1 00:00:00,420 --> 00:00:10,260 So thank you for the invitation. And I'm very happy to be here to to present the results of this of this project. 2 00:00:10,260 --> 00:00:18,210 But it's more an opportunity to discuss how, as a clinical academic, 3 00:00:18,540 --> 00:00:30,900 I'm trying to bridge what is the best available evidence to something that can really form the treatment decision making process and our patients. 4 00:00:31,800 --> 00:00:37,290 First of all, yes, I don't have any conflict of interest. 5 00:00:37,290 --> 00:00:39,930 I've never received money from the pharmaceutical industry. 6 00:00:40,260 --> 00:00:50,280 And I think it's important to say this because in psychiatry and psychopharmacology especially, there's a lot of ideology. 7 00:00:50,280 --> 00:00:57,840 And I think it's important to state that I don't have any conflict of interest on this matter. 8 00:00:58,320 --> 00:01:01,770 So this is a master's about evidence synthesis. 9 00:01:01,770 --> 00:01:08,250 And I think it's good to start with what we consider the the best in terms of the hierarchy of evidence. 10 00:01:08,610 --> 00:01:12,989 So this is what we usually teach our students. 11 00:01:12,990 --> 00:01:25,500 And the idea is that depending on the specific questions in this case, for instance, about therapy, which means whether a treatment is effective, 12 00:01:25,920 --> 00:01:31,950 we all know that level one is randomised controlled trials, evidence from randomised studies, 13 00:01:32,820 --> 00:01:38,490 level two is cohorts, observational studies, cohort studies first and then case control. 14 00:01:38,730 --> 00:01:46,620 And the idea is also that within level one, if we have a systematic review, a collection of many trials, 15 00:01:47,400 --> 00:01:55,020 this is likely better than having an individual study because of the external validity, the generalisability of findings. 16 00:01:55,350 --> 00:01:58,409 And also, I mean the more the merrier. 17 00:01:58,410 --> 00:02:04,980 So the statistical power, if you do a mat analysis is, is much, much more informative. 18 00:02:08,760 --> 00:02:12,150 And this is what standard meta analyses do. 19 00:02:12,630 --> 00:02:20,220 So you have a pool of studies comparing, let's say, treatment, eh, versus placebo. 20 00:02:20,220 --> 00:02:25,950 And in the end you have a pooled weighted estimate comparing a versus placebo. 21 00:02:27,240 --> 00:02:38,160 If we say that on a specific topic or specific question, we also have a pool of three different studies comparing A versus B. 22 00:02:38,370 --> 00:02:45,509 What standard meta analysis do is to basically be able to summarise and pool all the 23 00:02:45,510 --> 00:02:51,990 evidence a versus placebo in one single estimate A versus B in one single estimate. 24 00:02:52,800 --> 00:03:03,750 But the problem is that E as clinicians, basically, these cannot be really informative because we know what is one head to head comparisons. 25 00:03:03,750 --> 00:03:09,180 We know what the estimate between another head to head comparisons, but we don't have the full picture. 26 00:03:09,780 --> 00:03:20,640 So in a nutshell, what we do with the network match analysis is in a set of trials comparing over placebo and a set of trials comparing A versus B, 27 00:03:20,910 --> 00:03:28,590 we use a treatment in common like A to calculate the indirect evidence B versus placebo. 28 00:03:29,550 --> 00:03:38,520 This is the idea behind the approach so called network match analysis or multiple treatment meta analysis. 29 00:03:39,000 --> 00:03:47,520 And the reason why it is informative for for for clinicians is because if we have a network of 30 00:03:47,520 --> 00:03:56,400 experimental interventions which can be in this case antidepressants but can be any kind of treatment, 31 00:03:57,390 --> 00:04:05,549 we can summarise all available evidence into this network and the graphical representation of the network is also, I think, important. 32 00:04:05,550 --> 00:04:14,430 So in this specific situation, the size of the node is proportional to the number of patients randomised to a specific intervention for oxygen. 33 00:04:14,430 --> 00:04:23,639 Here you see many patients randomised to fluoxetine, not so many to be protean and the lines represent the number of direct comparisons. 34 00:04:23,640 --> 00:04:32,219 So the studies comparing one treatment versus the other and the width of the line is proportional to the number of trials comparing to treatment. 35 00:04:32,220 --> 00:04:38,220 So many studies comparing fluoxetine with certainly not so many comparing escitalopram 36 00:04:38,220 --> 00:04:43,680 with bupropion and actually no studies comparing reebok's acting with vitamins effect. 37 00:04:44,310 --> 00:04:51,210 So if the question is which is the difference between Reebok sitting and witness have been to licensed antidepressants on the market? 38 00:04:51,540 --> 00:04:54,660 I carry out a systematic review and the answer is no data. 39 00:04:55,980 --> 00:04:59,590 So with this network of treatment, I can kind. 40 00:04:59,690 --> 00:05:07,160 You lay out the indirect evidence for a boxer to move to a pin going by a common comparator, like, for instance, fluoxetine. 41 00:05:07,670 --> 00:05:11,780 So one of the advantages of network meta analysis is to fill the gaps. 42 00:05:12,500 --> 00:05:20,930 But I think that the most important advantage of network analysis is their comparison between direct and indirect evidence. 43 00:05:21,620 --> 00:05:25,519 So for comparison, like certainly versus fluoxetine with many, 44 00:05:25,520 --> 00:05:31,520 many trials comparing sensors in fluoxetine, we have a direct estimate, the solid red line. 45 00:05:32,060 --> 00:05:41,850 But using the network I can also calculate, for instance, via venlafaxine the indirect evidence or green dotted line. 46 00:05:42,830 --> 00:05:50,030 And what I can do is to compare the two and check the so-called consistency of the network. 47 00:05:50,420 --> 00:05:58,520 Consistency of the network is where the direct and indirect estimates are showing the same results or similar results. 48 00:05:59,300 --> 00:06:06,020 The other thing is that I can check the consistency of the estimates, certainly in fluoxetine, 49 00:06:06,260 --> 00:06:12,700 not only via venlafaxine, but I can go by any closed loop within the network. 50 00:06:12,710 --> 00:06:14,810 So via citalopram, by fluoxetine. 51 00:06:15,080 --> 00:06:24,710 These are three not loops, but I can also use fluoxetine escitalopram paroxetine and certainly in phone notes or any kind of number of nodes. 52 00:06:24,740 --> 00:06:35,330 Of course, the higher the number of nodes, the less informative is the comparison because I increase the uncertainty around the the estimate. 53 00:06:35,780 --> 00:06:43,910 So in this kind of network, I think, which is very well connected. 54 00:06:44,900 --> 00:06:49,040 Another important thing to consider is the so-called geometry of the network, 55 00:06:49,400 --> 00:06:58,760 because that comparison between direct and indirect, the consistency of the network can be checked only if I have closed loops. 56 00:06:59,390 --> 00:07:03,770 If I don't have closed loops within the network, I cannot check consistency. 57 00:07:04,400 --> 00:07:06,530 So the advantages of net from its analysis, 58 00:07:06,530 --> 00:07:13,009 I can compare interventions which haven't been directly compared in any trial, fill the gaps and Murtaza remarks. 59 00:07:13,010 --> 00:07:21,200 And then I can use it. I can have a comprehensive use of all available data, direct and indirect, as I showed. 60 00:07:21,200 --> 00:07:27,829 I can improve precision and I'll show you in a bit with the antidepressants met analysis how we can improve precision. 61 00:07:27,830 --> 00:07:33,350 But basically we add the confidence intervals so they they increase the precision. 62 00:07:33,530 --> 00:07:38,180 And I can also rank the treatments using different methods. 63 00:07:38,630 --> 00:07:45,170 One of the methods we use to rank the treatments from the estimates of measurement analysis is this one. 64 00:07:45,170 --> 00:07:51,320 So we estimate for each intervention, which is the probability to be in each position. 65 00:07:51,770 --> 00:07:57,740 This is a fictional example with only four treatments just to exemplifying the process, 66 00:07:58,070 --> 00:08:05,209 but basically using the estimates that we calculate from the natural naturellement analysis which can calculate for a we can sorry, 67 00:08:05,210 --> 00:08:15,430 we can estimate for a which is the privilege to be the first which is 25%, it is 50% for B, it is 25% for C and zero for D. 68 00:08:15,950 --> 00:08:24,320 So can I ask you the winner is exactly 50% probability the second Y. 69 00:08:25,630 --> 00:08:30,360 50% to be the second. The third. Y Yeah. 70 00:08:31,060 --> 00:08:35,230 And fourth, of course. D Also because you have 75%. 71 00:08:35,410 --> 00:08:44,500 And that's exactly the point. So we look not only at the actual ability to be in each position, but also what we call the cumulative probability. 72 00:08:44,980 --> 00:08:49,750 So the probability to be among the best three is 75%. 73 00:08:49,750 --> 00:08:54,490 For a 100% would be 100% for C, only 25% for D. 74 00:08:55,450 --> 00:09:00,579 And this is very important because one of the most common mistakes in natural rate analysis is to 75 00:09:00,580 --> 00:09:07,240 consider only the probability to be in a specific position and for treatment with a lot of uncertainty. 76 00:09:07,510 --> 00:09:11,260 You have a huge probably to be the first huge probability to be the last. 77 00:09:11,830 --> 00:09:17,740 So if you look only the probability to be one position, you don't really understand which is the wrong of a treatment. 78 00:09:17,930 --> 00:09:28,180 We have a great example. There was a paper published in the BMJ about pharmacological treatment of GAD, the generalised anxiety disorder, 79 00:09:28,480 --> 00:09:35,290 and they use the ranking in the wrong way probabilities to be the best so called, and they rank fluoxetine as the best treatment. 80 00:09:35,560 --> 00:09:40,450 And of course it was the wrong approach. We we wrote a letter to the BMJ. 81 00:09:40,750 --> 00:09:43,120 They didn't want to publish our reanalysis. 82 00:09:43,390 --> 00:09:51,580 And I was very happy now that a couple of weeks ago in The Lancet you find in that too much analysis about good done properly. 83 00:09:51,910 --> 00:10:02,020 Okay, so it's a very common mistake. So this is a sorry for the theoretical background, but it's just to be sure we are on the same page. 84 00:10:02,200 --> 00:10:10,179 So now the story about the antidepressants we started in 2009 publishing this a paper in The Lancet, 85 00:10:10,180 --> 00:10:18,460 which was the first at that time we called it multiple treatments, met analysis, the first natural met analysis in psychiatry. 86 00:10:19,720 --> 00:10:23,860 And at that time, I was finishing my residency in psychiatry. 87 00:10:24,130 --> 00:10:28,390 And one of the key questions, I was in Italy. I'm Italian, by the way. 88 00:10:28,510 --> 00:10:31,650 I'm sure you can pick up my terrible English accent. 89 00:10:32,110 --> 00:10:44,169 Anyway, I was finishing my writing residency and one of the key question was for patients we used to see in our clinics with the 90 00:10:44,170 --> 00:10:52,239 patient with depression was which is the antidepressant to prescribe to these patients and asking my senior colleagues, 91 00:10:52,240 --> 00:10:59,800 they said, well, usually the last one, the last marketed antidepressant is the best because they are all the same in terms of efficacy, 92 00:11:00,070 --> 00:11:08,500 slightly different terms of acceptability, but the newest is probably the best because they have less sexual dysfunction or they are better tolerated. 93 00:11:08,710 --> 00:11:13,810 I didn't buy the argument. I still was in favour of an evidence based approach. 94 00:11:13,810 --> 00:11:22,450 So what we did was to carry out this systematic review, which is actually a systematic review of 12 systematic reviews. 95 00:11:23,110 --> 00:11:28,030 And the clinical question was very clear to me. So I have a patient in front of me. 96 00:11:28,720 --> 00:11:34,510 The patient is depressed and the patient has agreed to start a pharmacological treatment. 97 00:11:34,840 --> 00:11:39,070 Okay. So which is, if any, the best treatment for this patient? 98 00:11:39,580 --> 00:11:47,170 So for this reason, what we looked at was only second generation antidepressants because according to the guidelines, 99 00:11:47,410 --> 00:11:54,370 we should prescribe a second generation antidepressant. We did not include placebo because the question was not whether to the presence work. 100 00:11:54,610 --> 00:11:58,990 The question was, which is because we already decided to prescribe an antidepressant. 101 00:11:59,320 --> 00:12:07,870 And of course, because the patient was depressed, the outcome was acute efficacy and acceptability of this treatment. 102 00:12:08,620 --> 00:12:12,669 So the project was called Mangi. 103 00:12:12,670 --> 00:12:20,260 We included 117 randomised controlled trial, 26,000 participants, about 12 drugs. 104 00:12:20,500 --> 00:12:22,930 And I think names are important. 105 00:12:23,290 --> 00:12:31,809 Names are important because especially in network meta analysis, you are working with a multidisciplinary team and there's a lot of email exchange. 106 00:12:31,810 --> 00:12:37,390 So it's good to have a title like Manga Meta Analysis of New Generation Antidepressants. 107 00:12:38,140 --> 00:12:43,180 As I said, we included only active drugs, no placebo, no all the comparator, 108 00:12:43,180 --> 00:12:47,170 because according to guidelines, second generation antidepressants should be preferred. 109 00:12:47,560 --> 00:12:50,469 And it was the first natural meta analysis. 110 00:12:50,470 --> 00:12:57,190 So we decided to keep it simple, but we wanted to look at two primary outcomes which were clinical informative. 111 00:12:57,220 --> 00:13:02,290 One was response. So after eight weeks of treatments, how many patients respond? 112 00:13:02,500 --> 00:13:08,590 50% reduction on a rating scale, which is pretty standard measure second was at eight weeks. 113 00:13:08,590 --> 00:13:13,540 How many patients are still on treatment? So called acceptability dropout rate? 114 00:13:15,160 --> 00:13:24,550 Of course, we had a lot of there was a lot of controversy around this met analysis because for the first time we proved that there. 115 00:13:25,000 --> 00:13:33,309 Differences between antidepressants. But a lot of criticism was about the exclusion of all comparators, the exclusion of placebo, 116 00:13:33,310 --> 00:13:37,840 because there's a lot of debate in the literature whether antidepressants really work. 117 00:13:38,260 --> 00:13:50,290 So after a few years, actually, we started in 2011, 2012, and we published a paper six years after in 2018. 118 00:13:50,290 --> 00:13:58,990 So it took a long, long time to do Greece out of that group of research and investigating specific efficacy of individual drugs for acute depression. 119 00:13:59,470 --> 00:14:04,300 It's it's a long title, but it's a very nice story about Griselda. 120 00:14:04,810 --> 00:14:14,470 I don't know if you're familiar. She's a fantastic woman that basically when I started the project, I knew it was something very long. 121 00:14:15,100 --> 00:14:21,340 And we basically what we needed word was a lot of determination, 122 00:14:21,550 --> 00:14:31,840 endurance and will to fight against all the difficult things we may have encountered or during and during the process. 123 00:14:32,050 --> 00:14:43,120 And Griselda is a lady that was basically it's a difficult story today because she had everything she should not have had. 124 00:14:43,390 --> 00:14:54,730 Like she was treated badly by the husband. And basically she was a slave of her husband, treated poorly and neglected, and in the end, 125 00:14:55,110 --> 00:15:00,280 the husband to punish her for nothing because she didn't do anything wrong. 126 00:15:00,730 --> 00:15:10,090 She he pretended that he was marrying their daughter and she was asked to be the witness of the marriage that something terrible. 127 00:15:10,570 --> 00:15:19,150 But in the end, all these was a sort of test of the husband to test how good and how strong she was. 128 00:15:19,540 --> 00:15:22,960 So terrible story. Nice, happy ending. 129 00:15:23,110 --> 00:15:30,460 But the very reason why I like the story of Griselda is not only because it's true that women are much better than men, 130 00:15:30,760 --> 00:15:40,570 but also because if you ask when I present this story to British people, they know, they say, well, I know Griselda, it's chose her. 131 00:15:41,050 --> 00:15:46,600 It's a novel, a story narrated by Chaucer in his Canterbury Tales. 132 00:15:47,020 --> 00:15:50,530 But that's where British people are wrong because it's Italian. 133 00:15:50,890 --> 00:15:54,640 So Tozer decided to steal the story from Boccaccio. 134 00:15:54,820 --> 00:16:03,340 And in this specific period of Brexit, I'm so proud to keep saying that these are actually Italy and not UK anyway. 135 00:16:03,580 --> 00:16:08,980 It's a personal revenge because last night I was very, very worried about the Parliament. 136 00:16:09,950 --> 00:16:15,159 Okay, so for Griselda, we had to rethink the project. 137 00:16:15,160 --> 00:16:18,819 And the first question was, do antidepressants work for depression? 138 00:16:18,820 --> 00:16:22,500 So are they an effective treatment for depression? 139 00:16:22,510 --> 00:16:26,620 First question if yes, which is if any of the best treatment. 140 00:16:26,830 --> 00:16:37,000 So this is why in grizzle that we have placebo, we included also not only the two previous outcomes, but we also included continuous outcomes. 141 00:16:37,270 --> 00:16:41,950 We wanted to expand the number of drugs. 142 00:16:41,950 --> 00:16:48,250 So not only the newer or the newest antidepressants on the market, but also all the comparators. 143 00:16:48,490 --> 00:16:58,090 And to be methodologically sound, we decided to include antidepressants which are in WTO essential list of medicine to be informative globally. 144 00:16:58,930 --> 00:17:06,550 Of course, we were very careful about the search strategy, including publish and publish and also all the regulatory stuff about drugs. 145 00:17:06,850 --> 00:17:19,450 And in the end, we have 21 active treatments and placebo, 5522 double blind randomised controlled trials, less than 120,000 participants. 146 00:17:19,870 --> 00:17:23,890 Just to give you an idea, this is the largest net promoter analysis in medicine. 147 00:17:25,540 --> 00:17:30,309 So as I said, the rationale is long lasting debate about efficacy of antidepressants. 148 00:17:30,310 --> 00:17:34,030 Are there really differences between these individual drugs? 149 00:17:34,540 --> 00:17:39,670 So we focussed again on the acute treatment of adults with unipolar, major depressive disorder. 150 00:17:39,910 --> 00:17:44,020 As you know, we need to have a specific question to answer. 151 00:17:44,290 --> 00:17:50,409 We have parallel projects about long term effects or treatment resistant depression. 152 00:17:50,410 --> 00:17:56,290 But this paper is only about the acute treatment of adults with unipolar depressive disorder. 153 00:17:56,620 --> 00:18:04,330 The reason why we, especially for not too much analysis we need to focus is because of an assumption we do with measurement analysis, 154 00:18:04,510 --> 00:18:06,010 which is called transitivity. 155 00:18:06,490 --> 00:18:16,420 So when we have the network and this is a good example, if we have a network of treatment to be able to do a network meta analysis, 156 00:18:16,660 --> 00:18:24,640 the assumption of transitivity is crucial because basically we assume that each patient included. 157 00:18:24,970 --> 00:18:31,600 In each of the treatment could be theory randomised to any of the other treatments. 158 00:18:32,350 --> 00:18:41,350 So that's the transitivity which allows us to use the indirect comparison to compare two interventions which haven't been compared. 159 00:18:41,710 --> 00:18:50,920 And the problem, I mean, how we violate transitivity is also important not only in terms of how we deliver the treatment. 160 00:18:50,920 --> 00:18:52,510 For instance, we include placebo, 161 00:18:52,750 --> 00:19:01,780 but we cannot have placebo as an injection compared to placebo as a tablet because especially for depression or pain, 162 00:19:02,380 --> 00:19:07,750 the way you administer placebo can have an effect on the outcome. 163 00:19:08,050 --> 00:19:17,200 And the other important thing, especially in the field of antidepressants, is the effect of the drugs should be stable over time. 164 00:19:17,980 --> 00:19:24,760 Because if you compare and all these are all new drugs, but for all comparisons like amitriptyline, 165 00:19:25,000 --> 00:19:32,620 if you compare a drug with placebo and in the seventies the response rates to placebo was 20% 166 00:19:33,100 --> 00:19:40,450 and you compare escitalopram with placebo in 2000 and the response rate of placebo is 40%. 167 00:19:40,870 --> 00:19:48,730 This is a violation of transitivity because you lumped together placebo assuming that we have the same effect modifiers. 168 00:19:48,940 --> 00:19:49,960 But this is different. 169 00:19:50,260 --> 00:20:01,990 So the first thing we had to do was to be sure that we were comparing similar with similar and that the transitivity assumption would be confirmed. 170 00:20:02,890 --> 00:20:05,920 So we go back to these in a moment. 171 00:20:05,920 --> 00:20:13,270 So all lies and second generation antidepressants, not only in Europe, for instance, in Japan, they don't have a vaccine, but they have been nasty. 172 00:20:13,570 --> 00:20:20,380 So this is why I mean, aspirin is in the list. We have full reference, first generation antidepressants and placebo. 173 00:20:20,710 --> 00:20:30,430 This is the paper which is open access, which means that you have to pay a lot to the Lancet $5,000. 174 00:20:31,670 --> 00:20:37,790 It's amazing. But it's open access, also the data set. 175 00:20:38,030 --> 00:20:43,190 So all the data we analysed, six years of work are freely available on the website. 176 00:20:43,430 --> 00:20:51,320 In order to incentivise people to replicate the analysis and find any mistakes we may have made. 177 00:20:52,670 --> 00:20:58,639 So as I said, only double blind randomised controlled trials published in unpublished antidepressants 178 00:20:58,640 --> 00:21:06,710 as monotherapy oral adults with a primary diagnosis of MDD major depressive disorder. 179 00:21:06,950 --> 00:21:09,589 We use standardised criteria for instance. 180 00:21:09,590 --> 00:21:20,780 For this reason we had to exclude a lot of Chinese studies and we also consider excluded people who had depression with cancer or are like, 181 00:21:20,990 --> 00:21:25,550 ah, rheumatoid arthritis of these kind of medical comorbidities. 182 00:21:26,240 --> 00:21:36,860 The search is up to January 2016 because it took us about a year and a half to analyse the data and find the right approach. 183 00:21:37,580 --> 00:21:47,270 And we searched manually all the international trial registries, the websites of regulatory agencies, and also we contacted all the study authors. 184 00:21:47,270 --> 00:21:56,239 So this is why it took us a lot of time. This is the list of 21 drugs up to here is the list of new generation 185 00:21:56,240 --> 00:22:00,590 antidepressants from my go into 40 sitting in a sabbatical order and these are the 186 00:22:00,590 --> 00:22:07,250 for reference drug I tripling compare mean among the TCAS trazodone and the 187 00:22:07,400 --> 00:22:11,720 incident because they are antidepressants with a specific side effect profile. 188 00:22:11,960 --> 00:22:20,540 For instance, trazodone is often used by GP's because of the A side effect profile in terms of insomnia. 189 00:22:21,470 --> 00:22:26,680 As I said, we included bipolar psychotic disorder, depression, treatment resistant depression, 190 00:22:26,690 --> 00:22:34,040 serious mental illness and this is the list of outcomes, the two historical one. 191 00:22:34,280 --> 00:22:42,560 Just to be sure, we could compare the results of 2018 with 20 or nine because things as we know can change in evidence based medicine. 192 00:22:42,800 --> 00:22:48,710 But we added continuous outcome change on symptoms, we added remission to response. 193 00:22:49,010 --> 00:22:54,380 And also not only dropouts, all college dropouts, but also dropouts due to adverse events. 194 00:22:55,190 --> 00:23:04,610 Because these is a review to inform clinical practice, we included data only within the therapeutic range of the antidepressants. 195 00:23:04,610 --> 00:23:10,459 We excluded low dose or too high dose in terms of statistical analysis. 196 00:23:10,460 --> 00:23:16,880 Of course, we use alterations in the dose mean differences because we have both continuous and dichotomous outcomes. 197 00:23:17,600 --> 00:23:24,980 We look at the heterogeneity. Of course the transitivity assumption was evaluated looking at the distribution of the 198 00:23:25,250 --> 00:23:31,760 clinical and methodological variables that could affect could mean effect modifiers. 199 00:23:32,510 --> 00:23:37,910 But transitivity we cannot test transitivity with a statistical method. 200 00:23:38,240 --> 00:23:46,040 So this is why all the discussion is mainly about the clinical issues that may have an effect as effect modifier. 201 00:23:46,730 --> 00:23:55,430 Why consistency? We have quite robust methods to check consistency and we use both the local method and the global method. 202 00:23:56,180 --> 00:24:05,570 I'm very proud of this because in the review team I wanted three statisticians to analyse the data independently using three different softwares, 203 00:24:05,780 --> 00:24:14,150 just to be sure that the results were consistent, consistent independently of the way or the statistical software you analyse them. 204 00:24:14,540 --> 00:24:20,390 And of course for the certainty of evidence we use the standard which is now agreed. 205 00:24:21,980 --> 00:24:31,940 We took the opportunity also to carry out subgroup and sensitivity analysis because for instance, we know that study can be an effect. 206 00:24:32,180 --> 00:24:39,310 Modifier sponsorship is a big issue in psychiatry, especially in antidepressants, literature, severity, 207 00:24:39,350 --> 00:24:47,630 baseline, the dosing schedule and also the small side effect, which is somehow a proxy of publication bias. 208 00:24:48,230 --> 00:24:52,670 And the novelty effect, because the same drug, when is the experimental drug, 209 00:24:52,670 --> 00:25:02,300 tend to be more effective than when the same drug is the comparator, which is a strange thing, which is not justified by the biology. 210 00:25:02,480 --> 00:25:07,550 So we look also for the novelty effect and in terms of sensitivity analysis, 211 00:25:07,820 --> 00:25:17,180 because we imputed response rate from the continuous data, we excluded studies which did not report a response rate. 212 00:25:17,450 --> 00:25:21,919 Of course we excluded study that did not use all accepted those in all arms. 213 00:25:21,920 --> 00:25:26,060 So not only the arms but all the study, the unpublished data. 214 00:25:26,420 --> 00:25:29,930 Yes. Because we look for unpublished data. 215 00:25:29,930 --> 00:25:39,990 Systematic. So if the same study reported, published and unpublished data, we gave preference to the unpublished data assuming they are more reliable. 216 00:25:40,440 --> 00:25:46,110 So we included only studies which provided also unpublished data and and so on. 217 00:25:47,520 --> 00:25:50,879 So this is the, the flow chart. 218 00:25:50,880 --> 00:25:54,690 We started from around 30,000 references. 219 00:25:54,960 --> 00:25:59,490 And from this side is the electronic data sets we had. 220 00:26:00,180 --> 00:26:07,140 We included about 420 studies from the electronic datasets databases. 221 00:26:07,590 --> 00:26:17,370 But we found 86 unpublished studies and also by a personal communication or hand searching of reference list, 222 00:26:17,520 --> 00:26:20,130 additional 15 randomised controlled trials. 223 00:26:20,610 --> 00:26:32,520 So in the end, the 522 are in the large majority from electronic databases, but a significant proportion of the data is unpublished as expected. 224 00:26:33,240 --> 00:26:42,330 This information, I think, is also important. We reported the number of studies per drug compared to placebo or another active comparator. 225 00:26:42,630 --> 00:26:49,680 And of course, drugs like fluoxetine or paroxetine, they have more than 110 trials. 226 00:26:50,070 --> 00:26:55,330 Most recent drug like does similar vaccine or level in nasty. 227 00:26:55,680 --> 00:26:59,520 They have only less than ten studies and that's expected. 228 00:26:59,730 --> 00:27:05,790 However, the the bottom line message is that the amount of evidence is different across different drugs. 229 00:27:06,930 --> 00:27:14,160 These are the two networks network network plots for efficacy and acceptability. 230 00:27:15,720 --> 00:27:18,930 Can I ask you what you think about this? 231 00:27:19,140 --> 00:27:28,140 I told you that the shape of the network is very important to to gather some information, to guide the interpretation of the results. 232 00:27:28,470 --> 00:27:31,830 So what do you think about this? To. 233 00:27:34,020 --> 00:27:37,960 So the colours are different. I know, but. Exactly. 234 00:27:38,310 --> 00:27:45,090 That's. That's one of the important things. So the shape of the network is basically identical. 235 00:27:45,570 --> 00:27:52,830 The only difference, if you can see, is that between Bupropion and Tripoli, we don't have studies reporting efficacy. 236 00:27:53,100 --> 00:27:56,580 We have one studies reporting data about acceptability. 237 00:27:56,850 --> 00:28:05,280 But apart from these and another couple of examples, the same studies reported data about response and dropouts, 238 00:28:05,700 --> 00:28:13,170 which is good for us because the initial assumption was the primary outcome is a combination of response and dropouts, 239 00:28:13,560 --> 00:28:21,690 because we want to know whether after eight weeks people who responded and whether they are still on treatment, 240 00:28:21,690 --> 00:28:25,770 which is a proxy of acceptability and tolerability as well. 241 00:28:26,460 --> 00:28:29,820 So the shape of the network is almost identical. 242 00:28:30,120 --> 00:28:38,490 And the other thing which is important from a methodological point of view, is that in the end the network is very well connected. 243 00:28:38,820 --> 00:28:41,880 So we have a lot of head to head comparisons. 244 00:28:42,540 --> 00:28:50,310 I don't know how familiar you are with the literature in psychiatry, but the great majority of trials are placebo controlled. 245 00:28:50,640 --> 00:28:55,320 We did a natural meta analysis about acute mania published in The Lancet a few years ago, 246 00:28:55,680 --> 00:29:03,000 and the shape of the network is called Star Shape Network because you have placebo in the middle. 247 00:29:03,210 --> 00:29:08,580 A lot of comparison with placebo, but only if you had to have this like like a star. 248 00:29:08,790 --> 00:29:13,739 So this is a very well-connected network, which means for us we can check inconsistency. 249 00:29:13,740 --> 00:29:17,820 So we have more, more. The results are much more robust. 250 00:29:18,000 --> 00:29:22,829 Sorry. So let's ask the first question. 251 00:29:22,830 --> 00:29:28,380 So whether antidepressants work. So the comparison here is drug versus placebo. 252 00:29:28,980 --> 00:29:34,560 I use the standard forest plots, but these are network meta analysis data. 253 00:29:34,830 --> 00:29:43,170 Okay. So including both direct and indirect and on the right favours drug on the left favours placebo. 254 00:29:43,560 --> 00:29:49,680 As you can see here, all the drugs are statistically significantly better than placebo. 255 00:29:49,680 --> 00:29:50,880 Everything is on the right. 256 00:29:51,240 --> 00:29:58,880 And also there are some differences between, for instance, the top one amitriptyline and the least effective Rybak sitting. 257 00:29:59,580 --> 00:30:05,310 This is a material difference because if you look at the confidence interval, they do not overlap. 258 00:30:05,790 --> 00:30:10,860 So in the end, we have completely different treatments in terms of efficacy versus placebo, 259 00:30:11,160 --> 00:30:16,559 but all antidepressants are statistically significantly better than placebo. 260 00:30:16,560 --> 00:30:19,860 In terms of efficacy, in terms of acceptability, 261 00:30:20,940 --> 00:30:28,230 the great majority of drugs are not statistically significantly different in terms of dropout rate or cause dropouts. 262 00:30:28,710 --> 00:30:35,490 A couple like accumulating and fluoxetine statistically significantly better compare mean significant, 263 00:30:35,700 --> 00:30:40,110 statistically significantly worse in terms of acceptability than placebo. 264 00:30:40,320 --> 00:30:45,720 But if you look at the confidence interval is very close to that line of no difference. 265 00:30:45,990 --> 00:30:50,940 So the figure of the paper is combining efficacy and acceptability. 266 00:30:51,210 --> 00:30:55,740 Of course, in order to make it visually like this, I had to swap. 267 00:30:56,040 --> 00:31:00,750 So you can see here that the confidence interval and the estimates are the other way around. 268 00:31:00,960 --> 00:31:04,740 So it's a bit complicated. So this is 1.5 to 2.5. 269 00:31:04,980 --> 00:31:14,730 These is the other way around. But because my paper was for clinicians, you give the message on the right favours drug on the left favours placebo. 270 00:31:17,010 --> 00:31:25,170 Okay. The big question is whether these differences translate into something which is clinically meaningful. 271 00:31:25,980 --> 00:31:33,660 But this is the second part of the presentation. So the remarks that this is the least effective. 272 00:31:34,590 --> 00:31:38,580 Can we try to combine the two outcomes in one graph? 273 00:31:39,450 --> 00:31:43,889 So this is what we did. So we have efficacy here. 274 00:31:43,890 --> 00:31:56,400 It's the same data, same results. So efficacy on the x axis one using the references, Reebok City, these are all the other data and this is placebo. 275 00:31:56,820 --> 00:32:00,480 So in terms of efficacy, so this is the line of no difference. 276 00:32:00,750 --> 00:32:05,040 Antidepressants are better than placebo and this is acceptability. 277 00:32:05,040 --> 00:32:10,830 This is the line of the difference. You can see that the great majority of antidepressants are similar to placebo. 278 00:32:11,190 --> 00:32:17,070 A few are less acceptable. But this is the graph. 279 00:32:17,610 --> 00:32:24,570 If we use all studies, what happens if we exclude placebo controlled trials? 280 00:32:25,470 --> 00:32:30,210 If you exclude placebo controlled trials, keep these as a reference. 281 00:32:30,540 --> 00:32:35,360 Look at what happens. So this is with placebo. 282 00:32:36,410 --> 00:32:40,940 If we exclude all placebo controlled trials. So we have only had two trials. 283 00:32:41,330 --> 00:32:46,430 This is the figure. So and the box it in is in the same place is always the reference. 284 00:32:46,820 --> 00:32:53,990 So basically if you include placebo in the network, all the differences between antidepressants shrink. 285 00:32:55,820 --> 00:33:03,740 So the problem is to interpret this. Of course, we reported the figure of the paper is this one. 286 00:33:03,750 --> 00:33:13,399 So we had all studies head to head. But because we want we need to interpret the data from a clinical point of view, which is the truth, of course. 287 00:33:13,400 --> 00:33:17,360 I don't know. I'm just trying to suggest some interpretation. 288 00:33:17,570 --> 00:33:22,219 So what we did was to analyse the question specifically, 289 00:33:22,220 --> 00:33:28,940 and the paper is now published in the Journal of Clinical Epidemiology, so International Journal of Epidemiology. 290 00:33:30,200 --> 00:33:37,040 So we look at the impact of placebo arms on the outcomes of antidepressant trials, 291 00:33:37,040 --> 00:33:43,310 and the outcomes are efficacy, of course, dropout and dropouts due to adverse events. 292 00:33:44,750 --> 00:33:55,220 So what we did was to, as an outcome, look at the probability of being allocated to placebo, the pie as dichotomous yes or no. 293 00:33:55,790 --> 00:33:59,120 Try continuous 0% is head to head. 294 00:33:59,570 --> 00:34:02,570 50% is two on placebo controlled trial. 295 00:34:02,870 --> 00:34:09,260 Anything in between these multi arm placebo controlled trial or as a continuous variable? 296 00:34:10,070 --> 00:34:16,550 Of course, because we had to have the same drug tested in placebo and in head to head alone, 297 00:34:17,000 --> 00:34:22,380 the number of studies is less than the original number of 522 and so on. 298 00:34:22,400 --> 00:34:26,660 We in the end included 700 and something arms. 299 00:34:27,290 --> 00:34:32,689 So these is for response and these is the overall results. 300 00:34:32,690 --> 00:34:38,840 So the light blue is placebo controlled trial to are placebo controlled trial. 301 00:34:39,170 --> 00:34:47,540 The very dark blue, almost black, let's say is head to head trials and in the middle is multi arm placebo controlled. 302 00:34:47,870 --> 00:34:51,790 Okay. So if you look at the whole can you touch this or. 303 00:34:52,010 --> 00:35:02,209 Yeah, okay. So if you look at the overall results being in a two arm placebo controlled trials of drug versus 304 00:35:02,210 --> 00:35:11,420 placebo response rate is less than being in a head to head trial without placebo first finding. 305 00:35:11,720 --> 00:35:15,590 But if you look at the specific individual drugs, 306 00:35:16,130 --> 00:35:23,360 this pattern is quite typical for many of these drugs escitalopram citalopram for toxicity and blah blah blah because of pain. 307 00:35:23,690 --> 00:35:26,660 But it's completely different for drugs like amitriptyline. 308 00:35:27,140 --> 00:35:34,190 So in amitriptyline trial, being in a placebo control, I mean typically versus placebo trial, 309 00:35:34,520 --> 00:35:40,790 the response rate is much higher than being in amitriptyline versus an active drug with our placebo. 310 00:35:40,820 --> 00:35:44,389 So the opposite. So they are ordered. 311 00:35:44,390 --> 00:35:48,650 I think from here to here is the order with the block. 312 00:35:48,650 --> 00:35:57,170 So the probability of receiving so the head to head trials and these is the all overall it's not a sabbatical order. 313 00:35:57,890 --> 00:36:06,110 If you look at the all cause dropouts, you can see that being in a placebo controlled trial to arm, 314 00:36:06,620 --> 00:36:11,390 you have a higher probability of dropping out than being in head to head. 315 00:36:12,500 --> 00:36:16,370 And again, there are differences between individual drugs. 316 00:36:17,360 --> 00:36:22,550 If you look at dropouts due to side effects, there's no difference. 317 00:36:23,360 --> 00:36:27,290 So in the end, I think these is the most important figure of the paper. 318 00:36:27,590 --> 00:36:42,049 If you plot, again, efficacy Y and dropout in the X axis and the black dots are the head to head trials and 319 00:36:42,050 --> 00:36:47,540 the blue dots are the placebo controlled trials for the same drug for the same drug. 320 00:36:47,960 --> 00:36:52,610 Being in a head to head trial, you have a higher probability to respond, 321 00:36:52,610 --> 00:36:58,220 less probability to dropout than being in a placebo controlled trial for the same drug. 322 00:36:58,790 --> 00:37:05,869 So it's not about the drug is about the study design. So in the paper, we suggest possible explanations. 323 00:37:05,870 --> 00:37:14,960 And these are the two extremes. One is, of course, in, plus to and placebo controlled trial versus head to head. 324 00:37:15,350 --> 00:37:22,820 We can measure the response to the drug, but of course, we cannot measure the context to our response, 325 00:37:22,970 --> 00:37:30,710 what we what we call the placebo response, because we cannot measure this in a head to head trial because there is no placebo. 326 00:37:31,220 --> 00:37:39,020 Okay. So the only way to understand whether the difference between the two is related to this placebo effect or 327 00:37:39,020 --> 00:37:46,520 sort of contextual response would be to cheat two patients would be to grant a placebo controlled trial, 328 00:37:46,760 --> 00:37:51,410 which is actually a head to head and vice versa, which is not ethical. 329 00:37:51,740 --> 00:37:58,399 Well, I thought it was not ethical until I went to Australia to present this data to a 330 00:37:58,400 --> 00:38:03,410 group of colleagues who are working on pain and they are doing this kind of trials. 331 00:38:03,740 --> 00:38:06,950 So they treat two patients. They say this is a placebo controlled trial. 332 00:38:07,280 --> 00:38:14,299 So ethically I think it's more difficult because, you know, with depression we have the issues about suicide and or self-harming. 333 00:38:14,300 --> 00:38:23,300 So it's not an easy thing to do. But I think this is the only way to methodologically sound answer the clinical question. 334 00:38:23,330 --> 00:38:28,879 So we are working on this. But there's another thing about antidepressant trials. 335 00:38:28,880 --> 00:38:36,440 I don't know how familiar you are with the literature, but if this is the typical trajectory in a trial, 336 00:38:36,800 --> 00:38:41,360 so this is the change in symptoms over time, over eight weeks, 337 00:38:41,750 --> 00:38:47,209 you start with the same severity and after eight weeks people randomised to 338 00:38:47,210 --> 00:38:53,120 placebo respond less than people randomised to the highest dose of the drug. 339 00:38:53,150 --> 00:38:56,600 This is the lower dose of the drug and this is the active comparator. 340 00:38:56,600 --> 00:39:01,639 Okay, this is a typical situation. But because they a sorry. 341 00:39:01,640 --> 00:39:07,010 And at the end the difference between placebo and the drug is five points on the madras 342 00:39:07,010 --> 00:39:13,550 which is a standardised rating scale to measure depressive symptoms because there's placebo. 343 00:39:13,700 --> 00:39:18,590 And we know that people in placebo controlled trials tend to drop out early. 344 00:39:18,920 --> 00:39:22,280 What happens is that they drop out, let's say, after one week. 345 00:39:23,090 --> 00:39:32,000 And in order not to miss these information, what researchers and trialists do is to take the last observation, 346 00:39:32,120 --> 00:39:37,549 and they carry forward this until the end of the study because as I said, 347 00:39:37,550 --> 00:39:43,340 in placebo controlled trials also people randomised to the drug drop out early so they carry 348 00:39:43,340 --> 00:39:49,460 out the last observation and in the end the difference they measure is one instead of five. 349 00:39:50,540 --> 00:39:58,999 Do you know how many? Oh, I know in psychiatry we are far behind other fields of medicine because I know that very, 350 00:39:59,000 --> 00:40:07,250 very I mean, more sophisticated measure methods to basically tackle the issue of missing data. 351 00:40:07,430 --> 00:40:12,200 But do you know how many studies in the sample of Griselda? 352 00:40:12,200 --> 00:40:18,860 So out of 522, use the LCF method to impute four missing data. 353 00:40:19,160 --> 00:40:25,340 95 So the thing is, these may be something which bias the results. 354 00:40:25,700 --> 00:40:29,810 So in the end, all this is why the league table, 355 00:40:29,840 --> 00:40:39,620 which reports response rate and dropout rate in one table is based on the head to head and not on the placebo controlled trial. 356 00:40:40,070 --> 00:40:42,560 So these are the drugs in alphabetical order, 357 00:40:42,740 --> 00:40:51,470 and the idea is to be able to compare the column defining versus zero defining and be able to estimate the head to head. 358 00:40:51,740 --> 00:40:57,470 So if we have fluoxetine versus circulating certainly is better than fluoxetine. 359 00:40:57,590 --> 00:41:04,579 Escitalopram is better than a box, sitting escitalopram is worse than citalopram and so on. 360 00:41:04,580 --> 00:41:13,370 So there are highlighted and bold, statistically significant results and for efficacy and the same for acceptability. 361 00:41:15,350 --> 00:41:20,299 I will move ahead. We can go back to the results if you want limitations. 362 00:41:20,300 --> 00:41:26,390 There are many limitations of the study. Of course, we may have missed studies, but there's no evidence of publication bias. 363 00:41:26,690 --> 00:41:30,650 I didn't mention the quality according to grade was moderate to low or very low. 364 00:41:30,890 --> 00:41:35,209 So there's a big debate in terms of how good are these results. 365 00:41:35,210 --> 00:41:43,790 But at the same time, this is the best evidence we have. We didn't do any cost effectiveness analysis because we just wanted to inform the globe. 366 00:41:44,270 --> 00:41:48,890 So in order to do the cost effectiveness analysis, you need to use data in individual countries. 367 00:41:49,250 --> 00:41:50,660 This is an ongoing project. 368 00:41:51,200 --> 00:42:01,250 Of course, we may have missed some potentially important effect modifiers and of course the number of outcomes reported here is limited. 369 00:42:01,490 --> 00:42:05,540 We are now finalising the manuscript about adverse events, 370 00:42:05,780 --> 00:42:13,310 but it took us about another year to develop the methodology to analyse the adverse events in the natural matter analysis. 371 00:42:13,580 --> 00:42:17,750 But I think this is the main problem and this is where the controversy was in the end. 372 00:42:17,750 --> 00:42:25,670 These are data which are aggregate. So it's not we're still not able to personalise treatment in depression, 373 00:42:25,670 --> 00:42:31,670 even though it's much better to have a league table like this, we are still talking about average. 374 00:42:33,300 --> 00:42:36,360 So this is why we need to move a step forward. 375 00:42:36,630 --> 00:42:40,020 And can we use natural methodologies to personalise twittering depression? 376 00:42:40,260 --> 00:42:47,050 Of course, I think the answer is yes. This is a paper we published a few months ago in psychotherapy and psychosomatic. 377 00:42:48,000 --> 00:42:53,460 And this is a sort of proof of concept whether we can predict treatment outcome using, 378 00:42:54,600 --> 00:42:59,910 I mean, clinical baseline and demographic characteristics of patients. 379 00:43:00,300 --> 00:43:03,960 So for this specific project, which is about a depression, 380 00:43:04,260 --> 00:43:11,940 we used an individual patient data network meta analysis which investigated the use of 381 00:43:12,240 --> 00:43:19,020 antidepressants alone versus psychotherapy alone versus the combination in chronic depression. 382 00:43:19,740 --> 00:43:25,050 We we decided to choose the example because we did a systematic review. 383 00:43:25,320 --> 00:43:30,210 We had only four studies and we managed to have the individual data of all the studies. 384 00:43:30,360 --> 00:43:37,080 And because it's four studies with three nodes, we have a good representation of the evidence to test the method. 385 00:43:37,560 --> 00:43:45,750 But basically what we did was depending on the demographic characteristic of these patients. 386 00:43:45,750 --> 00:43:51,180 So being female with severe depression, severe anxiety, very young, 31, 387 00:43:51,360 --> 00:43:56,759 no family history of depression, short duration of symptoms, the trajectory of symptoms, 388 00:43:56,760 --> 00:44:01,920 how the symptoms change over 12 weeks for treatment A, B, and C, 389 00:44:02,160 --> 00:44:08,760 it's clear because this is a change, that C is the best option for Helen, but for a men. 390 00:44:08,910 --> 00:44:19,440 So being men, nothing less, severe depression, severe anxiety, slightly old family history of depression, a longer duration of symptoms. 391 00:44:19,480 --> 00:44:26,970 Actually, option B is the best treatment. So with the individual patient data, we can show the patients, which is the trajectory. 392 00:44:27,450 --> 00:44:35,700 How fast is the outcome plus the outcome to to to to to B and also which is the best treatment. 393 00:44:36,240 --> 00:44:38,600 But these is only part of the project. 394 00:44:38,610 --> 00:44:48,060 The other big important issue is to incorporate patient preferences because we have a lot of adverse events and we know that some people can be. 395 00:44:49,440 --> 00:44:55,920 I mean, the burden of weight gain can be more for me or sexual dysfunction or headache or tremor. 396 00:44:56,220 --> 00:45:05,900 So the idea is to incorporate these in a treatment algorithm, and the treatment choice can inform really the ranking of treatment. 397 00:45:05,910 --> 00:45:12,990 So this is that how can I say the template and how the screen of the algorithm. 398 00:45:13,000 --> 00:45:18,510 So I mentioned Nick and this is his situation. 399 00:45:18,810 --> 00:45:27,480 But if the preference of Nick is these are sliding box so he doesn't want sedation, doesn't want weight gain and he's fine in terms of cardiac rate. 400 00:45:27,490 --> 00:45:33,450 So the cutesy prolongation is not an issue for him. This is the ranking of the treatment. 401 00:45:33,960 --> 00:45:45,300 But for many made with the same characteristics, same age, same severity of depression, anxiety, same family history, but with different preferences. 402 00:45:45,900 --> 00:45:50,540 So he doesn't want nausea, he doesn't want tremor, and he has a problem intimately discipline. 403 00:45:51,180 --> 00:45:55,530 The ranking is completely different. So this is where we are. 404 00:45:55,650 --> 00:45:59,610 We should go. And this is my Nature Reserve Professorship project. 405 00:45:59,940 --> 00:46:05,579 Be able to analyse individual patient data from the trials they showed combined with the 406 00:46:05,580 --> 00:46:13,320 observational data from real world patients and incorporate patient preferences into this algorithm. 407 00:46:15,300 --> 00:46:23,240 You are familiar with a master's, but I'm editor of these journal Evidence based Mental Health and we published this paper, of course, open access. 408 00:46:23,490 --> 00:46:28,740 So you can find all the things I mentioned about natural analysis there. 409 00:46:29,040 --> 00:46:30,929 And this is the promotional bit. 410 00:46:30,930 --> 00:46:41,459 We have this course in Oxford because I strongly believe that clinicians and clinical researchers should be aware of natural meta analysis, 411 00:46:41,460 --> 00:46:51,090 but especially critically appraise this kind of very trendy article in the scientific literature, but often very biased and very rubbish. 412 00:46:51,390 --> 00:46:51,810 Thank you.