machine 74 00:05:01,720 --> 00:05:05,560 that can eventually hopefully tell you about your disease as accurately 75 00:05:05,560 --> 00:05:09,120 as a doctor can, that can give you your treatment better 76 00:05:09,120 --> 00:05:13,480 than most doctors can, that can look faster, 77 00:05:13,480 --> 00:05:17,320 cheaper, at variations of symptoms, billions of variations of symptoms, 78 00:05:17,320 --> 00:05:20,360 as well as if you had one of the best doctors in the world 79 00:05:20,360 --> 00:05:21,360 in your pocket. 80 00:05:25,440 --> 00:05:29,320 That's the big vision, but you can use Babylon's 81 00:05:29,320 --> 00:05:31,240 smartphone app right now. 82 00:05:31,240 --> 00:05:35,000 Like others on the market, such as Ada and Your.MD, 83 00:05:35,000 --> 00:05:39,680 it uses AI to let you check your symptoms when you feel unwell. 84 00:05:41,280 --> 00:05:44,120 And if you want, it will book you in to see 85 00:05:44,120 --> 00:05:46,480 one of their doctors privately. 86 00:05:46,480 --> 00:05:50,040 So I can see you've completed our AI assessment. 87 00:05:50,040 --> 00:05:53,560 And in London, there's a scheme that offers Babylon's AI 88 00:05:53,560 --> 00:05:56,920 and its doctors on the NHS, entirely free. 89 00:06:00,320 --> 00:06:04,200 This is Lillie Road in Fulham, west London. 90 00:06:04,200 --> 00:06:08,360 To look at it, you'd have little clue that this nondescript GP 91 00:06:08,360 --> 00:06:11,920 surgery is at the centre of a hi-tech venture. 92 00:06:13,240 --> 00:06:17,400 In late 2017, it partnered with Babylon to set 93 00:06:17,400 --> 00:06:21,280 up a new NHS GP service - 94 00:06:21,280 --> 00:06:22,920 GP At Hand. 95 00:06:25,120 --> 00:06:28,640 Any adult who lives or works in inner London can download 96 00:06:28,640 --> 00:06:33,320 Babylon's GP At Hand app and use it to switch from their normal GP 97 00:06:33,320 --> 00:06:36,120 to being registered here at Lillie Road. 98 00:06:36,120 --> 00:06:38,520 So many Londoners have done this that this is now one 99 00:06:38,520 --> 00:06:41,760 of the fastest-growing GP surgeries in the country 100 00:06:41,760 --> 00:06:45,480 with some 30,000 patients. 101 00:06:45,480 --> 00:06:49,960 But most of those people will never actually come here to see a doctor, 102 00:06:49,960 --> 00:06:52,760 and that's because the only way to initiate contact with 103 00:06:52,760 --> 00:06:55,120 GP At Hand is online. 104 00:06:55,120 --> 00:06:58,280 Although appointments are available at one of their clinics 105 00:06:58,280 --> 00:07:01,000 if you really need to see a doctor face-to-face. 106 00:07:03,440 --> 00:07:07,440 We can be open 24 hours a day, seven days a week. 107 00:07:07,440 --> 00:07:12,440 That is three times as much coverage as your GP has for exactly the same 108 00:07:12,920 --> 00:07:16,720 price and, more importantly, often, you have to wait days or weeks 109 00:07:16,720 --> 00:07:20,040 to see a GP. With us, you're seen within minutes. 110 00:07:20,040 --> 00:07:23,600 That is the beauty, the amazing thing about technology. 111 00:07:27,520 --> 00:07:31,320 Dan Smith is one of many who downloaded the app. 112 00:07:31,320 --> 00:07:33,960 One of the big restrictions that people have around primary care 113 00:07:33,960 --> 00:07:35,280 is about time. 114 00:07:35,280 --> 00:07:38,080 I run my own business, I have a two-year-old daughter, 115 00:07:38,080 --> 00:07:39,720 I don't have time. 116 00:07:39,720 --> 00:07:43,600 So being able to just pick the phone up and have an appointment 117 00:07:43,600 --> 00:07:46,600 whenever you want to I think's really important. 118 00:07:48,400 --> 00:07:50,960 Using GP At Hand is pretty straightforward. 119 00:07:50,960 --> 00:07:54,160 You just put in, "I have a headache," and then it will just 120 00:07:54,160 --> 00:07:55,960 take you through a list of questions. 121 00:07:57,920 --> 00:08:02,040 The app uses AI to check patients' symptoms and advise 122 00:08:02,040 --> 00:08:06,280 whether they need to see a doctor - a process called triage. 123 00:08:07,400 --> 00:08:09,440 If it says "See a doctor," you can just go straight 124 00:08:09,440 --> 00:08:12,600 in and have a video appointment, typically within ten minutes. 125 00:08:14,040 --> 00:08:16,080 You said that you're not feeling very well. 126 00:08:16,080 --> 00:08:18,720 I can see that you've used the triage chat today. 127 00:08:18,720 --> 00:08:23,680 Dr Zahra Damji is one of around 200 GPs employed by Babylon in the UK. 128 00:08:25,000 --> 00:08:28,440 Could you just tell me a bit more about how you've been feeling? 129 00:08:28,440 --> 00:08:32,160 I'm able to spend a lot of time seeing patients digitally, 130 00:08:32,160 --> 00:08:34,800 perhaps more so than I would have done if I was just working 131 00:08:34,800 --> 00:08:37,800 within that traditional model, and actually having more patient 132 00:08:37,800 --> 00:08:41,880 contact means that, as a GP, I'm happier because, ultimately, 133 00:08:41,880 --> 00:08:44,640 what I want is to see patients. 134 00:08:44,640 --> 00:08:49,640 Babylon's GP At Hand app not only gives patients triage advice, 135 00:08:50,080 --> 00:08:54,480 it also provides diagnostic information to the online GP. 136 00:08:55,880 --> 00:08:58,520 The information comes up in a list format for us, 137 00:08:58,520 --> 00:09:01,000 so we try to work out from that information 138 00:09:01,000 --> 00:09:03,480 what we predict the diagnosis for them could be, 139 00:09:03,480 --> 00:09:05,560 but we're already one step ahead because we've got 140 00:09:05,560 --> 00:09:07,880 all of that information in front of us. 141 00:09:07,880 --> 00:09:10,520 It also allows us to focus our questions better. 142 00:09:10,520 --> 00:09:12,960 Have a feel of your neck for me. If you just kind 143 00:09:12,960 --> 00:09:15,600 of put your finger across your jawline, 144 00:09:15,600 --> 00:09:17,000 just press all the way round. 145 00:09:17,000 --> 00:09:20,280 It's a slightly strange experience, speaking to a doctor on the phone 146 00:09:20,280 --> 00:09:22,320 rather than face-to-face. 147 00:09:22,320 --> 00:09:25,360 Stop. Stop, keep it right there. 148 00:09:25,360 --> 00:09:27,600 OK, that's quite a good view. 149 00:09:27,600 --> 00:09:31,840 Dan has certainly put the service through its paces. 150 00:09:31,840 --> 00:09:35,520 I suddenly started feeling a strange feeling in my face. 151 00:09:35,520 --> 00:09:39,640 I looked in the mirror and half my face had dropped. 152 00:09:39,640 --> 00:09:43,640 And the immediate fear you have is, "Am I having a stroke?" 153 00:09:43,640 --> 00:09:47,640 I can hear your speech is slightly affected because of your mouth. 154 00:09:47,640 --> 00:09:51,160 I had a video call with a doctor within three minutes, 155 00:09:51,160 --> 00:09:54,200 and he did a couple of basic tests, and he quickly eliminated 156 00:09:54,200 --> 00:09:56,040 it being a stroke and, because of the way 157 00:09:56,040 --> 00:09:58,640 that it was presenting, that it was Bell's palsy. 158 00:09:58,640 --> 00:10:02,520 It tends to get better. It's quite a dramatic problem 159 00:10:02,520 --> 00:10:03,720 to have at the moment. 160 00:10:03,720 --> 00:10:06,600 It meant I didn't have to rush down to A&E and I could just 161 00:10:06,600 --> 00:10:09,280 take some steroids and some painkillers. 162 00:10:09,280 --> 00:10:14,280 Being able to speak to a doctor and have them to reassure 163 00:10:14,520 --> 00:10:17,640 you in minutes I thought was wonderful. 164 00:10:21,800 --> 00:10:25,400 Talking to a doctor via your smartphone certainly feels 165 00:10:25,400 --> 00:10:29,680 a lot more efficient than schlepping down to the local GP and waiting 166 00:10:29,680 --> 00:10:31,200 in a queue. 167 00:10:31,200 --> 00:10:35,480 But that is not artificial intelligence. 168 00:10:35,480 --> 00:10:39,480 Where AI comes into all of this is in the symptom checker. 169 00:10:39,480 --> 00:10:41,720 This is the first thing that you see when you open up one 170 00:10:41,720 --> 00:10:44,600 of these kinds of app, it's where the patient types 171 00:10:44,600 --> 00:10:47,960 in their symptoms to what's known as a chatbot. 172 00:10:47,960 --> 00:10:50,840 At this point, there is no doctor involved at all. 173 00:10:50,840 --> 00:10:52,720 The computer is doing all of the thinking. 174 00:10:52,720 --> 00:10:56,600 It's working out what might be wrong with you and giving you information 175 00:10:56,600 --> 00:10:58,800 about what to do next. 176 00:10:58,800 --> 00:11:03,240 And if the AI is working well, it will filter out all of the people 177 00:11:03,240 --> 00:11:08,120 with non-serious conditions, leaving human GPs to treat 178 00:11:08,120 --> 00:11:10,960 the patients who really need their expertise. 179 00:11:14,800 --> 00:11:17,520 So how does this kind of AI work? 180 00:11:18,920 --> 00:11:22,720 Dr Saurabh Johri leads Babylon's AI team. 181 00:11:24,080 --> 00:11:28,600 What we're trying to do is essentially replicate the kind 182 00:11:28,600 --> 00:11:32,280 of decision-making process that a human doctor takes 183 00:11:32,280 --> 00:11:34,600 when they try and diagnose a patient. 184 00:11:37,360 --> 00:11:40,000 If a user is sick, what we're trying to do is reason 185 00:11:40,000 --> 00:11:43,320 about what may be the underlying condition. 186 00:11:43,320 --> 00:11:47,720 For their AI to do this, to reason, it will have to master 187 00:11:47,720 --> 00:11:49,160 three key skills. 188 00:11:50,160 --> 00:11:55,120 First, it has to become as knowledgeable as a human doctor. 189 00:11:55,240 --> 00:11:59,080 Then it must learn to talk to you, ideally in a nice, 190 00:11:59,080 --> 00:12:01,360 human-like way. 191 00:12:01,360 --> 00:12:05,400 Finally, it must be able to use all that to deduce 192 00:12:05,400 --> 00:12:06,840 what might be wrong. 193 00:12:09,560 --> 00:12:12,320 And that's what we're trying to build, to replicate kind 194 00:12:12,320 --> 00:12:15,160 of human decision-making in terms of medical diagnosis. 195 00:12:16,360 --> 00:12:18,200 It's not easy. 196 00:12:19,880 --> 00:12:24,080 Babylon's team have trained their machines to hoover up as much 197 00:12:24,080 --> 00:12:26,160 medical knowledge as possible. 198 00:12:27,760 --> 00:12:31,400 We have created systems that can read papers, 199 00:12:31,400 --> 00:12:34,640 they can read web pages, and they are much faster than us. 200 00:12:34,640 --> 00:12:39,640 Imagine you had a magical box that you could put a book 201 00:12:40,200 --> 00:12:43,280 in and it was like eating the book and understanding it. 202 00:12:44,800 --> 00:12:46,840 That's what it does. 203 00:12:46,840 --> 00:12:49,880 But hoovering information isn't enough. 204 00:12:49,880 --> 00:12:53,680 The knowledge has to be organised into a knowledge graph - 205 00:12:53,680 --> 00:12:56,400 a vast web of interconnections. 206 00:12:57,560 --> 00:13:01,360 And then the chatbot must be taught to interpret the meaning 207 00:13:01,360 --> 00:13:03,160 of what the patient tells it. 208 00:13:04,360 --> 00:13:08,600 To a machine, a word is just a string of characters, there's no further meaning to it. 209 00:13:08,600 --> 00:13:10,920 In order to determine what something means, 210 00:13:10,920 --> 00:13:14,160 it is important to learn what it is similar to. 211 00:13:16,200 --> 00:13:20,400 The AI computer tries to capture the similarity between words 212 00:13:20,400 --> 00:13:23,560 in a kind of mathematical map. 213 00:13:23,560 --> 00:13:27,000 For instance, words like headache and migraine, 214 00:13:27,000 --> 00:13:31,480 with similar meanings, go in the head region. 215 00:13:31,480 --> 00:13:34,400 If I can arrange things in a way that the computer can easily figure 216 00:13:34,400 --> 00:13:38,920 out the two things are similar, then you're taking a large step 217 00:13:38,920 --> 00:13:42,720 towards the computer actually understanding what the word means. 218 00:13:42,720 --> 00:13:47,160 A map of phrases you might say to the chatbot works in exactly 219 00:13:47,160 --> 00:13:49,120 the same way. 220 00:13:49,120 --> 00:13:54,120 Those that mean something similar belong in the same part of the map, 221 00:13:55,520 --> 00:14:00,120 which helps the chatbot decide what questions to ask you next. 222 00:14:02,440 --> 00:14:05,080 We are at ground zero today. 223 00:14:05,080 --> 00:14:08,520 I bet, in two years' time, maybe five years' time - 224 00:14:08,520 --> 00:14:10,600 but it certainly isn't anything longer than that - 225 00:14:10,600 --> 00:14:14,800 machines can speak any language in the world in any local accent 226 00:14:14,800 --> 00:14:16,800 and understand any dialect. 227 00:14:16,800 --> 00:14:18,680 That is a game changer. 228 00:14:18,680 --> 00:14:21,720 The computing powers, the data sciences, 229 00:14:21,720 --> 00:14:26,440 are now becoming sophisticated enough for us to be able to do 230 00:14:26,440 --> 00:14:31,040 things that only a few years ago we used to dream about. 231 00:14:31,040 --> 00:14:34,880 Imagine everyone having powerful clinical artificial intelligence 232 00:14:34,880 --> 00:14:35,920 in their pocket. 233 00:14:35,920 --> 00:14:37,560 So how can I help you today? 234 00:14:37,560 --> 00:14:38,960 I've got a bad headache. 235 00:14:38,960 --> 00:14:40,920 Headaches are really common... 236 00:14:40,920 --> 00:14:43,720 Since their launch, Babylon have invested tens 237 00:14:43,720 --> 00:14:46,880 of millions of pounds on research and development. 238 00:14:46,880 --> 00:14:49,000 According to the World Health Organization, 239 00:14:49,000 --> 00:14:51,840 we need at least five million more doctors. 240 00:14:53,000 --> 00:14:55,840 They're trying to edge out their competitors to be the 241 00:14:55,840 --> 00:15:00,920 go-to health app in the potentially highly lucrative global market. 242 00:15:01,520 --> 00:15:03,080 But what about the rest of the world? 243 00:15:03,080 --> 00:15:06,200 They've signed up patients from Britain to Rwanda 244 00:15:06,200 --> 00:15:09,360 and partnerships from China to Saudi Arabia. 245 00:15:12,280 --> 00:15:17,360 Such is the current buzz about AI's potential to help us do many things 246 00:15:17,480 --> 00:15:21,640 better, faster and cheaper that the British Government are keen 247 00:15:21,640 --> 00:15:24,080 to get in on the act. 248 00:15:24,080 --> 00:15:28,120 The United Kingdom will incubate a whole new industry around AI 249 00:15:28,120 --> 00:15:30,160 in health care. 250 00:15:30,160 --> 00:15:34,840 But for all the excitement, some people have serious concerns 251 00:15:34,840 --> 00:15:38,680 about how this market functions. 252 00:15:38,680 --> 00:15:43,760 I think there is an inherent risk in the hype that is required 253 00:15:44,760 --> 00:15:47,840 to attract the kind of venture capital that these companies 254 00:15:47,840 --> 00:15:52,040 who are developing novel AI technologies are reliant on to get 255 00:15:52,040 --> 00:15:54,800 themselves off the ground, in many cases. 256 00:15:54,800 --> 00:15:58,800 There is a risk that behind the hype there won't turn out 257 00:15:58,800 --> 00:16:00,720 to be enough substance. 258 00:16:00,720 --> 00:16:03,480 I'm under no doubt that this potentially could really 259 00:16:03,480 --> 00:16:06,760 revolutionise health care, I'm really excited about that. 260 00:16:06,760 --> 00:16:11,400 But the real tragedy here would be if we over-promise 261 00:16:11,400 --> 00:16:13,920 and then funding dries up and we're unable to develop 262 00:16:13,920 --> 00:16:16,600 these tools, and therefore patients miss out on them. 263 00:16:19,920 --> 00:16:23,160 Personally, I find it hard not to be excited 264 00:16:23,160 --> 00:16:27,240 about the possibilities that AI offers for radically 265 00:16:27,240 --> 00:16:29,280 transforming health care. 266 00:16:29,280 --> 00:16:32,520 But it's also really important that we don't allow the hype 267 00:16:32,520 --> 00:16:35,760 around AI, which is currently pretty massive, 268 00:16:35,760 --> 00:16:38,600 to tarnish its true potential. 269 00:16:38,600 --> 00:16:40,840 And that means that the people who are involved here 270 00:16:40,840 --> 00:16:45,760 should be really responsible about the claims that they're making. 271 00:16:45,760 --> 00:16:49,400 And what's more, we should remember that changing how people interact 272 00:16:49,400 --> 00:16:53,040 with a trusted institution like the NHS is never 273 00:16:53,040 --> 00:16:54,760 going to be easy. 274 00:17:00,360 --> 00:17:05,000 Some critics have taken aim at Babylon's GP At Hand service, 275 00:17:05,000 --> 00:17:09,400 worried less about the AI and more about its rapid growth 276 00:17:09,400 --> 00:17:13,560 and impact on normal NHS GPs. 277 00:17:13,560 --> 00:17:17,040 I'm a GP from East London, and I've come here because 278 00:17:17,040 --> 00:17:20,240 I'm furious for my patients, my family and myself 279 00:17:20,240 --> 00:17:23,720 about the privatisation in the NHS. 280 00:17:23,720 --> 00:17:28,680 In protests in London in 2018, a range of concerns are voiced. 281 00:17:29,000 --> 00:17:32,200 But we're not against artificial intelligence when it 282 00:17:32,200 --> 00:17:33,880 helps patient care. 283 00:17:33,880 --> 00:17:36,680 But we're really worried that this system is set to undermine 284 00:17:36,680 --> 00:17:38,000 front line practice. 285 00:17:38,000 --> 00:17:42,080 There are lots of good, I've no doubt, digital interventions, 286 00:17:42,080 --> 00:17:43,440 but this is not one. 287 00:17:46,480 --> 00:17:51,360 Every GP is paid a fee by the NHS for each of its patients, 288 00:17:51,760 --> 00:17:56,160 so when London GPs like Dr Jackie Applebee lose patients to 289 00:17:56,160 --> 00:17:59,120 GP At Hand, they lose funding, too. 290 00:18:03,120 --> 00:18:06,160 So the big concerns that motivated the protest against GP At Hand 291 00:18:06,160 --> 00:18:09,080 was our real fear that it's going to destabilise general practice. 292 00:18:09,080 --> 00:18:12,880 Our patients are leaving, taking their capitation fee from us 293 00:18:12,880 --> 00:18:15,920 to GP At Hand so that we can't provide the care that we want 294 00:18:15,920 --> 00:18:18,520 to provide to the whole of our patient population. 295 00:18:20,400 --> 00:18:24,640 GP At Hand capitalises on the fact that people can now register 296 00:18:24,640 --> 00:18:27,280 with practices far away from home. 297 00:18:28,600 --> 00:18:31,960 It's partly this loophole that allows Babylon to attract 298 00:18:31,960 --> 00:18:36,240 digitally savvy young Londoners away from their old doctors. 299 00:18:37,880 --> 00:18:40,960 GP leaders hadn't seen this coming. 300 00:18:42,320 --> 00:18:44,960 GP At Hand certainly burst onto the scene in London last year 301 00:18:44,960 --> 00:18:48,000 and took us all by surprise because people didn't realise 302 00:18:48,000 --> 00:18:50,440 that there was this loophole in the contractual mechanisms 303 00:18:50,440 --> 00:18:52,840 of all GP surgeries that they could adopt. 304 00:18:52,840 --> 00:18:55,920 But they did, they saw it, and they grabbed it. 305 00:18:55,920 --> 00:18:58,600 When you take the fit, well people out of a practice 306 00:18:58,600 --> 00:19:01,000 and the money goes with them, you're left with far 307 00:19:01,000 --> 00:19:02,800 less money to look after the really sick, 308 00:19:02,800 --> 00:19:05,640 frail people that leaves the practice in jeopardy. 309 00:19:05,640 --> 00:19:08,600 And if that practice goes under, what's going to happen to those patients? 310 00:19:08,600 --> 00:19:10,760 That is a real risk to the individuals 311 00:19:10,760 --> 00:19:12,160 and to their communities. 312 00:19:16,040 --> 00:19:19,800 Unease about GP At Hand, which hopes soon to expand 313 00:19:19,800 --> 00:19:23,360 nationwide, is not just about money. 314 00:19:25,240 --> 00:19:28,640 Something that really concerns me about the GP At Hand type model 315 00:19:28,640 --> 00:19:31,080 is that, on their website, they say, if you're pregnant, 316 00:19:31,080 --> 00:19:33,680 if you're mentally ill, if you've got a complex long-term 317 00:19:33,680 --> 00:19:35,960 condition, we might not be the best service for you. 318 00:19:35,960 --> 00:19:38,600 So what they're doing is they're creating a two-tier 319 00:19:38,600 --> 00:19:41,560 health care service and, if this starts to come in, and more and 320 00:19:41,560 --> 00:19:46,560 more practices start to do that, then we lose our comprehensive NHS. 321 00:19:50,360 --> 00:19:52,960 I'm excited by innovation, but that excitement 322 00:19:52,960 --> 00:19:55,600 is really tempered by anxiety. 323 00:19:55,600 --> 00:19:59,520 My worry about this innovation and this rush to embrace technology 324 00:19:59,520 --> 00:20:02,360 is that we're leaving behind the very people who stand 325 00:20:02,360 --> 00:20:04,360 to need most from the health care service. 326 00:20:04,360 --> 00:20:07,000 Those who need most are going to get least. 327 00:20:09,880 --> 00:20:12,880 You know, people don't have to agree on everything all the time, 328 00:20:12,880 --> 00:20:15,120 what matters is patients in their tens of thousands 329 00:20:15,120 --> 00:20:16,360 are agreeing with this. 330 00:20:16,360 --> 00:20:19,200 As you and I are doing this interview, every three minutes, 331 00:20:19,200 --> 00:20:22,200 a new patient is joining Babylon. 332 00:20:22,200 --> 00:20:24,680 We have many patients with mental health issues, 333 00:20:24,680 --> 00:20:25,920 many elderly patients. 334 00:20:25,920 --> 00:20:28,920 Our fastest-growing group of people are the elderly. 335 00:20:33,240 --> 00:20:37,120 Fears about the impact of AI on health care go well 336 00:20:37,120 --> 00:20:40,560 beyond suggestions that it threatens the status quo 337 00:20:40,560 --> 00:20:42,720 and could increase inequality. 338 00:20:43,880 --> 00:20:47,600 People are asking, how do we know it's safe, 339 00:20:47,600 --> 00:20:50,200 and is it appropriately regulated? 340 00:20:51,760 --> 00:20:55,360 From the first day of medical school, the thing you're told is, first, do no harm. 341 00:20:55,360 --> 00:20:57,160 And that's exactly what patients expect. 342 00:20:57,160 --> 00:20:58,440 That's the right thing. 343 00:20:58,440 --> 00:21:01,400 And then, when you bring that to the fast-moving world 344 00:21:01,400 --> 00:21:04,680 of technology, the move fast and break things culture, 345 00:21:04,680 --> 00:21:06,440 clearly, there's a clash there. 346 00:21:06,440 --> 00:21:07,840 Chest X-ray. Is it really? Yes. 347 00:21:07,840 --> 00:21:12,440 And I wouldn't blame patients and doctors and nurses for finding 348 00:21:12,440 --> 00:21:15,680 it hard to trust these technologies at the moment 349 00:21:15,680 --> 00:21:18,920 because it's hard to find out how they've been assessed 350 00:21:18,920 --> 00:21:21,640 as being safe, who has determined that they are safe. 351 00:21:24,840 --> 00:21:27,880 Now, stay with me because at this point 352 00:21:27,880 --> 00:21:30,680 I'm going to beckon you in to the arcane world 353 00:21:30,680 --> 00:21:32,720 of health care regulation. 354 00:21:33,920 --> 00:21:38,320 Apps like Babylon's are regulated by the MHRA - 355 00:21:38,320 --> 00:21:42,080 the Medicines and Healthcare products Regulatory Agency. 356 00:21:42,080 --> 00:21:45,600 But what you might be surprised to learn is that they are classified 357 00:21:45,600 --> 00:21:50,000 in the lowest possible category of medical device 358 00:21:50,000 --> 00:21:55,040 alongside things like bandages or Zimmer frames and spectacles. 359 00:21:56,160 --> 00:21:59,160 In this category, it is the manufacturers themselves 360 00:21:59,160 --> 00:22:04,200 who vouch for the effectiveness and safety of their product, 361 00:22:04,280 --> 00:22:08,240 and that means that we only have Babylon's word for it 362 00:22:08,240 --> 00:22:10,920 that their app is safe. 363 00:22:10,920 --> 00:22:15,800 Now, there are a lot of very strict rules about what AI symptom checkers 364 00:22:16,480 --> 00:22:19,000 are and aren't allowed to do if they're to qualify 365 00:22:19,000 --> 00:22:22,920 for this quite quick and easy way to launch. 366 00:22:22,920 --> 00:22:26,960 They are allowed to collect data on symptoms and to provide 367 00:22:26,960 --> 00:22:30,960 information, but they're not allowed to give any feedback 368 00:22:30,960 --> 00:22:35,680 that could replace a human medic's advice or offer anything 369 00:22:35,680 --> 00:22:38,840 that could appear to be a diagnosis. 370 00:22:42,640 --> 00:22:46,240 In the case of Babylon, whose human GPs are inspected 371 00:22:46,240 --> 00:22:49,520 and regulated just like any other UK doctor, 372 00:22:49,520 --> 00:22:52,600 its patients could be forgiven for thinking that, 373 00:22:52,600 --> 00:22:56,640 with their AI chatbot, they're consulting something 374 00:22:56,640 --> 00:22:59,680 very like an online doctor. 375 00:22:59,680 --> 00:23:01,800 Is the pain in the lower left part of your tummy 376 00:23:01,800 --> 00:23:03,840 close to the worst pain that you've ever felt? 377 00:23:03,840 --> 00:23:08,440 The chatbot does seem to be working towards a diagnosis. 378 00:23:08,440 --> 00:23:10,360 ..Best describes the pain in the lower left... 379 00:23:10,360 --> 00:23:14,320 The GP At Hand website explains the symptom checker offers health 380 00:23:14,320 --> 00:23:18,160 information and fast, accurate advice. 381 00:23:18,160 --> 00:23:21,200 But, in fact, its terms and conditions say 382 00:23:21,200 --> 00:23:24,280 the output does not constitute medical advice, 383 00:23:24,280 --> 00:23:26,680 diagnosis or treatment. 384 00:23:28,720 --> 00:23:31,480 This is the first time I've heard that somebody believes 385 00:23:31,480 --> 00:23:34,480 it's confusing and, if it is, we need to fix it and we need 386 00:23:34,480 --> 00:23:37,680 to call on the regulators to help us to come up with the language 387 00:23:37,680 --> 00:23:40,520 that is not confusing for the users. 388 00:23:40,520 --> 00:23:44,640 Artificial intelligence is a new challenge for the regulators, 389 00:23:44,640 --> 00:23:47,040 and they need to figure out how to do it. 390 00:23:47,040 --> 00:23:50,440 So there is a contradiction between what we want to do 391 00:23:50,440 --> 00:23:54,320 and what we believe people want versus what the regulatory 392 00:23:54,320 --> 00:23:56,120 environment is ready to. 393 00:23:57,360 --> 00:24:01,000 After this interview was filmed, Babylon changed their website. 394 00:24:05,320 --> 00:24:09,400 Confusion about the regulation of AI health products 395 00:24:09,400 --> 00:24:11,600 is clearly a problem. 396 00:24:11,600 --> 00:24:14,160 But what about the science? 397 00:24:14,160 --> 00:24:17,360 Do such products stand up to traditional, 398 00:24:17,360 --> 00:24:19,440 scientific scrutiny? 399 00:24:20,400 --> 00:24:24,600 These are peer-reviewed medical journals, where professional experts 400 00:24:24,600 --> 00:24:29,240 anonymously review each other's work and make sure that it's good 401 00:24:29,240 --> 00:24:33,440 science - robust, accurate and significant. 402 00:24:33,440 --> 00:24:37,760 And, every year, up to 1 million research papers are published 403 00:24:37,760 --> 00:24:42,720 in journals like this, and yet, despite the frenzy around AI 404 00:24:44,320 --> 00:24:49,200 in health care, this is the sum total of all publications to date 405 00:24:50,320 --> 00:24:55,080 on peer-reviewed clinical trials that assess real time, 406 00:24:55,080 --> 00:24:59,880 real world impacts of this technology on patients. 407 00:24:59,880 --> 00:25:02,120 That's right - one paper. 408 00:25:05,320 --> 00:25:10,280 There are a handful more where AI and human doctors are set to work 409 00:25:10,480 --> 00:25:15,000 on the same past data, but basically the AI health field 410 00:25:15,000 --> 00:25:19,560 is booming and yet there's been little scientific proof 411 00:25:19,560 --> 00:25:21,880 that these products are effective. 412 00:25:23,040 --> 00:25:26,920 But this is a reflection of how the tech industry works. 413 00:25:26,920 --> 00:25:31,360 With huge investments at stake, companies often keep their methods 414 00:25:31,360 --> 00:25:35,680 as a closely guarded secret and, as far as they see it, 415 00:25:35,680 --> 00:25:38,920 they can't afford to wait for the long-winded process 416 00:25:38,920 --> 00:25:42,200 of getting a paper reviewed and published. 417 00:25:43,600 --> 00:25:46,040 At the moment, we see a bit of a Wild West. 418 00:25:46,040 --> 00:25:51,000 We see some companies that are going about things in a way 419 00:25:51,000 --> 00:25:56,000 that is highly informed by the kind of concept of evidence and clinical 420 00:25:56,200 --> 00:26:00,960 trials, but at the other end, we have the companies that are kind 421 00:26:00,960 --> 00:26:03,680 of assuming that the problems are much simpler than 422 00:26:03,680 --> 00:26:06,840 they often turn out to be, who are delivering things 423 00:26:06,840 --> 00:26:09,600 and putting them out on the market without the evidence 424 00:26:09,600 --> 00:26:13,760 to support their efficacy or their appropriateness. 425 00:26:13,760 --> 00:26:17,040 As one commentator put it, at the moment, we have no clear 426 00:26:17,040 --> 00:26:21,400 regulator, no clear trial structure and no clear accountability process, 427 00:26:21,400 --> 00:26:23,520 so that doesn't really inspire confidence. 428 00:26:25,480 --> 00:26:29,120 When it comes to Babylon's symptom checker, it hasn't 429 00:26:29,120 --> 00:26:33,920 been independently evaluated, nor is it required to be. 430 00:26:33,920 --> 00:26:36,120 Medical director Mobasher Butt 431 00:26:36,120 --> 00:26:39,440 explains that they do testing another way. 432 00:26:40,480 --> 00:26:43,320 Our approach to testing and evaluation is very different 433 00:26:43,320 --> 00:26:47,480 to traditional models of evaluation, and it has to be. 434 00:26:47,480 --> 00:26:50,680 So, when you're working in a very fast-paced technology 435 00:26:50,680 --> 00:26:55,400 environment, there needs to be the ability to constantly 436 00:26:55,400 --> 00:26:57,360 evaluate the things that we're doing. 437 00:26:57,360 --> 00:27:02,400 So, once we have released a product, the work doesn't stop there. 438 00:27:02,480 --> 00:27:05,280 We're constantly monitoring that, so we have a whole programme 439 00:27:05,280 --> 00:27:06,880 of post-market surveillance. 440 00:27:09,720 --> 00:27:13,160 Up until now, Babylon have never publicly proved 441 00:27:13,160 --> 00:27:15,640 that their app works. 442 00:27:15,640 --> 00:27:19,080 But they say they believe in transparency 443 00:27:19,080 --> 00:27:22,320 and know it's essential to win trust. 444 00:27:22,320 --> 00:27:27,000 So, in June 2018, Ali Parsa gathers his team to make 445 00:27:27,000 --> 00:27:28,880 a big announcement. 446 00:27:28,880 --> 00:27:32,280 We can say it's brilliant, we can run all the tests internally 447 00:27:32,280 --> 00:27:33,920 that says it's brilliant. 448 00:27:33,920 --> 00:27:37,760 It's time that we run these tests also publicly. 449 00:27:37,760 --> 00:27:42,640 At the end of this month, we're going to put our AI to test. 450 00:27:42,640 --> 00:27:46,400 We're going to sit the exam of the Royal College of GPs 451 00:27:46,400 --> 00:27:47,880 related to diagnostics. 452 00:27:47,880 --> 00:27:50,240 Of course, not everything, right? 453 00:27:50,240 --> 00:27:53,440 The second thing we're going to do is be blind-tested 454 00:27:53,440 --> 00:27:55,040 with a set of questions. 455 00:27:55,040 --> 00:27:58,080 Those are the symptoms that a patient will present 456 00:27:58,080 --> 00:28:01,920 itself to our machine, and we're going to test the machine 457 00:28:01,920 --> 00:28:04,120 against those questions. 458 00:28:04,120 --> 00:28:05,600 Hopefully, we perform well. 459 00:28:05,600 --> 00:28:09,800 Otherwise, we have to fire our AI team and get another team coming in! 460 00:28:12,080 --> 00:28:15,760 Babylon intends to publicly demonstrate that the latest model 461 00:28:15,760 --> 00:28:19,560 of their AI chatbot is a match for human doctors. 462 00:28:22,520 --> 00:28:25,040 If we do this, we are it. 463 00:28:28,120 --> 00:28:32,320 So, in a few weeks' time, when they pit the chatbot's 464 00:28:32,320 --> 00:28:36,240 triage and diagnostic skills against human GPs, 465 00:28:36,240 --> 00:28:38,560 will the AI succeed? 466 00:28:39,760 --> 00:28:43,520 And will sharing the evidence with the world be enough to secure 467 00:28:43,520 --> 00:28:46,080 the trust they need to stay ahead? 468 00:28:52,320 --> 00:28:54,560 One British company in the race to 469 00:28:54,560 --> 00:28:58,440 transform health care have already submitted their AI 470 00:28:58,440 --> 00:29:03,320 to a full, peer-reviewed trial by an independent research organisation. 471 00:29:05,120 --> 00:29:08,760 Kheiron Medical aim to use AI to revolutionise 472 00:29:08,760 --> 00:29:10,600 breast cancer diagnosis. 473 00:29:12,200 --> 00:29:15,280 Because their target customers are hospitals, 474 00:29:15,280 --> 00:29:20,120 they've taken a cautious approach, rigorously testing their AI 475 00:29:20,120 --> 00:29:22,840 before it's let loose on patients. 476 00:29:24,800 --> 00:29:28,280 We've been working on this for about two years now, 477 00:29:28,280 --> 00:29:32,440 acquiring data from clinical partners, building algorithms 478 00:29:32,440 --> 00:29:33,960 and testing them. 479 00:29:33,960 --> 00:29:37,600 We want to make sure we get it right because, 480 00:29:37,600 --> 00:29:40,080 in the end, if our technology is going to be in a situation 481 00:29:40,080 --> 00:29:43,280 where it's helping make life or death decisions, 482 00:29:43,280 --> 00:29:47,520 I think it's the only sort of ethical way to build 483 00:29:47,520 --> 00:29:49,160 this kind of technology. 484 00:29:50,920 --> 00:29:54,880 Kheiron's system that aims to use AI to detect cancer 485 00:29:54,880 --> 00:29:59,520 in breast X-ray images called mammograms uses the cutting edge 486 00:29:59,520 --> 00:30:02,160 technique known as deep learning. 487 00:30:03,840 --> 00:30:05,200 We say, "This is an image. 488 00:30:05,200 --> 00:30:07,640 "It has cancer in it. 489 00:30:07,640 --> 00:30:08,960 "In fact here's a million images. 490 00:30:08,960 --> 00:30:10,280 "These ones have cancer in. 491 00:30:10,280 --> 00:30:12,360 "Here's a million images that don't have cancer. 492 00:30:12,360 --> 00:30:15,040 "You figure out what the features are." 493 00:30:15,040 --> 00:30:19,960 With enough good image data, deep learning AI will establish 494 00:30:19,960 --> 00:30:23,320 its own methods for spotting cancers. 495 00:30:23,320 --> 00:30:27,320 Once you've actually trained it, you can then test it. 496 00:30:27,320 --> 00:30:30,040 That, and data that your model has never seen before, 497 00:30:30,040 --> 00:30:32,320 never been tested on, never been trained on, and that's 498 00:30:32,320 --> 00:30:34,880 what are used to validate your model. 499 00:30:36,720 --> 00:30:40,360 Mammograms are very information-dense images. 500 00:30:40,360 --> 00:30:42,200 Millions and millions of pixels. 501 00:30:42,200 --> 00:30:45,240 And, sometimes, the abnormality can just be a few pixels. 502 00:30:45,240 --> 00:30:49,120 Not only are you trying to find small amounts of pixel information 503 00:30:49,120 --> 00:30:52,480 within a million-pixel image, the proportion of women 504 00:30:52,480 --> 00:30:55,000 who actually have cancer that come to a screening programme 505 00:30:55,000 --> 00:30:57,000 is very small. It's about 1%. 506 00:30:57,000 --> 00:30:59,960 So, it's like trying to look for a needle within a haystack 507 00:30:59,960 --> 00:31:01,280 within a haystack. 508 00:31:03,720 --> 00:31:06,360 In Britain's breast screening programme, 509 00:31:06,360 --> 00:31:11,280 each mammogram is usually read by two trained experts to minimise 510 00:31:11,280 --> 00:31:13,400 the risk of missing cancers. 511 00:31:14,720 --> 00:31:18,000 If Kheiron's AI can do some of that work, 512 00:31:18,000 --> 00:31:23,040 it will save doctors precious time, and that could be life saving 513 00:31:23,720 --> 00:31:28,280 for the one in eight British women diagnosed with breast cancer. 514 00:31:30,400 --> 00:31:35,080 Breast surgeon Dr Liz O'Riordan, who advises Kheiron, 515 00:31:35,080 --> 00:31:37,360 speaks from her own experience. 516 00:31:38,520 --> 00:31:40,320 A couple of years ago, I was diagnosed with 517 00:31:40,320 --> 00:31:41,880 breast cancer myself. 518 00:31:41,880 --> 00:31:45,960 I've seen first-hand just how awful the full gamut of breast cancer 519 00:31:45,960 --> 00:31:47,800 treatment could be. 520 00:31:47,800 --> 00:31:50,760 If a AI and technology can reduce the time it takes for women 521 00:31:50,760 --> 00:31:53,280 to get their results, it means they get their biopsies 522 00:31:53,280 --> 00:31:55,120 quicker, they get their cancer diagnosis quicker, 523 00:31:55,120 --> 00:31:57,360 they start treatment quicker. 524 00:31:57,360 --> 00:32:01,000 Kheiron's clinical trial results prove their AI 525 00:32:01,000 --> 00:32:04,160 is outperforming humans. 526 00:32:04,160 --> 00:32:08,080 What we've managed to demonstrate with our first trial 527 00:32:08,080 --> 00:32:12,960 is that our system is both more sensitive and more specific 528 00:32:16,480 --> 00:32:19,680 at making that decision of whether there's a cancer or not 529 00:32:19,680 --> 00:32:20,720 in an image. 530 00:32:22,760 --> 00:32:27,600 So, in future, rather than two human radiologists scrutinising each 531 00:32:27,600 --> 00:32:32,480 image, one could be paired with an AI. 532 00:32:32,480 --> 00:32:37,040 Since 97% of British radiology departments are understaffed, 533 00:32:37,040 --> 00:32:39,200 this could be transformative. 534 00:32:41,040 --> 00:32:44,480 But, even with such promising trial results, 535 00:32:44,480 --> 00:32:49,560 deciding to trust AI is far from straightforward. 536 00:32:51,800 --> 00:32:54,680 to make me accept new AI technology and it's really hard 537 00:32:54,680 --> 00:32:56,680 because it's never been done before. 538 00:32:56,680 --> 00:32:59,200 I don't know what standards it should meet. 539 00:32:59,200 --> 00:33:01,880 Breast cancer is based on trials with 20, 540 00:33:01,880 --> 00:33:05,000 30 years follow-up that look at hundreds of thousands of women, 541 00:33:05,000 --> 00:33:07,000 and you don't have the time to get that follow-up to bring 542 00:33:07,000 --> 00:33:09,600 in new technology because it might change in a year - 543 00:33:09,600 --> 00:33:11,680 let alone 20 years. 544 00:33:11,680 --> 00:33:13,920 I'd want to know that it's been assessed by people 545 00:33:13,920 --> 00:33:15,960 who are independent, who have nothing to gain, 546 00:33:15,960 --> 00:33:17,840 no financial interest. 547 00:33:17,840 --> 00:33:20,360 And, if this is to save money to improve resources, 548 00:33:20,360 --> 00:33:23,240 is that money being ploughed back into NHS services instead 549 00:33:23,240 --> 00:33:26,000 of just going into the profit of the company? 550 00:33:26,000 --> 00:33:29,200 And I think it is being very sure it is patient-driven. 551 00:33:29,200 --> 00:33:32,960 And that would make me think, "OK, this is good." 552 00:33:32,960 --> 00:33:36,720 In Britain, many are recommending caution, 553 00:33:36,720 --> 00:33:40,400 but in some parts of the world, they're thinking more of 554 00:33:40,400 --> 00:33:42,200 opportunity than risk. 555 00:33:49,400 --> 00:33:54,360 This is Rwanda in Africa - one of the world's poorest countries. 556 00:34:00,440 --> 00:34:03,600 Yet, surprisingly, it's a global leader 557 00:34:03,600 --> 00:34:07,760 in experimenting with AI and tech to transform 558 00:34:07,760 --> 00:34:09,600 its citizens' health care. 559 00:34:10,600 --> 00:34:15,680 We generally have a strategy to attract proof of concept companies. 560 00:34:15,680 --> 00:34:20,160 If a company out there in the UK or the US is trying to test 561 00:34:20,160 --> 00:34:23,920 an idea - either for the first time and they are looking for a place to, 562 00:34:23,920 --> 00:34:26,040 you know, if you like, a laboratory to try it out, 563 00:34:26,040 --> 00:34:27,880 we open Rwanda for them. 564 00:34:27,880 --> 00:34:29,880 That has been our strategy. 565 00:34:29,880 --> 00:34:32,320 Be bold. Try something. Experiment. 566 00:34:32,320 --> 00:34:36,920 If it doesn't work, you know, move on and count your losses. 567 00:34:36,920 --> 00:34:40,640 This openness to being the testing ground for new tech means 568 00:34:40,640 --> 00:34:43,680 that Rwanda has welcomed in Babylon. 569 00:34:43,680 --> 00:34:45,760 Here, they're known as Babyl. 570 00:35:07,920 --> 00:35:12,320 Rwanda's 12 million people are served by fewer than 2,000 571 00:35:12,320 --> 00:35:16,160 doctors, but the vast majority of adult Rwandans 572 00:35:16,160 --> 00:35:17,640 have a mobile phone. 573 00:35:19,280 --> 00:35:23,320 If they sign up to Babyl, they can get fast triage and direct 574 00:35:23,320 --> 00:35:27,040 access to a doctor - just like they could in Britain. 575 00:35:43,960 --> 00:35:48,880 Ibrahim Misafiri lives 40 minutes' walk from his local health centre, 576 00:35:49,320 --> 00:35:51,480 which doesn't have a doctor. 577 00:36:11,000 --> 00:36:14,360 Because most Rwandans don't have a smartphone, 578 00:36:14,360 --> 00:36:19,000 people like Ibrahim call Babyl's HQ, where nurses use the same AI 579 00:36:19,000 --> 00:36:21,040 chatbot as UK users. 580 00:36:48,120 --> 00:36:51,400 Rwanda has universal health insurance. 581 00:36:51,400 --> 00:36:55,120 And, any consultation, whether in person or via Babyl, 582 00:36:55,120 --> 00:36:57,080 costs the same flat fee. 583 00:37:09,280 --> 00:37:13,640 It's the possibility of treating vastly more patients with the same 584 00:37:13,640 --> 00:37:17,000 number of doctors that makes this public health experiment 585 00:37:17,000 --> 00:37:20,680 in Rwanda so potentially significant for the world. 586 00:37:22,160 --> 00:37:26,800 In Babyl, we employ about 16 doctors and 16 nurses. 587 00:37:26,800 --> 00:37:30,280 And if we're operating at capacity, we can do about 2,000 consultations 588 00:37:30,280 --> 00:37:32,720 in one day. 589 00:37:32,720 --> 00:37:35,760 With the AI technology for the triage tool, 590 00:37:35,760 --> 00:37:38,400 if that can be given to non-clinical staff, 591 00:37:38,400 --> 00:37:40,240 are you looking at how the numbers are expanding? 592 00:37:40,240 --> 00:37:44,520 How the number of consultations and how we are using technology 593 00:37:44,520 --> 00:37:47,120 to stretch the limited resources that are available 594 00:37:47,120 --> 00:37:49,760 to be able to do more in Randa. 595 00:37:52,520 --> 00:37:54,800 Pretty impressive. 596 00:37:54,800 --> 00:37:58,640 But some people have nonetheless raised concerns that an AI system 597 00:37:58,640 --> 00:38:03,680 built and trained in the UK isn't yet much use on common tropical 598 00:38:04,200 --> 00:38:06,240 diseases like malaria, 599 00:38:06,240 --> 00:38:09,280 although Babylon are working on it. 600 00:38:09,280 --> 00:38:13,360 AIs are only as good as what you put into them. 601 00:38:13,360 --> 00:38:18,120 And incomplete training or biases like that can be very problematic. 602 00:38:22,680 --> 00:38:25,800 Yet we would do well to remember the bigger picture here. 603 00:38:25,800 --> 00:38:29,840 We are living in a world with a desperate shortage of doctors 604 00:38:29,840 --> 00:38:32,560 but where two thirds of the population owns 605 00:38:32,560 --> 00:38:34,280 a mobile phone. 606 00:38:34,280 --> 00:38:39,360 So, if this technology can put good, affordable health care within reach, 607 00:38:40,160 --> 00:38:43,120 surely the impetus is to continue working 608 00:38:43,120 --> 00:38:48,160 to make AI safer, more accurate, and more reliable. 609 00:38:54,400 --> 00:38:58,640 Back in London, the push to improve Babylon's AI continues 610 00:38:58,640 --> 00:39:00,760 at breakneck speed. 611 00:39:00,760 --> 00:39:03,920 It's not long before the big event where its performance 612 00:39:03,920 --> 00:39:07,320 will be compared to human doctors. 613 00:39:07,320 --> 00:39:11,160 Thinking about the event, is a mixture of kind of 614 00:39:11,160 --> 00:39:15,320 like extreme nervousness and also excitement. 615 00:39:15,320 --> 00:39:17,400 We want to show people that we're doing well, 616 00:39:17,400 --> 00:39:20,360 so, there are lots of new ideas rolling, to try and get 617 00:39:20,360 --> 00:39:23,640 to where we need to be for that. 618 00:39:23,640 --> 00:39:26,880 Babylon are trying to advance their app, 619 00:39:26,880 --> 00:39:30,640 going beyond offering triage advice to giving out 620 00:39:30,640 --> 00:39:32,960 diagnostic information, too. 621 00:39:37,680 --> 00:39:42,480 This has always been available to Babylon's GPs, but the aim 622 00:39:42,480 --> 00:39:46,320 is to share it more widely with patients. 623 00:39:46,320 --> 00:39:49,880 Medicine is fundamentally an uncertain discipline, 624 00:39:49,880 --> 00:39:53,760 so, if a patient comes to a doctor and tells the doctor 625 00:39:53,760 --> 00:39:57,520 about their symptoms, often the doctor will have a variety 626 00:39:57,520 --> 00:40:00,000 of different conditions that they may be thinking about, 627 00:40:00,000 --> 00:40:02,640 so they are uncertain about what the underlying condition is. 628 00:40:02,640 --> 00:40:05,600 So, we need to be able to build a system which is able to update 629 00:40:05,600 --> 00:40:08,320 that belief about what the underlying condition is, 630 00:40:08,320 --> 00:40:10,520 based on new information. 631 00:40:10,520 --> 00:40:15,440 The new diagnostic AI system needs to excel at calculating 632 00:40:15,440 --> 00:40:17,640 probabilities or likelihoods. 633 00:40:19,080 --> 00:40:22,920 It works by connecting three things - symptoms, 634 00:40:22,920 --> 00:40:25,480 diseases and risk factors. 635 00:40:26,600 --> 00:40:29,840 A symptom like knee pain could be caused 636 00:40:29,840 --> 00:40:31,960 by a disease like arthritis. 637 00:40:33,000 --> 00:40:37,960 How likely this is depends on the risk factors like age. 638 00:40:38,120 --> 00:40:40,360 Each likelihood can be scored. 639 00:40:41,840 --> 00:40:45,480 And, as the chatbot learns more about the patient, 640 00:40:45,480 --> 00:40:48,880 the scores may change. 641 00:40:48,880 --> 00:40:53,880 For example, a recent fall could suggest an injury 642 00:40:54,440 --> 00:40:59,440 is causing the pain, making arthritis a less likely 643 00:40:59,880 --> 00:41:00,880 diagnosis. 644 00:41:03,360 --> 00:41:07,200 With the new AI system, the team can see exactly 645 00:41:07,200 --> 00:41:10,080 how it comes up with each diagnosis. 646 00:41:10,080 --> 00:41:12,320 Around the same magnitude, then... 647 00:41:12,320 --> 00:41:17,360 In that sense, they think it's step ahead of human doctors. 648 00:41:19,440 --> 00:41:23,680 It's transparent and it's fully accountable, so we can understand 649 00:41:23,680 --> 00:41:26,120 precisely why we've given a diagnosis. 650 00:41:26,120 --> 00:41:28,160 That's not something humans can do. 651 00:41:28,160 --> 00:41:32,640 Humans, you know, probably have a similar 652 00:41:32,640 --> 00:41:35,640 decision-making process in their minds, but it's actually 653 00:41:35,640 --> 00:41:36,880 never written down. 654 00:41:36,880 --> 00:41:40,320 The fact is that here everything is. 655 00:41:40,320 --> 00:41:42,320 There's a trail about how a particular 656 00:41:42,320 --> 00:41:44,200 diagnosis has been arrived at. 657 00:41:44,200 --> 00:41:47,840 And we can revisit that trail at any point. 658 00:41:47,840 --> 00:41:52,240 Babylon use this transparency to try to improve their AI 659 00:41:52,240 --> 00:41:55,640 through repeated rounds of testing. 660 00:41:55,640 --> 00:41:57,800 So, I'm a 65-year-old man. 661 00:41:57,800 --> 00:42:00,640 OK. And how can I help you this morning? So... 662 00:42:00,640 --> 00:42:04,280 Here, two Babylon medics are role-playing a typical GP 663 00:42:04,280 --> 00:42:05,280 consultation. 664 00:42:06,320 --> 00:42:09,960 The pretend patient has been given a secret set of symptoms, 665 00:42:09,960 --> 00:42:11,400 known as a vignette. 666 00:42:12,960 --> 00:42:17,960 So, Doctor, I've a really bad cough I've had for a couple of days. 667 00:42:18,080 --> 00:42:19,960 OK. A couple of days. OK. 668 00:42:19,960 --> 00:42:22,560 When you're coughing, are you bringing up any phlegm at all? 669 00:42:22,560 --> 00:42:24,400 Yes, pretty nasty stuff. 670 00:42:25,560 --> 00:42:29,240 After asking all the questions he needs to, the doctor reports 671 00:42:29,240 --> 00:42:31,360 his most likely diagnoses. 672 00:42:32,720 --> 00:42:34,560 My differentials included bronchitis, 673 00:42:34,560 --> 00:42:37,000 as well as more general influenza. 674 00:42:38,440 --> 00:42:41,680 The pretend patient then enters exactly the same symptoms 675 00:42:41,680 --> 00:42:46,640 into the chatbot to see what the AI thinks is wrong with him. 676 00:42:46,960 --> 00:42:51,320 What it's come out with is bronchitis and pneumonia. 677 00:42:51,320 --> 00:42:56,000 This time, the human doctor and the AI largely agree, 678 00:42:56,000 --> 00:42:58,120 but not always. 679 00:42:58,120 --> 00:43:00,800 This is case vignette number 10. 680 00:43:00,800 --> 00:43:03,000 So, all of a sudden, I've just had this really, 681 00:43:03,000 --> 00:43:05,640 just like difficulty breathing. 682 00:43:05,640 --> 00:43:07,080 When did this start? 683 00:43:07,080 --> 00:43:08,840 About a half hour ago. 684 00:43:08,840 --> 00:43:11,320 It came on quite suddenly did it, or more gradual? 685 00:43:11,320 --> 00:43:14,760 No, it just came on really sudden. OK. And... 686 00:43:14,760 --> 00:43:19,080 Taking a thorough history is one of the key ways a human GP 687 00:43:19,080 --> 00:43:21,480 reaches a diagnosis. 688 00:43:21,480 --> 00:43:23,080 Any recent operation? 689 00:43:23,080 --> 00:43:24,720 Any long haul flights recently? 690 00:43:24,720 --> 00:43:26,960 Um, I've had my hip done. 691 00:43:26,960 --> 00:43:28,160 OK. When was that done? 692 00:43:28,160 --> 00:43:29,520 A couple of days ago. 693 00:43:29,520 --> 00:43:31,760 A couple of days ago? OK. Have you noticed any 694 00:43:31,760 --> 00:43:34,920 swelling of your legs or your calves at all? 695 00:43:34,920 --> 00:43:37,200 Erm, I think, yeah, one of them is a little bit, 696 00:43:37,200 --> 00:43:38,720 a little bit swollen. 697 00:43:40,040 --> 00:43:43,960 Having asked all the questions he needs to, the human concludes 698 00:43:43,960 --> 00:43:47,920 that the patient most likely has a pulmonary embolism, 699 00:43:47,920 --> 00:43:52,800 a blood clot in the lung, which is a life-threatening condition. 700 00:43:52,800 --> 00:43:57,240 But when it is time to enter the same symptoms into the chatbot, 701 00:43:57,240 --> 00:43:59,440 it has other ideas. 702 00:43:59,440 --> 00:44:03,760 So, with this one, it's come up with costochondritis. 703 00:44:03,760 --> 00:44:08,200 The AI has missed the most serious and most probable condition, 704 00:44:08,200 --> 00:44:10,240 given the patient's history. 705 00:44:11,600 --> 00:44:15,960 Though the chatbot is currently only licensed to provide information 706 00:44:15,960 --> 00:44:20,000 to patients, and is not to be relied on as a diagnosis, 707 00:44:20,000 --> 00:44:22,640 it seems an unnerving mistake. 708 00:44:24,880 --> 00:44:27,040 This is the whole process by which science works. 709 00:44:27,040 --> 00:44:31,400 You test new data and then you again try and validate, and so 710 00:44:31,400 --> 00:44:34,840 this is the process by which all of the teams work together to try 711 00:44:34,840 --> 00:44:39,400 and refine the process, to try and bump up that accuracy. 712 00:44:39,400 --> 00:44:42,720 But what does Ali Parsa think of his AI missing 713 00:44:42,720 --> 00:44:44,840 a potentially fatal condition? 714 00:44:46,480 --> 00:44:51,480 Look, we have extraordinarily unrealistic expectations, 715 00:44:52,280 --> 00:44:56,480 in terms of the accuracy, in terms of infallibility 716 00:44:56,480 --> 00:44:58,200 of our machines. 717 00:44:58,200 --> 00:45:00,640 And that is just silly. 718 00:45:00,640 --> 00:45:05,640 For any mistakes our machine makes today, I can name you a hundredfold 719 00:45:05,920 --> 00:45:08,920 the number of mistakes that will be made by human doctors 720 00:45:08,920 --> 00:45:11,760 of the same order of magnitude. 721 00:45:11,760 --> 00:45:13,640 We have to be realistic. 722 00:45:13,640 --> 00:45:16,640 What we are saying is that when you get a diagnosis 723 00:45:16,640 --> 00:45:20,760 from a machine, you have now narrowed the field for the doctor. 724 00:45:20,760 --> 00:45:23,600 What we are not saying is go and entirely, 725 00:45:23,600 --> 00:45:28,000 completely rely on this diagnostic and don't talk to a doctor. 726 00:45:31,440 --> 00:45:35,800 There is a tenacious band of Babylon critics who are waging a noisy 727 00:45:35,800 --> 00:45:39,440 Twitter war, exposing the AI chatbot's mistakes, 728 00:45:39,440 --> 00:45:44,120 and these range from the very serious to the somewhat ridiculous. 729 00:45:44,120 --> 00:45:48,360 For instance, one user told it, "I have a swollen elbow," 730 00:45:48,360 --> 00:45:52,160 which the chatbot eventually misunderstood as "red skin 731 00:45:52,160 --> 00:45:54,200 "around the genitals." 732 00:45:54,200 --> 00:45:56,640 Cue a whole range of jokes about how it doesn't know its arse 733 00:45:56,640 --> 00:45:58,520 from its elbow. 734 00:45:58,520 --> 00:46:02,360 The thing is, for the AI to improve, it needs to gather data 735 00:46:02,360 --> 00:46:04,600 from lots of use. 736 00:46:04,600 --> 00:46:07,600 But plenty of people find the idea of it learning on the job pretty 737 00:46:07,600 --> 00:46:12,320 troubling, especially when computers are generally assumed 738 00:46:12,320 --> 00:46:14,720 to be somehow infallible. 739 00:46:16,400 --> 00:46:21,280 The future of AI-powered health care may depend on this issue 740 00:46:21,280 --> 00:46:23,280 of how far to trust it. 741 00:46:26,160 --> 00:46:29,800 With just days to go to the big event, which will reveal 742 00:46:29,800 --> 00:46:33,920 to the world how Babylon's AI measures up to human doctors, 743 00:46:33,920 --> 00:46:38,960 the testing of human versus machine intensifies. 744 00:46:39,960 --> 00:46:42,120 Feeling pretty apprehensive about it, I think, 745 00:46:42,120 --> 00:46:44,440 not just me but the team. 746 00:46:44,440 --> 00:46:46,080 There's a lot to do. 747 00:46:46,080 --> 00:46:49,880 With every test, the team helps the app to improve. 748 00:46:51,320 --> 00:46:53,560 There aren't many other companies in this domain 749 00:46:53,560 --> 00:46:54,880 that are doing this. 750 00:46:54,880 --> 00:46:57,200 And somebody has to be the first, so... 751 00:46:57,200 --> 00:46:59,240 ..fingers crossed. 752 00:46:59,240 --> 00:47:01,640 I mean, I wouldn't trust claims without evidence, 753 00:47:01,640 --> 00:47:04,080 so we kind of have to make... 754 00:47:04,080 --> 00:47:05,320 ..evidence our claims. 755 00:47:09,000 --> 00:47:12,920 Finally, it's time for Babylon's big test, 756 00:47:12,920 --> 00:47:17,320 pitching the AI against human doctors, as before, 757 00:47:17,320 --> 00:47:19,560 but now under exam conditions. 758 00:47:21,200 --> 00:47:22,920 Could you give me your age and gender, please? 759 00:47:22,920 --> 00:47:25,240 I'm male, I'm 21 years old. Mmm-hmm. 760 00:47:25,240 --> 00:47:27,960 I'm feeling really short of breath. Mmm-hmm. 761 00:47:27,960 --> 00:47:32,560 Testing takes place behind closed doors at Babylon HQ. 762 00:47:32,560 --> 00:47:37,240 It doesn't involve real patients and it hasn't been designed and run 763 00:47:37,240 --> 00:47:40,680 by an independent research organisation. 764 00:47:40,680 --> 00:47:43,920 So prior to swimming earlier today, how were you feeling? 765 00:47:43,920 --> 00:47:48,520 But the team haven't chosen the questions or seen them 766 00:47:48,520 --> 00:47:52,480 in advance and won't be judging the answers themselves either. 767 00:47:55,920 --> 00:48:00,200 Babylon's big test starts with respected independent 768 00:48:00,200 --> 00:48:05,240 collaborators supplying 100 sets of patient symptoms that a GP 769 00:48:05,600 --> 00:48:08,160 might typically encounter. 770 00:48:08,160 --> 00:48:12,360 A pretend patient takes each one, presents it to up to seven 771 00:48:12,360 --> 00:48:17,440 independent human GPs and also to Babylon's AI, 772 00:48:17,640 --> 00:48:19,000 via the chatbot. 773 00:48:20,080 --> 00:48:24,320 Both human doctors and AI must respond to each case 774 00:48:24,320 --> 00:48:26,160 with triage advice... 775 00:48:27,280 --> 00:48:29,880 ..and their most likely diagnoses. 776 00:48:32,480 --> 00:48:37,320 An independent judge will mark the anonymised answers to see 777 00:48:37,320 --> 00:48:40,880 who scores highest for safety and accuracy, 778 00:48:40,880 --> 00:48:43,960 humans or machine? 779 00:48:43,960 --> 00:48:49,000 And the AI also has a go at some Royal College of GP exam 780 00:48:49,160 --> 00:48:51,520 questions on diagnosis. 781 00:48:58,280 --> 00:49:01,920 The Royal College of Physicians, one of the world's most 782 00:49:01,920 --> 00:49:05,400 respected medical institutions. 783 00:49:05,400 --> 00:49:09,840 So, a bold venue choice for Babylon to reveal their results 784 00:49:09,840 --> 00:49:10,880 to the world. 785 00:49:15,560 --> 00:49:20,200 Two years of development, tens of millions of pounds 786 00:49:20,200 --> 00:49:25,120 of investment and a moment of truth for Babylon's AI. 787 00:49:29,280 --> 00:49:32,240 The question is, how accurate is it? 788 00:49:32,240 --> 00:49:37,280 Can it ever, ever be as accurate as a human doctor in diagnosis? 789 00:49:38,920 --> 00:49:42,360 Before some of Britain's most senior medics, 790 00:49:42,360 --> 00:49:46,800 first up is the Royal College of GPs' exam result. 791 00:49:46,800 --> 00:49:48,760 So how did Babylon's AI do? 792 00:49:49,880 --> 00:49:53,400 In the relevant tests, the average pass mark for human 793 00:49:53,400 --> 00:49:54,480 doctors is 72%. 794 00:49:57,400 --> 00:50:02,480 Babylon's AI achieved 81%. 795 00:50:06,840 --> 00:50:10,160 But this really just is the start. 796 00:50:10,160 --> 00:50:13,640 Now, Babylon's human versus machine test. 797 00:50:15,680 --> 00:50:20,080 How good were the seven independent GPs at diagnosis? 798 00:50:21,240 --> 00:50:24,840 So some did OK, 64% - not, not that great. 799 00:50:24,840 --> 00:50:27,440 But some did really, really well - so 94%. 800 00:50:27,440 --> 00:50:31,120 But if you take an average of this, it's around 80%. 801 00:50:31,120 --> 00:50:36,160 So in Babylon's 100 patient tests, the human doctors on average managed 802 00:50:37,400 --> 00:50:42,120 to give an accurate diagnosis four out of five times. 803 00:50:42,120 --> 00:50:43,680 And how did our AI do? 804 00:50:45,720 --> 00:50:50,640 So, exactly the same as the doctor average, 80%. 805 00:50:51,880 --> 00:50:53,880 And what of safety? 806 00:50:53,880 --> 00:50:57,360 How safe were the AI's triage recommendations? 807 00:50:59,320 --> 00:51:04,040 Human doctors were able to give a safe triage outcome 93.1% 808 00:51:04,040 --> 00:51:05,040 of the time. 809 00:51:06,440 --> 00:51:09,760 In terms of the AI, we were able to give a safe triage 810 00:51:09,760 --> 00:51:12,120 outcome 97% of the time. 811 00:51:21,280 --> 00:51:24,720 As the event ends, Babylon release their research 812 00:51:24,720 --> 00:51:27,360 and results for the scrutiny of all. 813 00:51:29,000 --> 00:51:33,960 So far as Ali Parsa is concerned, Babylon have now verified that AI 814 00:51:34,240 --> 00:51:37,120 matches human GPs at diagnosis. 815 00:51:38,760 --> 00:51:41,160 We absolutely now have proof. 816 00:51:41,160 --> 00:51:43,960 How would that be a claim? 817 00:51:43,960 --> 00:51:47,920 Entirely independent people, from the United States 818 00:51:47,920 --> 00:51:50,360 all the way to the Royal College of Physicians, 819 00:51:50,360 --> 00:51:54,400 created vignettes that the machine sat and passed. 820 00:51:54,400 --> 00:51:58,640 The results were independently judged by... 821 00:52:00,080 --> 00:52:02,440 ..a doctor at Harvard. 822 00:52:02,440 --> 00:52:04,160 And we published the results. 823 00:52:04,160 --> 00:52:05,960 It is all in our paper. 824 00:52:05,960 --> 00:52:10,040 And, look - if we ran these tests again and the result 825 00:52:10,040 --> 00:52:12,880 is a little bit different, who cares? 826 00:52:12,880 --> 00:52:16,720 What really matters is that we're getting better and better. 827 00:52:16,720 --> 00:52:18,560 I mean, that is mind-boggling, 828 00:52:18,560 --> 00:52:20,960 the speed at which a machine can learn. 829 00:52:20,960 --> 00:52:25,720 They can never forget, they can never have all the other 830 00:52:25,720 --> 00:52:29,560 problems that we have as humans. 831 00:52:29,560 --> 00:52:32,360 It's an absolute fact, any way they can put 832 00:52:32,360 --> 00:52:34,000 it to test now, right? 833 00:52:36,040 --> 00:52:40,120 But some scientists question Babylon's methods, 834 00:52:40,120 --> 00:52:45,120 saying the study was too small for meaningful statistical analysis, 835 00:52:45,120 --> 00:52:50,040 as well as not being fully independent or peer-reviewed. 836 00:52:50,480 --> 00:52:53,120 So what does the medical establishment make of 837 00:52:53,120 --> 00:52:55,320 what they've seen? 838 00:52:55,320 --> 00:52:58,160 What they've done is they've done it in a lab environment. 839 00:52:58,160 --> 00:53:02,040 They've done it in vignettes, without real patients. 840 00:53:02,040 --> 00:53:07,080 I want to see that algorithm compete with a doctor over many, 841 00:53:08,120 --> 00:53:13,040 many cases, and dealing with Enid, who's 85, who's got multiple 842 00:53:13,040 --> 00:53:15,440 problems, hard of hearing. 843 00:53:15,440 --> 00:53:18,000 She's never going to use the phone. 844 00:53:18,000 --> 00:53:23,000 So I remain excited at the promise and I remain cynical at the moment. 845 00:53:25,200 --> 00:53:28,320 I think GPs will initially feel quite threatened by this, 846 00:53:28,320 --> 00:53:32,920 if the choice is a machine or a GP, but we all know it's not. 847 00:53:32,920 --> 00:53:36,160 The choice is, how do we improve what GPs can do already 848 00:53:36,160 --> 00:53:39,200 and which takes them a lot of time and often a lot of frustration 849 00:53:39,200 --> 00:53:41,880 and a great deal of stress, which worries me enormously 850 00:53:41,880 --> 00:53:45,080 about general practitioners in England. 851 00:53:45,080 --> 00:53:50,080 The strongest criticism comes from the Royal College of GPs, 852 00:53:50,080 --> 00:53:55,120 whose diagnosis exam Babylon say their AI effectively passed. 853 00:53:56,320 --> 00:53:57,920 Babylon don't have our exam. 854 00:53:57,920 --> 00:54:01,160 Babylon used sample old questions from our website, 855 00:54:01,160 --> 00:54:03,360 which they used, as far as we understand. 856 00:54:03,360 --> 00:54:06,440 Our exam is highly confidential. 857 00:54:06,440 --> 00:54:09,160 I had a phone call the night before the presentation to say 858 00:54:09,160 --> 00:54:11,680 that they were going to be making these claims and I did ask 859 00:54:11,680 --> 00:54:14,360 lots of questions. I asked about independent evaluation. 860 00:54:14,360 --> 00:54:15,800 I asked to see the data. 861 00:54:15,800 --> 00:54:17,600 I asked to see the methodology. 862 00:54:17,600 --> 00:54:20,200 Now, I didn't get to see those things, and for a very long time. 863 00:54:20,200 --> 00:54:23,320 However, it's interesting to see. 864 00:54:23,320 --> 00:54:26,480 It's innovative of them to try and evaluate it and I'm glad to see 865 00:54:26,480 --> 00:54:28,000 them trying to evaluate that. 866 00:54:28,000 --> 00:54:30,440 It's a good attempt in the right direction. 867 00:54:30,440 --> 00:54:32,240 But we have a very long way to go. 868 00:54:32,240 --> 00:54:35,280 What they have not done is published in any peer-reviewed journal. 869 00:54:35,280 --> 00:54:39,040 What they have not done is follow any robust scientific methodology. 870 00:54:39,040 --> 00:54:42,080 And what they have not done is allow others in to independently 871 00:54:42,080 --> 00:54:44,640 evaluate their research to date. 872 00:54:44,640 --> 00:54:46,680 All those things are possible and we look forward 873 00:54:46,680 --> 00:54:48,480 to them happening. 874 00:54:48,480 --> 00:54:49,800 We disagree with that. 875 00:54:49,800 --> 00:54:53,440 We think there is a very robust scientific approach and our 876 00:54:53,440 --> 00:54:56,240 scientists' reputation and name is on it. 877 00:54:56,240 --> 00:54:59,680 As any scientist knows, the peer review system was invented 878 00:54:59,680 --> 00:55:01,480 in the 18th century. 879 00:55:01,480 --> 00:55:03,080 It was... 880 00:55:03,080 --> 00:55:06,760 ..widely became common in the 19th and early parts of the 20th century, 881 00:55:06,760 --> 00:55:09,600 but it has a speed and at that time, science was progressing 882 00:55:09,600 --> 00:55:10,920 at a certain speed. 883 00:55:10,920 --> 00:55:15,480 So it takes normally 12-18 months from the time you do research 884 00:55:15,480 --> 00:55:17,760 for a peer-review paper to be published. 885 00:55:17,760 --> 00:55:21,320 That, in technology, is a very long time. 886 00:55:21,320 --> 00:55:25,280 Of course, we have to be in a hurry to bring accessible, 887 00:55:25,280 --> 00:55:28,080 affordable health care for everybody on earth. 888 00:55:28,080 --> 00:55:30,760 It's not an intellectual exercise. 889 00:55:30,760 --> 00:55:31,920 It's a mission. 890 00:55:36,000 --> 00:55:39,040 After the event, it's not just Ali Parsa answering 891 00:55:39,040 --> 00:55:40,840 questions about Babylon. 892 00:55:40,840 --> 00:55:45,560 On his first day at the despatch box, the incoming Health Secretary 893 00:55:45,560 --> 00:55:47,880 is in the firing line, too. 894 00:55:47,880 --> 00:55:51,280 Is the Secretary of State familiar with the GP At Hand online service, 895 00:55:51,280 --> 00:55:54,520 which has poached thousands of profitable patients from GPs 896 00:55:54,520 --> 00:55:58,920 all over London, to the alarm of the BMA and GPs generally? 897 00:55:58,920 --> 00:56:01,600 This is creating a two tier service for GPs. 898 00:56:01,600 --> 00:56:04,240 Will the Secretary of State investigate it? 899 00:56:04,240 --> 00:56:07,560 Erm, yes, I'm acutely aware of this, the question that he raises, 900 00:56:07,560 --> 00:56:12,000 not least because I'm a user of the Babylon service myself. 901 00:56:12,000 --> 00:56:13,680 They are my GP. 902 00:56:13,680 --> 00:56:17,240 Despite Matt Hancock's personal endorsement, 903 00:56:17,240 --> 00:56:19,880 NHS England are now investigating. 904 00:56:21,120 --> 00:56:24,520 There is no greater enthusiast for technology than me, 905 00:56:24,520 --> 00:56:27,240 Mr Speaker, as you well know. 906 00:56:27,240 --> 00:56:29,920 But the thing about new technology is that sometimes 907 00:56:29,920 --> 00:56:32,440 the rules need to be updated to take into account 908 00:56:32,440 --> 00:56:33,920 the changes in technology. 909 00:56:33,920 --> 00:56:35,720 So they've been properly evaluated... 910 00:56:35,720 --> 00:56:39,880 The regulations governing AI health products are being reviewed 911 00:56:39,880 --> 00:56:42,800 to ensure they are safe and effective. 912 00:56:42,800 --> 00:56:44,840 The answer, where there are challenges such as the ones 913 00:56:44,840 --> 00:56:47,920 she raises, is not to reject the technology. 914 00:56:47,920 --> 00:56:51,160 The opposite, it's to keep improving the technology so it gets better 915 00:56:51,160 --> 00:56:54,200 and better and make sure the rules keep up to pace. 916 00:56:54,200 --> 00:56:59,240 The NHS has appointed an independent evaluation of GP At Hand, 917 00:56:59,520 --> 00:57:02,760 and its impact on the health care system. 918 00:57:02,760 --> 00:57:06,000 For now, it has decided not to approve Babylon's 919 00:57:06,000 --> 00:57:08,440 expansion outside London. 920 00:57:08,440 --> 00:57:11,960 Yes, it's important to make sure that it works well 921 00:57:11,960 --> 00:57:15,000 and that the rules are right, but if we turn our backs 922 00:57:15,000 --> 00:57:16,960 on new technology, we're turning our backs 923 00:57:16,960 --> 00:57:18,080 on better care. 924 00:57:24,520 --> 00:57:29,440 Revolutions are famously hard to judge without the benefit 925 00:57:29,440 --> 00:57:31,840 of a lot of hindsight. 926 00:57:31,840 --> 00:57:34,000 So working out what to make of the one we're living 927 00:57:34,000 --> 00:57:36,880 through right now is not easy. 928 00:57:38,080 --> 00:57:41,200 For me, there is no question that we could all stand to benefit 929 00:57:41,200 --> 00:57:46,200 enormously from this revolution in AI health care. 930 00:57:46,640 --> 00:57:49,480 In fact, after a while, it might become immoral not to use 931 00:57:49,480 --> 00:57:53,520 AI, if it can outperform humans in so many different tasks. 932 00:57:53,520 --> 00:57:56,600 But at the same time, we might still feel, 933 00:57:56,600 --> 00:57:59,240 understandably, anxious about whether or not 934 00:57:59,240 --> 00:58:01,040 we can trust it. 935 00:58:01,040 --> 00:58:04,280 Because this new technology, these algorithms, they're only 936 00:58:04,280 --> 00:58:08,280 as good as the humans who are involved, the scientists, 937 00:58:08,280 --> 00:58:13,200 the inventors, the entrepreneurs, the politicians and the regulators. 938 00:58:13,200 --> 00:58:17,880 And, in this age of human and machine, it's only right 939 00:58:17,880 --> 00:58:19,920 that we should keep an eye on both. 940 00:58:21,920 --> 00:58:24,400 Businesses, technology businesses in the UK, 941 00:58:24,400 --> 00:58:29,240 are creating an environment which is safe and secure for investment... 942 00:58:29,240 --> 00:58:31,680 Secretary of State, Matt Hancock. 943 00:58:35,080 --> 00:58:37,600 Thanks, Ali, and it is great 944 00:58:37,600 --> 00:58:40,320 to be here to support GP At Hand.