1 00:00:07,240 --> 00:00:11,280 Greetings and welcome to EHA Unplugged, the official podcast 2 00:00:11,280 --> 00:00:15,120 channel of the European Hematology Association EHA. 3 00:00:15,560 --> 00:00:19,360 Hi everyone, My name is Wilson Lim, a medical writer for EHA. 4 00:00:19,640 --> 00:00:22,760 I'm sitting here today with Professor Gregor Horman to talk 5 00:00:22,760 --> 00:00:26,320 about AI, artificial intelligence in hematology. 6 00:00:27,200 --> 00:00:29,920 Professor Orman, would you introduce yourself please? 7 00:00:31,400 --> 00:00:32,040 Yeah. Hi. 8 00:00:32,080 --> 00:00:35,280 My name is Gregor Herman. I'm a clinical pathologist or 9 00:00:35,280 --> 00:00:38,440 specialist in laboratory medicine working at the MLL of 10 00:00:38,440 --> 00:00:40,480 the Munich Leukemia Laboratory in Germany. 11 00:00:41,440 --> 00:00:44,400 The M&L is one of the largest laboratories focusing on 12 00:00:44,400 --> 00:00:47,880 patients with leukemia and lymphoma and diagnosis of bone 13 00:00:47,880 --> 00:00:51,160 marrow and blood samples. So when we talk about artificial 14 00:00:51,160 --> 00:00:54,720 intelligence, AI, the first thing that comes to mind is all 15 00:00:54,720 --> 00:00:58,920 the large language models such as ChatGPT and and Gemini. 16 00:01:00,240 --> 00:01:04,480 Many, many people are unaware of the other models that can be 17 00:01:04,480 --> 00:01:06,760 applied that are applied in healthcare. 18 00:01:07,160 --> 00:01:09,360 Can you share with us some examples of that? 19 00:01:10,000 --> 00:01:14,960 It's a huge spectrum of AI to be applied in healthcare, for 20 00:01:14,960 --> 00:01:19,000 example, where we are mainly focusing on is of course using 21 00:01:19,080 --> 00:01:24,960 AI for our diagnostic processes. And there we are mainly applying 22 00:01:25,000 --> 00:01:29,400 a particular kind of AI machine learning or more specifically 23 00:01:29,400 --> 00:01:34,240 deep learning, where we try to get the information out of our 24 00:01:34,240 --> 00:01:37,360 data and use them in a scientific way. 25 00:01:37,640 --> 00:01:45,160 With those AI machine approaches, for example, we are 26 00:01:45,160 --> 00:01:50,760 doing soda morphology or we are doing chromosome on Banach 27 00:01:50,760 --> 00:01:53,600 analysis. Both of those techniques focus 28 00:01:53,600 --> 00:01:58,000 on imaging. So our technicians, our soda 29 00:01:58,000 --> 00:02:02,520 geneticists, our haematologists have to assess those images and 30 00:02:02,520 --> 00:02:06,280 tell what kind of cells are these, what type of chromosomes 31 00:02:06,280 --> 00:02:09,880 are these? And we can use here deep neural 32 00:02:09,880 --> 00:02:15,720 networks, a kind of machine learning approach to align those 33 00:02:15,840 --> 00:02:20,200 pictures to the classes of cells or the classes of chromosomes. 34 00:02:21,120 --> 00:02:26,840 So we extract the features out of the picture and then have the 35 00:02:26,840 --> 00:02:30,720 network aligning the cell to a particular class. 36 00:02:30,720 --> 00:02:35,160 For example telling this is a blast and this is very helpful. 37 00:02:35,160 --> 00:02:38,680 As to the cells are pre annotated, of course you need 38 00:02:38,680 --> 00:02:42,640 all the data digital. First of all, we have to 39 00:02:43,880 --> 00:02:48,520 photograph all the cells from the peripheral blood smear or 40 00:02:48,520 --> 00:02:52,640 from the bone marrow smear or have the metaphases digital 41 00:02:52,880 --> 00:02:57,320 available and then we can apply those AI models to help us in 42 00:02:57,320 --> 00:02:59,400 the classification of the particular cells. 43 00:03:00,120 --> 00:03:03,760 This is nothing that the EI is doing alone. 44 00:03:04,320 --> 00:03:09,680 This is definitely a Co working between the EI and our trained 45 00:03:09,680 --> 00:03:12,640 personal. So each and every cell is then 46 00:03:12,880 --> 00:03:17,920 reviewed by a trained medical personnel and then we have a 47 00:03:17,920 --> 00:03:23,080 final report to be assigned. Still, this addition of AI was 48 00:03:23,080 --> 00:03:26,680 very helpful in our processes, in particular inside genetics. 49 00:03:26,680 --> 00:03:29,480 This has really streamlined our reporting. 50 00:03:29,560 --> 00:03:33,200 We are much faster now. We can deliver the report faster 51 00:03:33,200 --> 00:03:37,040 to the patient and therefore help the patient care directly. 52 00:03:37,600 --> 00:03:40,640 So does AI help? Yes. 53 00:03:40,840 --> 00:03:48,320 So we can use health data to predict and help us with certain 54 00:03:48,320 --> 00:03:52,160 tasks. So what I've just elaborated on 55 00:03:52,160 --> 00:03:56,440 was a rather easy task. Just tell me what is this cell? 56 00:03:56,440 --> 00:04:00,600 What is this chromosome? So we have a picture as an input 57 00:04:00,880 --> 00:04:03,840 and the classification as an output, rather straightforward. 58 00:04:04,440 --> 00:04:07,400 Of course, we can also ask more complex questions. 59 00:04:08,320 --> 00:04:11,000 For example, is this picture associated with a certain 60 00:04:11,000 --> 00:04:15,680 genotype or is this picture associated with a certain 61 00:04:16,120 --> 00:04:19,600 outcome of the patient? So then we are talking about the 62 00:04:19,600 --> 00:04:21,760 CS classification prognostication. 63 00:04:22,840 --> 00:04:26,400 EI can also be used for this. Again, we typically are then 64 00:04:26,400 --> 00:04:28,000 using machine learning approaches. 65 00:04:28,480 --> 00:04:32,440 We can use this with supervised or unsupervised machine 66 00:04:32,440 --> 00:04:34,480 learning. And a particular unsupervised 67 00:04:34,480 --> 00:04:38,040 machine learning also helps us to identify new classes of the 68 00:04:38,040 --> 00:04:40,360 CS where we would not have thought about. 69 00:04:40,840 --> 00:04:43,120 We have so much data now available in the area of 70 00:04:43,120 --> 00:04:47,800 genomics, where we struggle as humans to interpret all those 71 00:04:47,800 --> 00:04:51,080 data, to find patterns in this large field of data. 72 00:04:51,600 --> 00:04:56,640 And here AI is helpful to point to certain classes that are 73 00:04:56,640 --> 00:05:00,200 biologically similar depending on the genotype, on the other 74 00:05:00,200 --> 00:05:03,640 features. And then we as the humans of the 75 00:05:03,640 --> 00:05:06,840 medical doctors have to figure out if those classes are then 76 00:05:06,960 --> 00:05:10,280 meaningful in terms of treatment, in terms of 77 00:05:10,280 --> 00:05:12,240 prognostication, and how to deal with this. 78 00:05:13,240 --> 00:05:16,960 From all accounts, this also leads to a huge improvement in 79 00:05:16,960 --> 00:05:19,360 efficiency. How long would you say it takes 80 00:05:19,360 --> 00:05:23,000 to process for example image and karyotyping? 81 00:05:23,080 --> 00:05:26,160 Would you say 10 seconds more or less? 82 00:05:26,920 --> 00:05:29,360 Even faster. Now we are in the range of one 83 00:05:29,360 --> 00:05:31,680 second. So it's really just get it 84 00:05:31,680 --> 00:05:35,480 started and it's done. So that's not a bottleneck. 85 00:05:35,840 --> 00:05:41,200 The system is really fast. And to get this into context, a 86 00:05:41,200 --> 00:05:45,040 well trained technician who is doing this karyotyping leads 87 00:05:45,040 --> 00:05:49,560 about one minute for one metaphase and you have to do 88 00:05:49,560 --> 00:05:52,160 multiple of those to assess one case. 89 00:05:52,440 --> 00:05:56,240 So you typically need half an hour to assess a case. 90 00:05:56,680 --> 00:06:02,320 And now this is pre prepared. It still takes some time, but 91 00:06:02,320 --> 00:06:06,360 for reviewing and this is much faster than doing everything 92 00:06:06,360 --> 00:06:08,840 from the scratch. Do you think we can still 93 00:06:08,840 --> 00:06:11,520 improve this? There's always room for 94 00:06:11,520 --> 00:06:15,800 improvement, so in the past we have been focusing on normal 95 00:06:15,800 --> 00:06:19,720 chromosomes. The mast majority of those are 96 00:06:19,720 --> 00:06:24,440 normal, and we also can assign a numeric operation, meaning just 97 00:06:24,440 --> 00:06:27,600 one chromosome in addition of 1 chromosome missing. 98 00:06:27,760 --> 00:06:31,280 This nicely works, however, when we're talking about structural 99 00:06:31,280 --> 00:06:34,920 operations, for example, a translocation that leads to a 100 00:06:34,920 --> 00:06:37,320 fusion of the chromosomes. BCR Able would be such an 101 00:06:37,320 --> 00:06:38,920 example with the philadelphic chromosome. 102 00:06:39,680 --> 00:06:44,320 This needs to be specifically trained to also recognise those 103 00:06:44,320 --> 00:06:47,600 structurally abnormal chromosomes, and that's what 104 00:06:47,600 --> 00:06:50,200 we're currently doing. We're training the system with 105 00:06:50,720 --> 00:06:54,520 the most frequent and most relevant structural operations 106 00:06:54,640 --> 00:06:57,120 to also have them automatically key assigned. 107 00:06:57,440 --> 00:07:01,080 You mentioned morphology. Are there any morphology AI 108 00:07:01,080 --> 00:07:03,640 models as advanced as the Kyo Typing 1? 109 00:07:04,400 --> 00:07:07,800 Well, the current typing was the first one out there and it is 110 00:07:07,800 --> 00:07:10,920 really something that we have in place in our routine diagnostic 111 00:07:10,920 --> 00:07:12,640 application for a couple of years now. 112 00:07:12,880 --> 00:07:16,360 Therefore, we have a lot of experience with this and also 113 00:07:16,360 --> 00:07:20,200 other labs are already using this approach and we know it's 114 00:07:21,080 --> 00:07:26,160 really helpful also in in their hands in the field of 115 00:07:26,400 --> 00:07:30,960 sodomorphology. We are a little bit beyond this. 116 00:07:30,960 --> 00:07:34,720 We need to imply this in our own lab right now. 117 00:07:34,720 --> 00:07:37,680 This is perfectly fine. And we're currently in this 118 00:07:37,680 --> 00:07:41,120 transition phase to see how robust the workflow is to be 119 00:07:41,120 --> 00:07:42,880 also applicable for example, in other labs. 120 00:07:43,640 --> 00:07:47,480 So if any of our listeners wants want to get to know more about 121 00:07:47,600 --> 00:07:52,520 the Kyo typing process, it's AI. For example the benefits or what 122 00:07:52,520 --> 00:07:57,400 it take to set it up? How and where can they find more 123 00:07:57,400 --> 00:08:01,520 about it? You can contact us, we can get 124 00:08:01,520 --> 00:08:05,160 in touch. There is always room for 125 00:08:05,160 --> 00:08:08,360 collaboration. We have a designated AI team in 126 00:08:08,360 --> 00:08:12,320 our laboratory and currently there are five people working 127 00:08:12,320 --> 00:08:15,480 full time on AI applications in this diagnostic field. 128 00:08:15,480 --> 00:08:18,640 In medicine. We are enlarging in this team 129 00:08:18,640 --> 00:08:25,720 right now up to 10 people that will work at AI and for example, 130 00:08:25,720 --> 00:08:28,800 with the Carrot typing model. This has been a cooperation with 131 00:08:28,800 --> 00:08:32,000 Meta systems, is also available via them directly. 132 00:08:33,000 --> 00:08:38,159 So shifting gear a bit here, can you tell us a bit about how we 133 00:08:38,159 --> 00:08:41,200 can use AI to determine patient prognosis? 134 00:08:41,799 --> 00:08:44,200 Yes. So when we're talking about 135 00:08:44,200 --> 00:08:51,560 prognosis and hematology, we now rely on rather well simple 136 00:08:51,560 --> 00:08:54,640 models because we need to handle those. 137 00:08:54,840 --> 00:08:58,440 For example, if you think about acute mild leukemia, then we 138 00:08:58,440 --> 00:09:01,640 have the ELN certification in three different risk groups of 139 00:09:01,640 --> 00:09:05,520 this of utmost importance for treatment of patients with AML. 140 00:09:07,680 --> 00:09:12,200 The patients are assigned to this risk groups based typically 141 00:09:12,200 --> 00:09:18,720 on one marker 1 genetic operation found inside genetics 142 00:09:18,720 --> 00:09:22,280 or molecular genetics and this makes the risk group for this 143 00:09:22,280 --> 00:09:25,840 patient. This is straightforward 144 00:09:26,240 --> 00:09:29,400 physician can deal with this type of information perfectly 145 00:09:29,400 --> 00:09:32,360 fine. However, we are losing kind of 146 00:09:32,360 --> 00:09:35,440 complexity because in many instances we have combinations 147 00:09:35,440 --> 00:09:39,480 of multiple mutations and together with other operations 148 00:09:40,240 --> 00:09:44,360 and to assess this type of complexity here AI can help. 149 00:09:45,040 --> 00:09:49,600 There are a number of studies out there that just asked are 150 00:09:49,600 --> 00:09:53,800 there biological subclasses within AML that can be assigned 151 00:09:53,800 --> 00:09:56,680 to a different biology or to different outcome? 152 00:09:57,800 --> 00:10:02,000 And this more evidence based specification taking into 153 00:10:02,000 --> 00:10:05,280 account the complexity of the disease will be an interesting 154 00:10:05,280 --> 00:10:09,000 way to move the field forward. So the principle behind this is 155 00:10:09,000 --> 00:10:13,840 it applicable more widely, for example to MDSMPNS etcetera. 156 00:10:14,880 --> 00:10:18,000 Yes, it's the principle is applicable for all those types 157 00:10:18,000 --> 00:10:23,120 of diseases. For MD's, we now have the IPSSM 158 00:10:23,120 --> 00:10:26,320 that has come out in parallel. There have been multiple 159 00:10:26,440 --> 00:10:31,440 publications on other approaches, including also EI 160 00:10:31,680 --> 00:10:34,920 for certification of those patients and machine learning is 161 00:10:34,920 --> 00:10:39,640 an issue there. Moving forward, to what extent 162 00:10:39,760 --> 00:10:42,480 do you think you'll be able to implement this in your 163 00:10:42,480 --> 00:10:48,480 laboratory, for example? Yeah, they can be implemented 164 00:10:48,480 --> 00:10:50,560 and they have already been implemented to some extent. 165 00:10:50,800 --> 00:10:54,160 The important thing is you have to carefully validate your 166 00:10:54,160 --> 00:10:56,080 model. You need to know the 167 00:10:56,080 --> 00:10:59,560 performance, how well is it working, what are problems, what 168 00:10:59,560 --> 00:11:03,120 are cases where it doesn't work. And this is really careful 169 00:11:03,120 --> 00:11:06,920 validation process that is of utmost importance before you can 170 00:11:06,920 --> 00:11:08,640 apply this into a clinical practice. 171 00:11:11,760 --> 00:11:15,760 When we are talking about what we are currently doing, this are 172 00:11:16,040 --> 00:11:21,680 rather narrow questions that we are asking, for example, how to 173 00:11:21,680 --> 00:11:26,240 assign the class of the cell. And this is what we will call a 174 00:11:26,320 --> 00:11:30,080 uni model AI, just one input and one output format. 175 00:11:30,680 --> 00:11:33,960 In the future, there are now proposals and models out there 176 00:11:34,120 --> 00:11:36,840 that can deal with multiple formats multimodal AI. 177 00:11:37,480 --> 00:11:40,640 This will be a very interesting topic to see how the field is 178 00:11:40,640 --> 00:11:44,200 moving with this. We are not there as a current 179 00:11:44,400 --> 00:11:47,600 state to implement this in routine practice, but this will 180 00:11:47,600 --> 00:11:49,720 be of utmost importance to move forward. 181 00:11:50,440 --> 00:11:54,160 So what is synthetic data? It's a pretty hot topic at the 182 00:11:54,160 --> 00:11:56,440 moment. Would you be able to explain 183 00:11:56,720 --> 00:12:00,640 more about it? Yeah, that's a interesting and 184 00:12:00,640 --> 00:12:04,920 rather new field within generation of data with AI. 185 00:12:04,920 --> 00:12:08,440 So large language models are kind of generative AI, and this 186 00:12:08,440 --> 00:12:10,960 is also something a little bit in this direction. 187 00:12:11,200 --> 00:12:14,760 So we're not just analysing data, but we're producing data 188 00:12:14,760 --> 00:12:17,200 with this. What does this mean? 189 00:12:18,040 --> 00:12:21,760 You have a large data set. Big data is still an issue 190 00:12:21,760 --> 00:12:23,320 there. You have a large data set of 191 00:12:23,320 --> 00:12:28,760 real patients for example, and you then train a model to 192 00:12:28,920 --> 00:12:34,200 produce patients data sets that are like the real ones but not 193 00:12:34,400 --> 00:12:38,400 perfect copy of the real ones. So we've slide modifications and 194 00:12:38,400 --> 00:12:41,840 those new data sets are then called synthetic data sets. 195 00:12:43,200 --> 00:12:46,280 What are pros of this of this approach? 196 00:12:47,040 --> 00:12:53,840 You have better patient privacy as the real patient data are 197 00:12:53,840 --> 00:12:56,000 then not needed for sharing any more. 198 00:12:56,000 --> 00:13:00,120 You can share those slightly modified data, but we can also 199 00:13:00,120 --> 00:13:03,880 augment the data set, increase the number at least to some 200 00:13:03,880 --> 00:13:07,560 extent and to make some studies more powerful with this. 201 00:13:08,080 --> 00:13:11,200 That's an emerging field. We will have to see for what 202 00:13:11,200 --> 00:13:14,080 applications is really useful in clinical practice and where the 203 00:13:14,080 --> 00:13:15,440 limitation of this technique are. 204 00:13:16,280 --> 00:13:20,080 So for all our listeners who was also interested in synthetic 205 00:13:20,080 --> 00:13:24,440 data, we will have a specific podcast on that with Professor 206 00:13:24,440 --> 00:13:27,600 Matteo de la Porta. If you're interested, please 207 00:13:27,600 --> 00:13:31,200 keep an eye out for today episode by subscribing to the 208 00:13:31,200 --> 00:13:35,960 EHA Unplugged Podcast channel. So taking all the things that 209 00:13:36,040 --> 00:13:39,080 you've spoken about, along with the language models, being able 210 00:13:39,080 --> 00:13:43,520 to suggest reports and other communications, this is moving 211 00:13:43,520 --> 00:13:46,720 towards an AI doctor. What is your take on this? 212 00:13:48,160 --> 00:13:53,800 Well, we should not overestimate what AI is currently able to do. 213 00:13:54,560 --> 00:13:57,600 So those models are helpful in many instances. 214 00:13:58,880 --> 00:14:05,800 They can well help us get rid of the boring stuff to do and then 215 00:14:05,800 --> 00:14:11,160 focus on the complex things. Right now, AI will not be able 216 00:14:11,160 --> 00:14:14,960 to solve all the very complicated cases and really get 217 00:14:14,960 --> 00:14:17,440 up with much new information on a single case. 218 00:14:17,880 --> 00:14:21,400 This is really something where the human is still in the front. 219 00:14:22,680 --> 00:14:26,600 However, it is a matter of working with the eye, not 220 00:14:26,600 --> 00:14:30,480 against the eye, so it should really help us in our process 221 00:14:30,680 --> 00:14:33,800 and not replace what we are doing in this combination of 222 00:14:33,800 --> 00:14:37,200 human intelligence together with the eye is definitely the way to 223 00:14:37,200 --> 00:14:41,040 go for it. So what you're saying is that we 224 00:14:41,040 --> 00:14:44,800 shouldn't be afraid of AI, but rather we should embrace this 225 00:14:44,800 --> 00:14:47,600 change and move forward with it? Absolutely agree. 226 00:14:47,600 --> 00:14:53,840 So don't be afraid, try to learn about AI and here EHA can be 227 00:14:53,840 --> 00:14:58,200 very helpful with those training activities, trying to learn 228 00:14:58,200 --> 00:15:03,400 about their pros, the cons, the limitations, how you can use AI, 229 00:15:03,560 --> 00:15:08,600 how you can use AI in a way that was really helpful for you in 230 00:15:08,600 --> 00:15:12,280 your daily practice. And yeah, just you this 231 00:15:12,280 --> 00:15:15,960 opportunity, it's a great area where we're leaving right now 232 00:15:15,960 --> 00:15:20,200 and we should not only focus on negative sides, but really get 233 00:15:20,200 --> 00:15:23,120 into the doing with this. There was a wonderful quote at 234 00:15:23,120 --> 00:15:27,120 the end of your talk here at the EHA 2024 Congress. 235 00:15:27,640 --> 00:15:30,520 Would you be able to share it again with our listeners here? 236 00:15:30,760 --> 00:15:34,280 Yeah, the quote is from Eric Topple, one of the most famous 237 00:15:34,280 --> 00:15:39,960 physicians dealing with AI. And he said AI will not replace 238 00:15:39,960 --> 00:15:44,520 physicians, but physicians using AI will replace physicians not 239 00:15:44,520 --> 00:15:49,360 using AI. It's a very interesting quote, 240 00:15:49,600 --> 00:15:52,640 yes. Thank you very much, Professor 241 00:15:52,640 --> 00:15:56,800 Holman for this wonderful talk and also thank you for everyone 242 00:15:56,800 --> 00:15:59,640 listening. If you're interested in more 243 00:15:59,640 --> 00:16:04,400 podcasts, do stay tuned and check out our EHA Unplugged 244 00:16:04,400 --> 00:16:07,160 channel. Until then, have a nice one.