1 00:00:03,080 --> 00:00:06,640 Greetings and welcome to EHA Unplug, the official podcast 2 00:00:06,640 --> 00:00:10,240 channel of the European Hematology Association EHA. 3 00:00:10,600 --> 00:00:13,600 Hello, I am Isabella Rivera, a medical writer for EHA. 4 00:00:14,400 --> 00:00:18,680 Today we're thrilled to have Doctor Jan Moritz Medeka, a 5 00:00:18,840 --> 00:00:21,600 distinguished hematologist and AI specialist. 6 00:00:22,880 --> 00:00:26,680 We're going to be talking today about AI to accelerate clinical 7 00:00:26,680 --> 00:00:28,480 trials. Welcome, Doctor. 8 00:00:28,480 --> 00:00:30,920 Medeka, can you please introduce yourself? 9 00:00:31,800 --> 00:00:35,760 Hi, yeah, thanks for having me. I'm, I'm Moritz Malika from the 10 00:00:35,760 --> 00:00:38,880 University Hospital in Dresden. I'm working there as a 11 00:00:38,880 --> 00:00:40,960 hematologist and have a. Research Group. 12 00:00:41,600 --> 00:00:47,840 On AI in hematology. So we want to understand how you 13 00:00:47,840 --> 00:00:51,840 can accelerate clinical trials by generating data with AI. 14 00:00:52,680 --> 00:00:55,720 Can you please explain first what is synthetic data? 15 00:00:56,400 --> 00:00:58,960 Synthetic data is as the name says. 16 00:00:59,240 --> 00:01:01,960 Data which is. Made-up and. 17 00:01:02,200 --> 00:01:06,520 The idea is that we face a. Big problem with clinical trials 18 00:01:06,640 --> 00:01:11,880 in general, but especially in hematology. 19 00:01:11,880 --> 00:01:14,400 Where a lot of the diseases are rare diseases for. 20 00:01:14,400 --> 00:01:16,520 Example acute myeloid leukemia is a rare. 21 00:01:16,520 --> 00:01:20,360 Disease per seat. And we already know that. 22 00:01:20,360 --> 00:01:24,560 There are several subgroups with targeted therapies, so the. 23 00:01:24,560 --> 00:01:28,360 Number of patients we can include in clinical trials are 24 00:01:28,360 --> 00:01:34,040 even lower and therefore we need to think about new ways how we 25 00:01:34,040 --> 00:01:37,160 can. Accelerate how we. 26 00:01:37,160 --> 00:01:43,200 Perform clinical trials in order to get sufficient data to get 27 00:01:43,200 --> 00:01:46,320 the knowledge about novel. Therapies. 28 00:01:47,640 --> 00:01:52,120 So what types of synthetic data are you generating that could be 29 00:01:52,120 --> 00:01:55,480 used for clinical trials? So the idea. 30 00:01:55,480 --> 00:01:57,280 Is that you have synthetic? Data which is. 31 00:01:57,560 --> 00:02:00,680 Reassembling the original cohort in a A. 32 00:02:00,720 --> 00:02:05,320 Most perfect way. But without being able. 33 00:02:05,320 --> 00:02:07,920 To get back. To the original data. 34 00:02:09,320 --> 00:02:12,400 So you're generating images. This can be used. 35 00:02:12,480 --> 00:02:14,000 Used both for. Images. 36 00:02:14,000 --> 00:02:17,560 But also for tabular. Data and what we did in recent 37 00:02:17,560 --> 00:02:22,160 study and also other groups have shown is that you can reassemble 38 00:02:22,600 --> 00:02:27,600 real clinical trial data in terms of baseline data, but also 39 00:02:27,600 --> 00:02:34,520 in terms of outcome results based on cohorts of of. 40 00:02:34,520 --> 00:02:36,720 Real. Patients and the synthetic data 41 00:02:36,720 --> 00:02:40,560 is reassembling all the information about the original 42 00:02:40,560 --> 00:02:42,600 patient which is like. Age. 43 00:02:42,600 --> 00:02:46,200 Sex. But also molecular changes and 44 00:02:46,200 --> 00:02:49,840 then the outcome after a given intervention. 45 00:02:51,200 --> 00:02:53,480 And how do you generate the synthetic data? 46 00:02:53,800 --> 00:02:57,600 So for this. We used guns, Generative 47 00:02:57,600 --> 00:03:04,280 Adversarial Networks, which is a known form of AI. 48 00:03:04,280 --> 00:03:06,600 Which has been. Used for image generation. 49 00:03:06,720 --> 00:03:08,160 Quite successful. Already. 50 00:03:08,440 --> 00:03:12,600 But can also. Be used for tabular data and in 51 00:03:12,600 --> 00:03:15,600 our specific work we use two types of these, so. 52 00:03:15,920 --> 00:03:21,200 Nflow and C Top Gun. So these are types of deep 53 00:03:21,200 --> 00:03:23,480 learning and it's generative modelling. 54 00:03:23,560 --> 00:03:28,040 You have described 4 scenarios in which synthetic data 55 00:03:28,360 --> 00:03:33,840 generation can be useful and I would like to walk through this 56 00:03:33,960 --> 00:03:39,720 for the audience. So first, AI generated data 57 00:03:40,560 --> 00:03:48,000 could help to prevent privacy concerns about clinical trial 58 00:03:48,000 --> 00:03:50,480 date data of the patients. Exactly so. 59 00:03:51,840 --> 00:03:55,560 The generated synthetic data can be shared without any. 60 00:03:55,560 --> 00:03:57,840 Privacy issues, so we checked that. 61 00:03:58,240 --> 00:04:02,120 There was no single patient in the synthetic cohort. 62 00:04:02,120 --> 00:04:05,560 Which was exactly. Matching the original patient. 63 00:04:06,760 --> 00:04:10,800 So there's no way that you can go back to the original patient 64 00:04:10,800 --> 00:04:12,920 which. Of course is the case if you. 65 00:04:13,000 --> 00:04:17,360 Just have an anonymized data, so this makes. 66 00:04:17,360 --> 00:04:20,399 It really easy. You can just upload these data 67 00:04:20,399 --> 00:04:22,880 and make them available. Also for other research. 68 00:04:22,880 --> 00:04:27,760 Group and see public. So how does this compare to the 69 00:04:27,760 --> 00:04:31,960 de identification or anonymisation of patient data? 70 00:04:31,960 --> 00:04:37,880 So anonymisation of these data isn't 100% sure, and with the 71 00:04:37,880 --> 00:04:41,720 synthetic data it's 100% sure. So there's no way that you can 72 00:04:41,720 --> 00:04:44,640 go back to the original patient. Because see, there's. 73 00:04:44,640 --> 00:04:49,840 No original data coming from one single patient. 74 00:04:50,360 --> 00:04:54,320 So it's basically 100% sure that you can't go back to the. 75 00:04:56,200 --> 00:04:58,600 Characteristics of. One single patient. 76 00:04:59,280 --> 00:05:02,600 So this is a big step forward for this kind of data. 77 00:05:02,720 --> 00:05:07,560 Also, you would create new patients and this would lead to 78 00:05:07,600 --> 00:05:12,120 either cohort augmentation, so some synthetic patients among 79 00:05:12,120 --> 00:05:17,440 patients that are real or completely a substitution and 80 00:05:17,440 --> 00:05:23,600 you have a complete new cohort. How robust is this process? 81 00:05:25,040 --> 00:05:28,000 So for the. Generation it's. 82 00:05:28,000 --> 00:05:30,960 Very robust. We showed and also other groups. 83 00:05:30,960 --> 00:05:33,120 Matteo de la Porta did this for MD's. 84 00:05:33,400 --> 00:05:37,280 Showed that it's. Reassembling the whole. 85 00:05:37,280 --> 00:05:40,440 Crestrix of the. Patients for a given disease. 86 00:05:40,440 --> 00:05:45,400 So it's not only age and sex in terms of the distribution, but 87 00:05:45,400 --> 00:05:49,520 it also comes down to the molecular changes, which is very 88 00:05:49,520 --> 00:05:52,640 important because we know that there are certain relationships 89 00:05:52,640 --> 00:05:55,800 between certain mutations. So for example, if you have an 90 00:05:55,920 --> 00:05:59,480 NPM 1 mutation, you're more likely to also have inflat 3 91 00:05:59,480 --> 00:06:02,720 mutation. And these correlations are 92 00:06:02,720 --> 00:06:04,800 preserved in the synthetic. Data set. 93 00:06:05,120 --> 00:06:06,840 And this. Holds also true. 94 00:06:06,840 --> 00:06:10,160 For correlation between age and certain cytogenetic 95 00:06:10,160 --> 00:06:13,120 abnormalities like complex. Karyotypes with with. 96 00:06:13,120 --> 00:06:16,120 Higher age and so on so. Biologically. 97 00:06:16,120 --> 00:06:19,040 Important correlations are. Preserved. 98 00:06:19,040 --> 00:06:22,760 Within this data and even more important, this also. 99 00:06:22,760 --> 00:06:26,680 Holds true for the. Correlation to the outcome data, 100 00:06:26,680 --> 00:06:29,760 so known factors for. Poor or good? 101 00:06:29,760 --> 00:06:33,080 Prognosis are also preserved in the synthetic. 102 00:06:33,080 --> 00:06:35,040 Data so if. You have a patient with some 103 00:06:37,200 --> 00:06:39,200 complex karyotype. He. 104 00:06:39,480 --> 00:06:41,360 Also has a worse prognosis in the. 105 00:06:41,360 --> 00:06:45,240 Synthetic group of patients, and this is super important because 106 00:06:46,120 --> 00:06:51,680 if we want to use him as a real control group, we have to make 107 00:06:51,680 --> 00:06:53,400 sure that. This is also. 108 00:06:53,400 --> 00:06:55,000 Like in the. Real world. 109 00:06:56,560 --> 00:06:59,720 So there's also A use for exploratory analysis. 110 00:07:00,000 --> 00:07:04,840 So synthetically enlarged data sets, what is the interest? 111 00:07:04,880 --> 00:07:09,080 Of this, we know that there are several important subgroups when 112 00:07:09,080 --> 00:07:12,520 we stay in the case of acute myeloid leukemia, certain 113 00:07:12,520 --> 00:07:14,960 molecular subgroups which are recurrent. 114 00:07:14,960 --> 00:07:16,720 But are rare. And. 115 00:07:16,960 --> 00:07:20,200 So for these group. Of patients, it's of course very 116 00:07:20,200 --> 00:07:23,720 important to know. How they fare with a certain. 117 00:07:23,720 --> 00:07:27,480 Therapy and. With a synthetic. 118 00:07:28,600 --> 00:07:31,720 Generation of data you can. Substitute or. 119 00:07:31,720 --> 00:07:34,720 Enlarge. These rare subgroups. 120 00:07:36,560 --> 00:07:40,160 How reliable is this process? We have all heard about 121 00:07:40,280 --> 00:07:47,400 hallucinations in AI. Can this generative AI generate 122 00:07:48,200 --> 00:07:50,360 misleading data? Well. 123 00:07:51,520 --> 00:07:53,880 Of course, you always have to show it. 124 00:07:53,880 --> 00:07:57,600 For a given group, but for the examples I mentioned. 125 00:07:57,840 --> 00:08:02,600 This is very robust, so. There are some small 126 00:08:02,600 --> 00:08:05,920 differences, but in general. This is. 127 00:08:05,920 --> 00:08:11,560 Very robust in terms of how the baseline distribution of. 128 00:08:11,560 --> 00:08:16,200 Characteristics is but also. For the outcome data, however, 129 00:08:16,840 --> 00:08:20,840 especially when the outcome events get lower in the training 130 00:08:21,240 --> 00:08:24,680 group. This leads to a lower. 131 00:08:24,680 --> 00:08:28,040 Robustness of the model. So there is still open question 132 00:08:28,040 --> 00:08:32,880 where we have to work on. And the last case you mentioned 133 00:08:33,000 --> 00:08:36,360 where this kind of synthetic data can be used is in model 134 00:08:36,360 --> 00:08:39,840 training or benchmarking. Can you explain this further? 135 00:08:40,360 --> 00:08:42,720 So the. Idea is that. 136 00:08:42,720 --> 00:08:46,960 We also have to generate. Data for rare subgroups. 137 00:08:47,320 --> 00:08:50,720 And for that we can use synthetic data to enlarge the 138 00:08:50,720 --> 00:08:52,440 group of. Patients, basically. 139 00:08:53,880 --> 00:08:59,440 For in clinical trials this generated patients are called 140 00:08:59,440 --> 00:09:02,360 now synthetic patients. Can you summarize what the 141 00:09:02,360 --> 00:09:06,720 advantages to use this kind of synthetic patients? 142 00:09:06,720 --> 00:09:09,880 We have to point out that. So far synthetic. 143 00:09:10,280 --> 00:09:13,400 Patients have not been used within clinical trials and. 144 00:09:13,760 --> 00:09:16,800 I believe that there's still a lot of work to do, also from a 145 00:09:16,800 --> 00:09:19,360 regulatory. Aspect to. 146 00:09:19,360 --> 00:09:21,280 Show that this is really feasible. 147 00:09:21,520 --> 00:09:25,880 On the other hand, we know that the generation works and the. 148 00:09:25,880 --> 00:09:29,200 Problem is so big that. We have to think of new ways how 149 00:09:29,200 --> 00:09:31,960 we can accelerate clinical testing. 150 00:09:32,400 --> 00:09:37,320 So for this example we used. Only intensively treated acute 151 00:09:37,320 --> 00:09:40,960 myeloid leukemia patients almost with. 152 00:09:41,040 --> 00:09:43,120 Our targeted therapy, so this has to. 153 00:09:43,120 --> 00:09:45,920 Be shown with. Targeted therapies this has. 154 00:09:45,920 --> 00:09:48,160 To be shown with. Real life data. 155 00:09:48,520 --> 00:09:51,280 And with certain intervention and there's a lot of. 156 00:09:51,280 --> 00:09:53,440 Work to do. Before we can really. 157 00:09:53,440 --> 00:09:57,480 Say OK. So we one group of patients 158 00:09:57,480 --> 00:10:02,680 within a clinical trial for in the end leading to an approval 159 00:10:02,680 --> 00:10:07,440 or identification of certain efficacy of a of a given drug. 160 00:10:08,840 --> 00:10:13,080 So really is work in progress, how closely does these cohorts 161 00:10:13,080 --> 00:10:17,800 match real world cohorts? OK. 162 00:10:17,800 --> 00:10:19,560 That's a difficult question because. 163 00:10:19,560 --> 00:10:24,080 It hasn't been shown yet. So for the moment you don't know 164 00:10:24,080 --> 00:10:26,200 really how well and it's being tested. 165 00:10:26,400 --> 00:10:28,920 We know how well it matches. For. 166 00:10:29,160 --> 00:10:31,880 A given group of patients we studied. 167 00:10:32,040 --> 00:10:34,400 Or other groups. Studied and there it is working 168 00:10:34,400 --> 00:10:36,360 well. However, we have to. 169 00:10:36,720 --> 00:10:39,080 Yeah. Prove this for real world 170 00:10:39,080 --> 00:10:42,600 cohorts. And and certain interventions. 171 00:10:43,480 --> 00:10:47,960 So the idea would be to use these cohorts mostly in the 172 00:10:47,960 --> 00:10:51,960 control group in a clinical trial at the beginning at least. 173 00:10:52,040 --> 00:10:54,000 Yes, I, I. I think this is the most 174 00:10:54,000 --> 00:10:57,680 realistic and and closest scenario that we use this as a 175 00:10:57,680 --> 00:11:01,240 control group, an enlarged control group. 176 00:11:01,800 --> 00:11:03,640 Yes. With a certain intervention. 177 00:11:04,840 --> 00:11:07,720 Which would be interesting because then you don't have to 178 00:11:07,720 --> 00:11:10,400 assign patients to a control group. 179 00:11:10,640 --> 00:11:13,520 So you mentioned the regulatory considerations. 180 00:11:13,720 --> 00:11:17,280 We have some tools at the moment, like GDPR or HIPAA. 181 00:11:18,920 --> 00:11:22,840 Are they enough to address the potential issues that this kind 182 00:11:22,840 --> 00:11:26,880 of synthetic patients would bring? 183 00:11:30,440 --> 00:11:32,320 Well, I think this. Requires. 184 00:11:33,240 --> 00:11:37,400 Complete new thinking about the approval process and the. 185 00:11:37,440 --> 00:11:40,080 The way we look at. Clinical trials. 186 00:11:40,640 --> 00:11:43,080 We need a. Way to prove that this is. 187 00:11:43,080 --> 00:11:46,040 Really working and and I envision this in the 1st. 188 00:11:46,040 --> 00:11:48,680 Place to go. Alongside clinical. 189 00:11:48,680 --> 00:11:51,160 Trials and this is also. What we are currently doing, so 190 00:11:51,160 --> 00:11:54,480 we're going back to clinical. Trials we already. 191 00:11:54,480 --> 00:12:01,400 Did and we are going to see how many patients we can substitute 192 00:12:01,560 --> 00:12:06,360 by synthetic patients. To get the results, we. 193 00:12:06,360 --> 00:12:10,560 Already know, so for me this is an important step. 194 00:12:10,880 --> 00:12:15,520 To show how. We can incorporate A synthetic 195 00:12:15,520 --> 00:12:20,720 patients in real clinical trials and I think this step has to be 196 00:12:20,960 --> 00:12:24,840 done in close contact. With the authorities. 197 00:12:24,840 --> 00:12:28,680 That we together define criteria. 198 00:12:28,680 --> 00:12:33,560 Where we say OK this is. Working and we can move on 199 00:12:33,560 --> 00:12:40,040 further. So if you had a crystal ball and 200 00:12:40,040 --> 00:12:44,320 could see what the future holds in five years time, do you think 201 00:12:44,320 --> 00:12:46,800 it's going to be used a lot? Yeah, what 5? 202 00:12:46,800 --> 00:12:51,520 Years is close in in clinical testing, but I I know. 203 00:12:51,840 --> 00:12:56,200 That the way we. Did clinical trials in rare 204 00:12:56,200 --> 00:12:58,240 disease like acute myeloid leukemia in the. 205 00:12:58,240 --> 00:13:01,720 Past is over. So there won't be clinical. 206 00:13:01,720 --> 00:13:07,880 Trials with like. 1000 patient randomized with a certain drug 207 00:13:07,880 --> 00:13:12,520 because in the meanwhile other drugs are coming up we we gain 208 00:13:12,600 --> 00:13:16,520 further knowledge. So this is getting harder and 209 00:13:16,520 --> 00:13:18,720 harder to perform these large trials. 210 00:13:18,720 --> 00:13:21,840 So we have to. Adopt new methods and I think 211 00:13:21,840 --> 00:13:24,880 synthetic data is 1 very promising way. 212 00:13:25,120 --> 00:13:28,720 But there are several steps. Before we can really say, OK, 213 00:13:28,880 --> 00:13:32,480 that this is working and enhancing our clinical testing. 214 00:13:32,960 --> 00:13:39,080 Are there other uses of this data that you are now developing 215 00:13:39,440 --> 00:13:47,040 for prognostic capabilities in hematological malignancies? 216 00:13:47,880 --> 00:13:53,840 So coming back to images, there's another interesting 217 00:13:53,840 --> 00:13:57,800 field where we can enhance the amount of images. 218 00:13:57,800 --> 00:14:02,600 For rare subgroups. Of patients, so basically doing 219 00:14:02,600 --> 00:14:05,800 the same thing for small subgroups within clinical. 220 00:14:05,800 --> 00:14:07,760 Trials. We can also do this for. 221 00:14:07,760 --> 00:14:10,600 Images of of bone marrow smears, for example. 222 00:14:11,360 --> 00:14:15,440 This is already, it's been used. Well, I wouldn't say it's been 223 00:14:15,440 --> 00:14:19,320 used, but it's like we and others have shown that this is 224 00:14:19,320 --> 00:14:25,240 working and that you can generate synthetic images based 225 00:14:25,240 --> 00:14:29,200 on real bone marrow smears and that you can't really tell which 226 00:14:29,200 --> 00:14:33,040 is a synthetic or. A real bone marrow smear 227 00:14:33,040 --> 00:14:36,080 picture. And these have been used to 228 00:14:36,560 --> 00:14:40,000 train. Well, it hasn't been used, but 229 00:14:40,000 --> 00:14:40,840 that's. The idea. 230 00:14:40,840 --> 00:14:43,280 So the idea. Is that you can generate more 231 00:14:43,280 --> 00:14:47,280 data and then increase model development. 232 00:14:49,520 --> 00:14:53,880 So in the future you think that you will go beyond the control 233 00:14:53,880 --> 00:14:58,040 group, so you would use the synthetic data also in the test 234 00:14:58,040 --> 00:15:00,720 group. The improvement of the 235 00:15:01,200 --> 00:15:06,520 techniques is so fast that we can really, like simulate what 236 00:15:06,520 --> 00:15:09,600 is going on within a single cancer. 237 00:15:09,600 --> 00:15:13,440 Cell with a. Certain intervention and with 238 00:15:13,440 --> 00:15:18,200 that I believe that in I don't know how many years, but in the 239 00:15:18,200 --> 00:15:19,880 future. We will be able to. 240 00:15:19,880 --> 00:15:24,480 Also simulate what will happen to a certain group of patients 241 00:15:24,480 --> 00:15:27,440 when we. Give the drug X. 242 00:15:28,480 --> 00:15:34,080 Yes, I'm I'm a biologist by training and I find difficult to 243 00:15:34,080 --> 00:15:40,520 imagine that you will be able to reproduce this complex system in 244 00:15:40,520 --> 00:15:43,520 its totality, but I look forward to it. 245 00:15:43,600 --> 00:15:46,240 It would be fantastic. Yeah. 246 00:15:46,400 --> 00:15:51,000 Agree on that but like. Five years back, you probably 247 00:15:51,000 --> 00:15:52,960 would. Have said that in predicting 248 00:15:52,960 --> 00:15:56,720 protein structure which is now working almost. 249 00:15:56,720 --> 00:16:00,280 Perfectly. So yeah. 250 00:16:01,320 --> 00:16:04,080 I think it's really promising and I really hope this works 251 00:16:04,080 --> 00:16:10,920 because it would help enormously in accelerating clinical trials 252 00:16:10,920 --> 00:16:15,640 and treating the other patients. So thank you very much, Doctor 253 00:16:15,640 --> 00:16:18,800 Mideke, for sharing your insights and your experience 254 00:16:18,800 --> 00:16:22,160 with us. Thank you to everybody for 255 00:16:22,160 --> 00:16:24,640 listening. And if you like this episode, 256 00:16:25,040 --> 00:16:27,600 don't forget to like it and share it with your colleagues. 257 00:16:28,040 --> 00:16:30,560 And stay tuned for our next episode. 258 00:16:31,040 --> 00:16:32,320 Thank you. Thank you.