1 00:00:00,330 --> 00:00:06,120 So thank you all very much for coming and thanks for the invitation to speak this evening. 2 00:00:06,600 --> 00:00:10,260 I've been asked to talk on this topic, Applied Digital Health, 3 00:00:10,290 --> 00:00:20,640 and what I've done is put a presentation together that talks about how it provides opportunities for better patient care. 4 00:00:20,670 --> 00:00:24,570 I think that will be self-evident from the presentation that I make. 5 00:00:25,200 --> 00:00:34,320 But perhaps more importantly for us at least, why it may assist better or more clinical research. 6 00:00:35,280 --> 00:00:42,330 I'm Richard Hope's I'm a professor of primary care and head of departments here at the University of Oxford. 7 00:00:42,960 --> 00:00:49,980 I'm a clinical scientist, and my main interests are actually cardiovascular disease and cardiovascular research. 8 00:00:50,460 --> 00:01:03,420 But I've always had an interest in better utilisation of health data and interventions that improve patient care as well as beyond drugs. 9 00:01:03,720 --> 00:01:08,140 Hence, I think, why I've been asked to give this talk this evening. 10 00:01:08,160 --> 00:01:12,420 Although methodologically, I think there are all sorts of challenges. 11 00:01:12,430 --> 00:01:21,030 I'm going to touch on a few of them this evening about how you can utilise data in a way that's more reliable. 12 00:01:21,780 --> 00:01:28,530 And I think if you want to ask difficult questions that are methodological in nature, 13 00:01:28,860 --> 00:01:38,790 then I shall immediately transfer that to the methodology in the room who actually do a lot of the work in relation to how the department 14 00:01:38,790 --> 00:01:52,950 uses big data or routine data and therefore are fully aware of the limitations and lengths you have to go to to try and make it worth doing. 15 00:01:53,820 --> 00:02:08,160 Okay. That's really by way of introduction and what I am also happy doing once I've worked to change my slides is taking questions during the session. 16 00:02:08,430 --> 00:02:14,250 So I think I've been asked to talk for an inordinate length of time, 40 minutes or so. 17 00:02:14,670 --> 00:02:23,460 So but I'm quite happy. If you want to interrupt me, please do so because I think it'll it'll break up the presentation as well. 18 00:02:24,270 --> 00:02:31,440 All you got to do is raise a hand or shout, okay, so this is really what I'm going to talk about, 19 00:02:32,520 --> 00:02:36,989 why one might invest in a presence in Applied Digital Health. 20 00:02:36,990 --> 00:02:41,729 And then I'm going to talk about the research impact in those areas of how it 21 00:02:41,730 --> 00:02:47,640 might assist in relation to disease burden or predicting disease outbreak, 22 00:02:47,910 --> 00:02:55,350 how it can improve drug development, how it can assist in better disease prognosis and risk estimation. 23 00:02:55,350 --> 00:03:06,110 And that's a very important topic because for most major disorders, the biggest prognostic indicator is the point at which is diagnosed. 24 00:03:06,110 --> 00:03:12,660 So if you diagnosed important disease earlier for many important diseases, that is the biggest. 25 00:03:13,440 --> 00:03:21,870 That's going to have the biggest impact on prognosis. I'm also a little bit about system change, behaviour change. 26 00:03:22,140 --> 00:03:26,070 We spend a lot of time thinking about how we can get patients to change behaviour. 27 00:03:26,070 --> 00:03:33,209 It's difficult but important, but actually increasingly it's how you can get practitioners and health systems to change. 28 00:03:33,210 --> 00:03:38,580 That's just as relevant and I think there are examples where digital health can provide that as well. 29 00:03:38,970 --> 00:03:44,400 And as I said, it's really all of the above in relation to health care impact. 30 00:03:44,400 --> 00:03:50,940 And as I say, that should be self-evident as I work through. So why is this become more possible? 31 00:03:50,940 --> 00:03:54,900 And it's simply the degree of digitisation. 32 00:03:55,070 --> 00:04:06,270 Digitisation of the clinical record has been particularly pronounced in primary care and in fact particularly advanced in the United Kingdom. 33 00:04:07,050 --> 00:04:17,190 So you can actually see that that pre 2000, in fact most UK general practice was fully computerised. 34 00:04:17,220 --> 00:04:23,670 And what's happened since then is that lots of health systems have caught up and it's 35 00:04:23,670 --> 00:04:30,000 this recognition that an awful lot of the clinical record is now held digitally. 36 00:04:30,450 --> 00:04:37,110 That has meant that people have started thinking that can't be utilised more than it has been historically. 37 00:04:38,040 --> 00:04:43,109 Just in terms of the UK and primary care, I'm really only talking about primary care, 38 00:04:43,110 --> 00:04:49,530 but you could imagine similar examples could occur in specialist settings. 39 00:04:50,100 --> 00:04:59,820 But in primary care, the principal software provider that provides the architecture for primary care records is themis they. 40 00:05:00,030 --> 00:05:04,740 About 55% of the market. And the next big is this TPP. 41 00:05:05,640 --> 00:05:13,650 And we've got relationships with both of these commercial operations and indeed are looking at ways that we 42 00:05:13,650 --> 00:05:20,070 can increase our relationship with them with regard to co-investing in things that they want and we want, 43 00:05:20,400 --> 00:05:28,320 then there may be opportunities for collaboration there and it's already impacts on clinical practice, 44 00:05:28,320 --> 00:05:37,709 at least because the what's what's happened at an increasing trend is that that information is 45 00:05:37,710 --> 00:05:44,280 provided back to practitioners that can help guide them in terms of guideline recommendations, 46 00:05:44,280 --> 00:05:52,079 usually as a prompt. And of course the difficulty of that is that there are now so many prompts within GP systems. 47 00:05:52,080 --> 00:05:53,430 It's a bit like guidelines. 48 00:05:54,150 --> 00:06:03,780 You'll have rooms full of guidelines, you'll now have patient consultations where you might get two or three prompts for every single consultation. 49 00:06:04,260 --> 00:06:14,160 What happens, GP's ignore them. And so thinking about how you can present information that's useful for practice in a way that you could take it up, 50 00:06:14,790 --> 00:06:19,050 I think is, is of itself an interesting research challenge. 51 00:06:20,220 --> 00:06:27,750 It's a sort of thing that Facebook and companies like that have transformed social the way we live our lives socially, 52 00:06:28,110 --> 00:06:31,170 the way we interact with shops, the way we bank. 53 00:06:31,710 --> 00:06:39,450 So in our social lives, it's doing this massive revolution about the way digital applications have changed, the way we behave. 54 00:06:40,080 --> 00:06:42,870 And really what I think is going to happen over the next ten, 55 00:06:43,080 --> 00:06:53,070 15 years is that degree of penetration into practice and the relationships patients have with health systems is going to go the same sort of way. 56 00:06:53,580 --> 00:06:58,319 And one of the good things about Facebook is that they've demonstrated what the 57 00:06:58,320 --> 00:07:02,760 risks and problems associated with these sorts of mass changes in behaviour are, 58 00:07:03,420 --> 00:07:11,700 and we should be thinking about how we can safely develop these systems that maximise, gain and minimise risk. 59 00:07:12,480 --> 00:07:22,560 And even if you think that it's been demonstrated that you can manipulate public voting through mass online campaigns, 60 00:07:22,950 --> 00:07:27,720 in some ways that's a good thing because if if you can do it for the way people vote, 61 00:07:28,410 --> 00:07:35,129 shouldn't we be thinking about things that have a desirable consequence in terms of the way people live their lives and living healthier 62 00:07:35,130 --> 00:07:45,930 lives and accepting more evidence based interventions rather than the sorts of interventions you'd see on the front page of the Daily Mail. 63 00:07:46,020 --> 00:07:47,700 So that's a challenge to us, really, 64 00:07:47,700 --> 00:07:58,620 as to how we can canvass this change in the way people interact with themselves and each other through digital applications. 65 00:07:58,620 --> 00:08:05,309 I think is is going to be quite exciting. So in terms of UK practice, 66 00:08:05,310 --> 00:08:14,610 there is an awful lot of information which is actually available within the clinical record in relation to activity prescribing referral patterns. 67 00:08:15,180 --> 00:08:20,940 This is a typical screen where you'll have demographic and clinical information 68 00:08:20,940 --> 00:08:29,460 provided and increasingly you can actually link this personal level clinical record, 69 00:08:30,120 --> 00:08:33,359 electronic clinical record into other datasets. 70 00:08:33,360 --> 00:08:40,860 And for the UK most notably there are going to be hospital activity data through has or onus particularly 71 00:08:40,860 --> 00:08:47,729 for mortality data but also deprivation scores and you can even triangulate for disease registries, 72 00:08:47,730 --> 00:08:55,740 etc. So it's not just the single clinical record that one would want to be looking into, it's other national datasets as well. 73 00:08:57,120 --> 00:09:05,279 So for the rest of the talk, I'm really just going to run through a few examples and you you can then discuss 74 00:09:05,280 --> 00:09:09,210 whether you think that they're useful things to be doing or not at a later point. 75 00:09:10,410 --> 00:09:13,170 So retrospective clinical epidemiology, 76 00:09:13,170 --> 00:09:22,139 that probably is the thing that's been done the longest that in Pharmacoepidemiology and probably digital clinical records have been 77 00:09:22,140 --> 00:09:32,100 used for about 20 years or so in those two areas and it's just a more efficient way of doing things like epidemiological studies, 78 00:09:32,100 --> 00:09:39,960 which often used to rely on just physical records being filled in on cohorts and then followed up for extended periods of time. 79 00:09:40,230 --> 00:09:46,410 The fact that you can actually start exploring this digitally obviously is hugely less time intensive. 80 00:09:46,440 --> 00:09:59,550 This is a US use. The US are coming to this light but are investing huge dollars, hundreds of millions of dollars in their ability to mine health. 81 00:09:59,920 --> 00:10:04,870 Data in the future. And obviously, you've got the mega data giants, 82 00:10:04,900 --> 00:10:16,270 Google and Apple and and Amazon are spending huge sums of money in what they see is going to be the next big digital transformation, 83 00:10:16,720 --> 00:10:21,459 which relates to health care. But this is just looking at Medicare and Medicaid. 84 00:10:21,460 --> 00:10:28,930 It's easier to use Medicare Medicaid data because it's a single registration, it's universal care. 85 00:10:28,940 --> 00:10:35,080 So it's effectively the equivalent of the National Health Service, but just for a small proportion of the US population. 86 00:10:35,770 --> 00:10:45,430 But what they are able to do looking at these data is look at an outcome that you wouldn't want, which is an early 30 day readmission, 87 00:10:46,680 --> 00:10:55,580 i.e. you've been in hospital for something important, in this case a stroke, and you've been discharged home if you've had a stroke. 88 00:10:55,600 --> 00:10:59,430 What you don't want when you're discharged home is to have to go back into hospital within 30 days. 89 00:10:59,440 --> 00:11:04,090 That's a bad clinical outcome. There are usually issues that you can do that prevent that. 90 00:11:04,570 --> 00:11:10,120 And so what they've been able to do by mapping this is to look at the worst quartile, 91 00:11:10,120 --> 00:11:15,849 which is the bits in dark there of the US that have the worst rates of readmission. 92 00:11:15,850 --> 00:11:20,440 And what you could imagine you might do with that then is think about preventative programmes that are focussed. 93 00:11:20,830 --> 00:11:26,050 What order did you put your initiatives in? Could be dictated by clinical need. 94 00:11:26,260 --> 00:11:29,889 Another important thing that has come from retrospective epidemiology are the 95 00:11:29,890 --> 00:11:38,500 observations that Multimorbidity have become the norm in most developed health systems, 96 00:11:38,500 --> 00:11:42,850 i.e. that patients don't have single disorders anymore but they have multiple disorders. 97 00:11:43,480 --> 00:11:46,210 I don't touch that anymore in case it starts flicking around. 98 00:11:46,420 --> 00:11:52,510 So this is just looking at people with diabetes and showing their own than the 20% of them. 99 00:11:52,510 --> 00:11:55,210 This is Scottish data used in primary care. 100 00:11:56,110 --> 00:12:04,120 It's all the primary care records in Scotland and it's showing that only under 20% of patients only have diabetes. 101 00:12:04,600 --> 00:12:12,879 And if you look at people under 65, then the mean number of additional conditions is about three and over 65. 102 00:12:12,880 --> 00:12:21,070 It's six and a half of the conditions. And so what you could imagine from that is why it's important to think about how interventions 103 00:12:21,070 --> 00:12:26,950 may interact negatively or positively in relation to different sorts of conditions. 104 00:12:28,270 --> 00:12:36,220 And you start to get a feel for why the evidence base needs to be less, you need disease focus. 105 00:12:37,480 --> 00:12:45,160 And I think it's only through having large digitised records that these sorts of observations can occur. 106 00:12:45,880 --> 00:12:54,480 What we're quite interested in, though, at the moment it's all been about talking about multimorbidity, counting it, viewing it as a big issue. 107 00:12:54,490 --> 00:13:00,220 There's very little original research around how that may impact on health care, 108 00:13:00,700 --> 00:13:04,750 but you could immediately begin to think that is it the way they cluster? 109 00:13:05,350 --> 00:13:10,870 Or is it the order that you get diseases in? Or is it a disease exposure issue? 110 00:13:12,010 --> 00:13:21,940 And those are sorts of questions that you might pose of a routine data set and come up with some preliminary questions about that then 111 00:13:21,940 --> 00:13:30,160 might guide you to think about an intervention that you could try to see if it altered trajectory compared to a comparator population. 112 00:13:31,690 --> 00:13:38,020 So that's that's the classic use of routine data sets and still quite important because that last one, 113 00:13:38,530 --> 00:13:43,569 I think you could immediately think about how it could guide you into hypothesis 114 00:13:43,570 --> 00:13:48,550 testing in an efficient way that may come up with some empirical subsequent testing, 115 00:13:49,450 --> 00:13:55,960 but in other important areas, disease, surveillance. And again, this might be more practical from a health system point of view, 116 00:13:56,470 --> 00:14:02,530 but you could imagine how it might impact on the research programme and for that 117 00:14:03,130 --> 00:14:10,450 showing data from the GP War College or GP's Research Surveillance Centre, 118 00:14:10,450 --> 00:14:14,770 Research and Surveillance Centre, which has recently moved to Oxford. 119 00:14:15,310 --> 00:14:25,330 And what that currently do is it provides the best early warning service to the National Health Service about rising rates of of influenza, 120 00:14:25,630 --> 00:14:27,700 and it uses sentinel practices for that. 121 00:14:27,700 --> 00:14:37,450 So it's general practices that get a little bit of additional money to to record things more than in routine practice that is collected twice weekly. 122 00:14:37,870 --> 00:14:42,579 This is just looking at time of year showing that most flu, unsurprisingly, 123 00:14:42,580 --> 00:14:47,860 is in the first part of the year, but occasionally you'll get spikes out of sequence. 124 00:14:48,610 --> 00:14:53,349 And the other thing that they can do because they actually collect serology in these practices, 125 00:14:53,350 --> 00:14:59,440 which again is not done routinely, is look at what influenza subtype it is, which helps the. 126 00:14:59,530 --> 00:15:09,670 Guide how the NHS orders the antibody that goes into the flu, immunisations, it ultimately does. 127 00:15:10,210 --> 00:15:15,580 But what we're quite interested in doing this thinking that could you start mapping disease 128 00:15:15,580 --> 00:15:21,490 burden better at a local and at a regional level using these sorts of surveillance systems 129 00:15:21,970 --> 00:15:30,820 that might give the NHS more granularity about what may become big burden problems or 130 00:15:30,850 --> 00:15:36,669 what may predict a surge in requirements for acute admission in advance of it occurring. 131 00:15:36,670 --> 00:15:45,370 Because at the moment the NHS tends to be very responsive. It can't really predict where demand is going to occur and at what time and where you 132 00:15:45,370 --> 00:15:51,340 might be able to develop intelligent systems around those sorts of service issues. 133 00:15:52,780 --> 00:15:59,800 So another area which is historically has been used for for pharmacovigilance, so a drugs introduced, 134 00:16:00,190 --> 00:16:04,600 you can monitor its safety subsequently once it starts being used in clinical practice. 135 00:16:05,680 --> 00:16:13,210 But I think that's been extended now to think that because the data are more complete and more reliable could 136 00:16:13,300 --> 00:16:18,520 actually start to substitute for what has historically been the way that drug development has occurred, 137 00:16:19,000 --> 00:16:28,750 which is through the randomised controlled trial. And what I've done here is I've plotted the Re-ly trial which is looking at a direct 138 00:16:28,750 --> 00:16:33,069 oral anticoagulants or an alternative to warfarin that was on the bottom there. 139 00:16:33,070 --> 00:16:42,910 The pivotal trial that resulted in Dabigatran, the first of these new warfarin drugs, if you like, these anticoagulants. 140 00:16:44,110 --> 00:16:52,509 That was a pivotal trial against warfarin and showing the benefits in terms of reduced rates of 141 00:16:52,510 --> 00:17:01,360 mortality and less ischaemia stroke and particularly bleeding into the brain with some of the downside, 142 00:17:01,360 --> 00:17:05,409 which is that where there was more GI bleeding than warfarin. 143 00:17:05,410 --> 00:17:14,700 So some of the upside, some of the downside and the top one looks basically at its use in clinical practice, much larger data set. 144 00:17:14,710 --> 00:17:18,370 So that's looking at exposure in hundreds of thousands of people. 145 00:17:18,940 --> 00:17:22,960 This is a huge city, but it was around 20,000 people. 146 00:17:23,350 --> 00:17:33,969 These are tens of thousands of people up the top there in routine data use and it basically is identical results. 147 00:17:33,970 --> 00:17:41,800 And that's very reassuring because it's triangulating what may be a more select population in a clinical trial, in a more routine setting. 148 00:17:42,280 --> 00:17:49,690 And the drug regulators are now wanting real world data to complement complement the CT data. 149 00:17:50,530 --> 00:17:57,579 This is looking at comparison between the other direct oral anticoagulants. 150 00:17:57,580 --> 00:17:59,710 In fact, there's been a fourth one now, 151 00:17:59,920 --> 00:18:07,060 but it gives you an idea of the fact that basically all of them are at least as safe in terms of bleeding as warfarin, 152 00:18:07,540 --> 00:18:11,320 and perhaps some of them are a bit safer. But of course, 153 00:18:11,710 --> 00:18:16,930 there's going to be all sorts of bias in that because the reason why people prescribe a 154 00:18:16,930 --> 00:18:21,820 particular drug to a particular patient may result in some spectrum bias in the observation. 155 00:18:22,660 --> 00:18:28,989 But if you think that the clinical perception out there is that this is the safest when it comes to bleeding, 156 00:18:28,990 --> 00:18:31,570 that's what physicians think, rightly or wrongly. 157 00:18:31,960 --> 00:18:40,630 Then you might hypothesise that actually these patients are more likely to be higher bleed risk because it's perceived as a safer one. 158 00:18:41,140 --> 00:18:51,610 So the fact that it actually comes out despite that that probable spectrum bias as less bleeding, I think is is reassuring. 159 00:18:52,510 --> 00:18:53,709 So it's still observational. 160 00:18:53,710 --> 00:19:03,640 It's still going to caveat it, but it would provide quite a lot of reassurance to practitioners and that those last two were US data. 161 00:19:03,640 --> 00:19:09,190 This is UK data is actually based on key research showing essentially the same thing. 162 00:19:09,190 --> 00:19:18,009 And I think in medicine when you observe what you would anticipate, then that's very reassuring. 163 00:19:18,010 --> 00:19:23,650 So on the bottom here, Rivaroxaban, as you saw on that first side, 164 00:19:24,640 --> 00:19:31,690 is likely to cause more gastric bleeding in the RC to put it into general population. 165 00:19:31,690 --> 00:19:42,880 And yes, it does look as if it causes a bit more GI bleeding than warfarin and apixaban less so. 166 00:19:42,970 --> 00:19:49,000 So again triangulate data from randomised controlled trials. 167 00:19:50,560 --> 00:19:59,350 But that I think is is how it's been used historically and I think that the potential for routine data is that you could actually start doing. 168 00:19:59,420 --> 00:20:08,900 Things that currently aren't done. And two areas of drug development that it may be useful for is looking at subgroup subpopulations, 169 00:20:09,650 --> 00:20:18,260 where it's much more difficult potentially to do a randomised controlled trial or even having a much lower cost way of looking at new indications. 170 00:20:18,270 --> 00:20:21,500 And I'm going to show you a few examples now. 171 00:20:22,430 --> 00:20:25,520 So one of them is Metformin, 172 00:20:25,520 --> 00:20:37,729 which is a drug that's used in diabetes and was contraindicated for use in heart failure because there was a low risk of a serious complication, 173 00:20:37,730 --> 00:20:42,350 lactic acidosis. But in fact, when the FDA, 174 00:20:42,950 --> 00:20:52,910 the regulators in the US looked at all the observational data in relation to people who for whatever reason had been put on metformin, 175 00:20:53,420 --> 00:21:01,850 even though they weren't supposed to be. In fact, the data were very reassuring with all of the studies showing, if anything, safety. 176 00:21:02,450 --> 00:21:09,230 And as a consequence, in the end of what was observational data, the FDA removed that risk label. 177 00:21:09,260 --> 00:21:17,630 You'd never do a randomised controlled trial of that sort of indication where you were concerned about risk in the population. 178 00:21:19,370 --> 00:21:30,169 Here's another one where in fact is the traditional use where you do a clinical trial of various drugs used to treat hypertension, 179 00:21:30,170 --> 00:21:40,730 used in combination at a low level and resulted in a noninferiority of a more intensive regime against the less intensive regime. 180 00:21:41,180 --> 00:21:48,860 And then subsequent to that, a group actually mimicked the trial in routine data sets, 181 00:21:48,860 --> 00:21:52,639 because all of these drugs are routinely available and therefore a lot of 182 00:21:52,640 --> 00:21:56,420 prescriptions occur in clinical practice and showed essentially the same result. 183 00:21:57,140 --> 00:22:05,740 And this trial would have cost about $300 million and that study cost about $4 million. 184 00:22:05,750 --> 00:22:15,440 So you can start to see why people think that efficient study design might be more possible with digital record studies. 185 00:22:16,400 --> 00:22:25,010 And here's another one, which is the other way round, which is that this study had been announced, which is looking at a diabetes drug. 186 00:22:25,190 --> 00:22:30,620 And the observation that these drugs which are used for diabetes might actually 187 00:22:30,620 --> 00:22:34,760 be useful in heart failure as well in terms of reducing hospitalisation. 188 00:22:35,570 --> 00:22:46,250 And that observation came out of observational studies showing a hazard ratio of six, 189 00:22:46,580 --> 00:22:51,800 so a 40% improvement in terms of hospitalisation for heart failure. 190 00:22:52,640 --> 00:22:59,810 And the observational data was sufficient to actually get an RC t done to get the formal indication, 191 00:23:00,140 --> 00:23:07,040 almost identical results, but obviously hugely more expensive to do the randomised controlled trial. 192 00:23:08,030 --> 00:23:16,550 No one has to say though there are risks with the way you do this data because methodologically there are all sorts of challenges you're going to do. 193 00:23:17,360 --> 00:23:25,879 And this is an interesting one because basically what this group did is looked at a restrictive analysis of routine 194 00:23:25,880 --> 00:23:39,170 data to compare of how Pravastatin performed in randomised controlled trials in terms of the amount of mortality, 195 00:23:39,350 --> 00:23:42,860 cardiovascular mortality that was reduced by being on Pravastatin. 196 00:23:43,430 --> 00:23:52,400 And when they just did the simplest type of comparison, which is looking at everybody who's on Pravastatin versus people who weren't, 197 00:23:53,360 --> 00:24:02,870 then you've got this huge difference between observed or actual trial results and what was observed in the routine data set. 198 00:24:03,320 --> 00:24:11,600 But as they did more restrictive analyses doing things like only looking at people who were new onset Pravastatin rather than prevalent, 199 00:24:12,320 --> 00:24:21,290 looking at active comparators, looking at comparators where they started to standardise the patients to be more similar, 200 00:24:22,070 --> 00:24:29,360 then each incremental step they did resulted in a result that was much closer to that observed in the randomised trial. 201 00:24:29,510 --> 00:24:35,840 So it this demonstrates why you need to think methodologically about how you approach these questions. 202 00:24:36,620 --> 00:24:38,210 And here's another example. 203 00:24:39,230 --> 00:24:48,860 There was a group that published the CV, the real data set, which again looked at a type of antidiabetic drug, SGA SGLT2 inhibitors, 204 00:24:49,550 --> 00:24:58,700 which appeared to have this huge effect on vascular mortality with a 50% reduction based on the big data studies. 205 00:24:59,490 --> 00:25:06,690 But actually the this was actually demonstrated in a trial. 206 00:25:06,990 --> 00:25:14,080 It's lost the plot. Sorry, but but you can see in the clinical trial it was only a 13% risk reduction. 207 00:25:14,110 --> 00:25:17,190 Still a worthwhile risk reduction, but nothing like the 50%. 208 00:25:18,030 --> 00:25:22,200 And why did they make the mistake in the CVD rule group? 209 00:25:22,200 --> 00:25:30,930 And they did another basic error in terms of how you would look at pharmaco safety in a trial. 210 00:25:31,440 --> 00:25:43,050 The immortal term bias effect. Because what they were doing is looking at people who were randomised to either initial t one or other 211 00:25:43,620 --> 00:25:53,760 oral glucose lowering drug and then looking for outcomes which was like a death or a vascular event. 212 00:25:54,780 --> 00:26:02,970 But so if you were on an alternative or sl t two, then that's a reasonable comparison. 213 00:26:03,780 --> 00:26:11,040 But there would have been other people who were in this group who were on an old drug 214 00:26:11,550 --> 00:26:17,370 but were randomised or given in clinical practice in this trial to one inhibitor. 215 00:26:17,400 --> 00:26:28,110 And how are you going to count those events where for that bottom group, if that individual died, it counts in the analysis. 216 00:26:28,770 --> 00:26:36,030 But in this group, if they don't die because they live to be able to go into another drug, 217 00:26:36,360 --> 00:26:43,020 then that survival will be attributed to the oestriol t two drug in the analysis is what the mistake they made, 218 00:26:43,500 --> 00:26:49,110 rather than attributing some of that survival to the older drug as well. 219 00:26:49,650 --> 00:26:55,770 So you've got to think about these things if you're going to use the databases sensibly. 220 00:26:57,150 --> 00:27:05,130 So another area, which is going to be a big growth area in medicine because it's about trying to predict risk coming to validated risk scores. 221 00:27:06,270 --> 00:27:12,060 And I'm just used for an example there, all of the ones generated from the Q research database. 222 00:27:12,540 --> 00:27:17,489 There are lots of other risk scores that are used internationally from other databases, 223 00:27:17,490 --> 00:27:26,520 but it gives you a flavour of how you could use a large dataset and actually look at the factors that 224 00:27:26,520 --> 00:27:32,610 are independently associated with outcomes that you can then end up with a weighted composite score. 225 00:27:33,030 --> 00:27:45,299 And indeed, the key Risk three calculator is currently used by the National Health Service as its mandated risk score for predicting the ten year 226 00:27:45,300 --> 00:27:52,830 risk of somebody developing vascular event and then using that to determine whether you should treat them with a statin or not. 227 00:27:53,880 --> 00:27:56,430 But the other thing that you can do with these sorts of data, 228 00:27:56,430 --> 00:28:04,440 so using it in that in a clinical threshold setting is think about the way you present risk to people. 229 00:28:05,370 --> 00:28:07,290 So in this instance, 230 00:28:08,790 --> 00:28:19,770 a ten year Q risk score of 21% means that you've got a 21% likelihood of having an event within a ten year cycle across the population. 231 00:28:20,880 --> 00:28:30,450 But that means that that individual has got a relative risk compared to other people of a similar age and similar agenda of nearly fourfold. 232 00:28:31,110 --> 00:28:36,389 Or you can calculate that into a lifetime risk of 74, 233 00:28:36,390 --> 00:28:46,470 i.e. this 55 year old is on a population basis on average is going to be more like a 74 year old in terms of their outcomes. 234 00:28:47,040 --> 00:28:53,609 And so that immediately then throws up how you present risk may be quite important to helping people 235 00:28:53,610 --> 00:29:00,270 understand that personal risk and then making decisions about what they do about that perceived risk. 236 00:29:01,260 --> 00:29:09,660 You've also got all of these issues is that this is all based upon mean population effects, and you're attributing that to an individual, 237 00:29:10,530 --> 00:29:16,290 which may obviously underestimate the risk in some individuals or overestimate 238 00:29:16,290 --> 00:29:20,040 the risk in some individuals because you're just looking at a mean effect. 239 00:29:20,730 --> 00:29:23,430 So what we're quite interested in looking at, 240 00:29:23,430 --> 00:29:31,260 can you supplement these scores perhaps with biomarkers that may produce a more predictive individualised risk? 241 00:29:33,360 --> 00:29:40,049 So another area that we're pretty plain on, which is part of this behaviour change, 242 00:29:40,050 --> 00:29:47,640 is can you provide contextual information back to practitioners in this instance that helps them change behaviour? 243 00:29:48,570 --> 00:29:58,830 So this is what the GP currently provides back in relation to flu and it's producing all sorts of pieces of information about how this. 244 00:29:59,330 --> 00:30:07,040 Compared with the mean practice and talks about how much money they're losing. 245 00:30:08,030 --> 00:30:18,770 But it's obviously is a pretty complex slide and it does change behaviour, but you'd have to look at that in a practice meeting probably and think, 246 00:30:18,800 --> 00:30:23,510 digest it and think about the implications for your practice and then you might do something about it, 247 00:30:24,770 --> 00:30:29,480 but it's certainly not something that you're going to look at during a consultation to change your behaviour. 248 00:30:30,650 --> 00:30:36,770 Another thing that we've done in terms of routine data is is look at workload. 249 00:30:37,600 --> 00:30:46,399 It may be hard to believe, but since 1948 when the NHS was formed, there was only one analysis, 250 00:30:46,400 --> 00:30:52,370 objective analysis of GP workload and that just looked at the numbers of consultations. 251 00:30:53,510 --> 00:30:55,639 So there was no other objective evidence. 252 00:30:55,640 --> 00:31:04,940 And yet we had this huge concern in British general practice that they were under huge pressure and they were buckling under the workload increases. 253 00:31:05,510 --> 00:31:09,559 And yet the NHS had, you know, 254 00:31:09,560 --> 00:31:17,600 they didn't know whether it was just perceptions of working harder or actual working harder and we managed to using routine data, 255 00:31:18,200 --> 00:31:23,540 look at a seven year period, 2014 sorry, 256 00:31:23,540 --> 00:31:35,240 27 to 2014 and showed that objectively during that period there was this really remarkably consistent increase, 257 00:31:35,660 --> 00:31:43,969 annual increase in consultations, the 16% which were associated also with an increased length of consultations. 258 00:31:43,970 --> 00:31:45,740 So they weren't just seeing patients more, 259 00:31:45,740 --> 00:31:52,280 they were actually spending more on each average consultation and they were doing things to try and reduce the burden, 260 00:31:52,280 --> 00:31:54,950 like increasing them by telephone consultation. 261 00:31:54,950 --> 00:32:03,769 So I think the good thing about these data is it really did demonstrate that they their perceptions were borne out by reality and did 262 00:32:03,770 --> 00:32:13,970 actually change NHS England a view to this in what they then committed to and to try to increase the number of practitioners in the UK. 263 00:32:14,930 --> 00:32:21,440 We've had a number of follow on studies for that and we're still trying to look at issues like complexity. 264 00:32:21,440 --> 00:32:28,610 Can you actually weight workload by complexity of disorders that they're looking after within the population? 265 00:32:30,170 --> 00:32:35,750 Now this is another big data project that's emerged from from the department here in Oxford, 266 00:32:36,440 --> 00:32:41,870 which relates to open prescribing actually is not uncommon. 267 00:32:42,860 --> 00:32:45,079 But what this is doing is using a different data set. 268 00:32:45,080 --> 00:32:51,860 This is looking at the prescription pricing authority which previously was sent manually to CCGs, 269 00:32:51,860 --> 00:33:01,310 and then a pharmacist would pore over it all these tables that were done and look at the differences between practices 270 00:33:01,880 --> 00:33:08,150 and then would target practices to go into and say what they should be doing differently to try and save the NHS money. 271 00:33:08,750 --> 00:33:13,760 But what using big data techniques this group has done, led by Ben Goldacre, 272 00:33:13,760 --> 00:33:22,700 is to provide all of this online real time so that you can immediately start to focus on any particular comparison 273 00:33:22,820 --> 00:33:33,559 that this is looking at basically the generic prescribing of a progesterone drug rather than proprietary prescribing, 274 00:33:33,560 --> 00:33:40,730 and it divides up all the CCG by those doing a lot of the non generic prescribing, 275 00:33:41,060 --> 00:33:48,469 which is you're not supposed to do that sort of thing, but they can also do some time trend stuff which is quite interesting because the NHS 276 00:33:48,470 --> 00:33:53,840 every now and then tries to stop things so they'll have a campaign that will try and 277 00:33:53,840 --> 00:33:59,930 stop GP's using the proprietary drug and you can actually look at whether there are 278 00:33:59,930 --> 00:34:04,159 interventions that they did that are then associated with the change in activity. 279 00:34:04,160 --> 00:34:12,710 So the blue is basically the national trend and then the red is looking at individual CCGs. 280 00:34:12,860 --> 00:34:18,860 So here's one that dropped, had more desirable practice much more quickly. 281 00:34:19,460 --> 00:34:25,490 And you'd postulate did they do something then that resulted in this sudden and very rapid change? 282 00:34:26,060 --> 00:34:37,910 And here's one where they didn't do anything and then obviously they did something that resulted in here, but that is sort of three years difference. 283 00:34:38,780 --> 00:34:44,659 And so you could imagine now that you could think about a CG intervention and 284 00:34:44,660 --> 00:34:49,940 test it in a cheap way of following whether there's been desirable trend trend. 285 00:34:50,720 --> 00:34:59,020 And the group are actually now looking at the use of prompts where practices are. 286 00:34:59,090 --> 00:35:06,530 Actually asking for this information when the prescribing is starting to change to what they'd expect it to. 287 00:35:08,900 --> 00:35:16,370 The final area where I think is going to be a big change in terms of routine data relates to pragmatic trial platforms. 288 00:35:16,370 --> 00:35:23,210 And I think that that should have been self-evident from what I was talking about in real world studies that are looking at retrospective use. 289 00:35:23,840 --> 00:35:29,870 But potentially you could actually randomise people and do most of their phenotypic follow up. 290 00:35:29,870 --> 00:35:35,030 That's the expensive bit of trials is following a patient to capture all of the clinical information. 291 00:35:35,840 --> 00:35:40,040 Then you could imagine that you might be able to start doing these things prospectively. 292 00:35:40,940 --> 00:35:48,560 And one example, for example, might be that you could look at these drugs again, these are a diabetic drug, 293 00:35:49,370 --> 00:35:54,500 but do look as if they may be positive when it comes to vascular risk reduction. 294 00:35:55,310 --> 00:36:04,250 This trial will never happen. Nobody is going to have to do a pivotal primary prevention studies because the cost of it is just too great. 295 00:36:04,580 --> 00:36:14,810 But it's a really important question, and potentially you could get a reliable answer to it through using routine data sets. 296 00:36:17,510 --> 00:36:21,110 Okay. Now, although that's all I want to talk about in terms of data, 297 00:36:21,110 --> 00:36:35,750 I think it is worth me just closing by saying that there are things that we could be thinking about also in terms of kit of things that link to data. 298 00:36:36,950 --> 00:36:41,570 And one example is things that measure devices, that measure things. 299 00:36:42,170 --> 00:36:49,020 And one good example that we've been working on could be in relation to blood pressure management. 300 00:36:49,850 --> 00:36:55,760 And what we postulated was that there was quite good evidence to show that self-monitoring. 301 00:37:00,790 --> 00:37:07,720 Sorry. It's really slow and I'll change it. But self-monitoring actually does have an appreciable effect on blood pressure, 302 00:37:07,870 --> 00:37:20,260 with about a four over 1.5 reduction of blood pressure in patients who self-monitor compared to people who are randomised to non-intervention. 303 00:37:20,500 --> 00:37:22,930 And that's a worthwhile clinical difference. 304 00:37:23,590 --> 00:37:31,390 And we postulated that rather than just self-monitoring, if you provided patients with an algorithm to allow them to actually change their treatment, 305 00:37:31,990 --> 00:37:37,030 so they monitored and self-managed that that might be a desirable thing. 306 00:37:37,600 --> 00:37:42,129 And in fact you had a big treatment effect very similar to the self-monitoring, 307 00:37:42,130 --> 00:37:51,520 but it's on top of self-monitoring and it actually was a bigger treatment effect at 12 months, persisted even after the original trial intervention. 308 00:37:52,060 --> 00:37:58,900 And if you think that doesn't sound very much five millimetres of mercury, then look at the epidemiology. 309 00:37:58,900 --> 00:38:02,980 You can predict that it will have actually quite a significant effect on outcomes. 310 00:38:03,430 --> 00:38:10,060 The trouble is you'd have to do follow up of these cohorts for 15 to 20 years before you could demonstrate that. 311 00:38:10,720 --> 00:38:13,750 But of course, within the dates that you could do that length of follow up, 312 00:38:13,750 --> 00:38:18,970 but you're not going to wait 20 years before you think an intervention is worth doing. 313 00:38:19,510 --> 00:38:24,970 If there are surrogate markers and blood pressure is a very good surrogate marker of clinical outcomes, 314 00:38:26,200 --> 00:38:30,519 we've also done some work in more challenging settings. 315 00:38:30,520 --> 00:38:36,100 So this is in South Africa in townships looking at whether prompts, 316 00:38:36,700 --> 00:38:45,819 using simple messaging on smartphones actually results in desirable change in amongst the 317 00:38:45,820 --> 00:38:53,320 group that often the even more lacking in concordance with medication than in the UK. 318 00:38:54,070 --> 00:39:01,270 And the one thing that is quite good about the world today is that even in really poor settings, everybody's got a smartphone. 319 00:39:02,020 --> 00:39:08,650 I don't know how they pay the I think they're all given free the roaming they get roaming charges. 320 00:39:09,100 --> 00:39:13,870 But it does mean you can think about digital interventions even in poor settings. 321 00:39:14,410 --> 00:39:15,760 And did that make a difference? 322 00:39:15,760 --> 00:39:26,079 Yes, it may have been only two millimetres of mercury difference in the intervention of prompted intervention, but that, as I already demonstrated, 323 00:39:26,080 --> 00:39:35,440 is a worthwhile intervention and that was associated presumably with the fact that 25% more adherence to treatment within that group. 324 00:39:36,280 --> 00:39:43,300 And I think when you think about the future in relation to wearables, different devices linking in to the clinical record, 325 00:39:43,990 --> 00:39:49,660 then clearly there are going to be a lot of opportunities for more active research downstream. 326 00:39:51,130 --> 00:39:58,090 So I think hopefully you'll agree that there are all sorts of exciting new things we can think about 327 00:39:58,090 --> 00:40:03,820 in relation to the use of big data and the linkage of data and information in the clinical record. 328 00:40:04,300 --> 00:40:09,160 And the already there have been some worthwhile research examples that would encourage you in that. 329 00:40:10,000 --> 00:40:18,819 And we currently practice from that building which is the renovated outpatient building on the ARO key site. 330 00:40:18,820 --> 00:40:27,820 And west of efficiently convinced about this is that we are hopefully going to be investing or getting funding 331 00:40:27,820 --> 00:40:32,740 to invest in a new building which will probably be about that sort of size for the rest of the department. 332 00:40:33,430 --> 00:40:39,280 That's the current building, which will be largely around big investment in applied digital health. 333 00:40:40,000 --> 00:40:45,250 So watch this space and look at for future employment opportunities. 334 00:40:46,330 --> 00:40:47,050 Thank you very much.