Latest / EHA Unplugged / The Role of Artificial Intelligence in Hematology (featuring Dr Gregor Hormann)

Transcript

The Role of Artificial Intelligence in Hematology (featuring Dr Gregor Hormann) · EHA Unplugged

• 0:07
  Greetings and welcome to EHA Unplugged, the official podcast
• 0:11
  channel of the European Hematology Association EHA.
• 0:15
  Hi everyone, My name is Wilson Lim, a medical writer for EHA.
• 0:19
  I'm sitting here today with Professor Gregor Horman to talk
• 0:22
  about AI, artificial intelligence in hematology.
• 0:27
  Professor Orman, would you introduce yourself please?
• 0:31
  Yeah. Hi.
• 0:32
  My name is Gregor Herman. I'm a clinical pathologist or
• 0:35
  specialist in laboratory medicine working at the MLL of
• 0:38
  the Munich Leukemia Laboratory in Germany.
• 0:41
  The M&L is one of the largest laboratories focusing on
• 0:44
  patients with leukemia and lymphoma and diagnosis of bone
• 0:47
  marrow and blood samples. So when we talk about artificial
• 0:51
  intelligence, AI, the first thing that comes to mind is all
• 0:54
  the large language models such as ChatGPT and and Gemini.
• 1:00
  Many, many people are unaware of the other models that can be
• 1:04
  applied that are applied in healthcare.
• 1:07
  Can you share with us some examples of that?
• 1:10
  It's a huge spectrum of AI to be applied in healthcare, for
• 1:14
  example, where we are mainly focusing on is of course using
• 1:19
  AI for our diagnostic processes. And there we are mainly applying
• 1:25
  a particular kind of AI machine learning or more specifically
• 1:29
  deep learning, where we try to get the information out of our
• 1:34
  data and use them in a scientific way.
• 1:37
  With those AI machine approaches, for example, we are
• 1:45
  doing soda morphology or we are doing chromosome on Banach
• 1:50
  analysis. Both of those techniques focus
• 1:53
  on imaging. So our technicians, our soda
• 1:58
  geneticists, our haematologists have to assess those images and
• 2:02
  tell what kind of cells are these, what type of chromosomes
• 2:06
  are these? And we can use here deep neural
• 2:09
  networks, a kind of machine learning approach to align those
• 2:15
  pictures to the classes of cells or the classes of chromosomes.
• 2:21
  So we extract the features out of the picture and then have the
• 2:26
  network aligning the cell to a particular class.
• 2:30
  For example telling this is a blast and this is very helpful.
• 2:35
  As to the cells are pre annotated, of course you need
• 2:38
  all the data digital. First of all, we have to
• 2:43
  photograph all the cells from the peripheral blood smear or
• 2:48
  from the bone marrow smear or have the metaphases digital
• 2:52
  available and then we can apply those AI models to help us in
• 2:57
  the classification of the particular cells.
• 3:00
  This is nothing that the EI is doing alone.
• 3:04
  This is definitely a Co working between the EI and our trained
• 3:09
  personal. So each and every cell is then
• 3:12
  reviewed by a trained medical personnel and then we have a
• 3:17
  final report to be assigned. Still, this addition of AI was
• 3:23
  very helpful in our processes, in particular inside genetics.
• 3:26
  This has really streamlined our reporting.
• 3:29
  We are much faster now. We can deliver the report faster
• 3:33
  to the patient and therefore help the patient care directly.
• 3:37
  So does AI help? Yes.
• 3:40
  So we can use health data to predict and help us with certain
• 3:48
  tasks. So what I've just elaborated on
• 3:52
  was a rather easy task. Just tell me what is this cell?
• 3:56
  What is this chromosome? So we have a picture as an input
• 4:00
  and the classification as an output, rather straightforward.
• 4:04
  Of course, we can also ask more complex questions.
• 4:08
  For example, is this picture associated with a certain
• 4:11
  genotype or is this picture associated with a certain
• 4:16
  outcome of the patient? So then we are talking about the
• 4:19
  CS classification prognostication.
• 4:22
  EI can also be used for this. Again, we typically are then
• 4:26
  using machine learning approaches.
• 4:28
  We can use this with supervised or unsupervised machine
• 4:32
  learning. And a particular unsupervised
• 4:34
  machine learning also helps us to identify new classes of the
• 4:38
  CS where we would not have thought about.
• 4:40
  We have so much data now available in the area of
• 4:43
  genomics, where we struggle as humans to interpret all those
• 4:47
  data, to find patterns in this large field of data.
• 4:51
  And here AI is helpful to point to certain classes that are
• 4:56
  biologically similar depending on the genotype, on the other
• 5:00
  features. And then we as the humans of the
• 5:03
  medical doctors have to figure out if those classes are then
• 5:06
  meaningful in terms of treatment, in terms of
• 5:10
  prognostication, and how to deal with this.
• 5:13
  From all accounts, this also leads to a huge improvement in
• 5:16
  efficiency. How long would you say it takes
• 5:19
  to process for example image and karyotyping?
• 5:23
  Would you say 10 seconds more or less?
• 5:26
  Even faster. Now we are in the range of one
• 5:29
  second. So it's really just get it
• 5:31
  started and it's done. So that's not a bottleneck.
• 5:35
  The system is really fast. And to get this into context, a
• 5:41
  well trained technician who is doing this karyotyping leads
• 5:45
  about one minute for one metaphase and you have to do
• 5:49
  multiple of those to assess one case.
• 5:52
  So you typically need half an hour to assess a case.
• 5:56
  And now this is pre prepared. It still takes some time, but
• 6:02
  for reviewing and this is much faster than doing everything
• 6:06
  from the scratch. Do you think we can still
• 6:08
  improve this? There's always room for
• 6:11
  improvement, so in the past we have been focusing on normal
• 6:15
  chromosomes. The mast majority of those are
• 6:19
  normal, and we also can assign a numeric operation, meaning just
• 6:24
  one chromosome in addition of 1 chromosome missing.
• 6:27
  This nicely works, however, when we're talking about structural
• 6:31
  operations, for example, a translocation that leads to a
• 6:34
  fusion of the chromosomes. BCR Able would be such an
• 6:37
  example with the philadelphic chromosome.
• 6:39
  This needs to be specifically trained to also recognise those
• 6:44
  structurally abnormal chromosomes, and that's what
• 6:47
  we're currently doing. We're training the system with
• 6:50
  the most frequent and most relevant structural operations
• 6:54
  to also have them automatically key assigned.
• 6:57
  You mentioned morphology. Are there any morphology AI
• 7:01
  models as advanced as the Kyo Typing 1?
• 7:04
  Well, the current typing was the first one out there and it is
• 7:07
  really something that we have in place in our routine diagnostic
• 7:10
  application for a couple of years now.
• 7:12
  Therefore, we have a lot of experience with this and also
• 7:16
  other labs are already using this approach and we know it's
• 7:21
  really helpful also in in their hands in the field of
• 7:26
  sodomorphology. We are a little bit beyond this.
• 7:30
  We need to imply this in our own lab right now.
• 7:34
  This is perfectly fine. And we're currently in this
• 7:37
  transition phase to see how robust the workflow is to be
• 7:41
  also applicable for example, in other labs.
• 7:43
  So if any of our listeners wants want to get to know more about
• 7:47
  the Kyo typing process, it's AI. For example the benefits or what
• 7:52
  it take to set it up? How and where can they find more
• 7:57
  about it? You can contact us, we can get
• 8:01
  in touch. There is always room for
• 8:05
  collaboration. We have a designated AI team in
• 8:08
  our laboratory and currently there are five people working
• 8:12
  full time on AI applications in this diagnostic field.
• 8:15
  In medicine. We are enlarging in this team
• 8:18
  right now up to 10 people that will work at AI and for example,
• 8:25
  with the Carrot typing model. This has been a cooperation with
• 8:28
  Meta systems, is also available via them directly.
• 8:33
  So shifting gear a bit here, can you tell us a bit about how we
• 8:38
  can use AI to determine patient prognosis?
• 8:41
  Yes. So when we're talking about
• 8:44
  prognosis and hematology, we now rely on rather well simple
• 8:51
  models because we need to handle those.
• 8:54
  For example, if you think about acute mild leukemia, then we
• 8:58
  have the ELN certification in three different risk groups of
• 9:01
  this of utmost importance for treatment of patients with AML.
• 9:07
  The patients are assigned to this risk groups based typically
• 9:12
  on one marker 1 genetic operation found inside genetics
• 9:18
  or molecular genetics and this makes the risk group for this
• 9:22
  patient. This is straightforward
• 9:26
  physician can deal with this type of information perfectly
• 9:29
  fine. However, we are losing kind of
• 9:32
  complexity because in many instances we have combinations
• 9:35
  of multiple mutations and together with other operations
• 9:40
  and to assess this type of complexity here AI can help.
• 9:45
  There are a number of studies out there that just asked are
• 9:49
  there biological subclasses within AML that can be assigned
• 9:53
  to a different biology or to different outcome?
• 9:57
  And this more evidence based specification taking into
• 10:02
  account the complexity of the disease will be an interesting
• 10:05
  way to move the field forward. So the principle behind this is
• 10:09
  it applicable more widely, for example to MDSMPNS etcetera.
• 10:14
  Yes, it's the principle is applicable for all those types
• 10:18
  of diseases. For MD's, we now have the IPSSM
• 10:23
  that has come out in parallel. There have been multiple
• 10:26
  publications on other approaches, including also EI
• 10:31
  for certification of those patients and machine learning is
• 10:34
  an issue there. Moving forward, to what extent
• 10:39
  do you think you'll be able to implement this in your
• 10:42
  laboratory, for example? Yeah, they can be implemented
• 10:48
  and they have already been implemented to some extent.
• 10:50
  The important thing is you have to carefully validate your
• 10:54
  model. You need to know the
• 10:56
  performance, how well is it working, what are problems, what
• 10:59
  are cases where it doesn't work. And this is really careful
• 11:03
  validation process that is of utmost importance before you can
• 11:06
  apply this into a clinical practice.
• 11:11
  When we are talking about what we are currently doing, this are
• 11:16
  rather narrow questions that we are asking, for example, how to
• 11:21
  assign the class of the cell. And this is what we will call a
• 11:26
  uni model AI, just one input and one output format.
• 11:30
  In the future, there are now proposals and models out there
• 11:34
  that can deal with multiple formats multimodal AI.
• 11:37
  This will be a very interesting topic to see how the field is
• 11:40
  moving with this. We are not there as a current
• 11:44
  state to implement this in routine practice, but this will
• 11:47
  be of utmost importance to move forward.
• 11:50
  So what is synthetic data? It's a pretty hot topic at the
• 11:54
  moment. Would you be able to explain
• 11:56
  more about it? Yeah, that's a interesting and
• 12:00
  rather new field within generation of data with AI.
• 12:04
  So large language models are kind of generative AI, and this
• 12:08
  is also something a little bit in this direction.
• 12:11
  So we're not just analysing data, but we're producing data
• 12:14
  with this. What does this mean?
• 12:18
  You have a large data set. Big data is still an issue
• 12:21
  there. You have a large data set of
• 12:23
  real patients for example, and you then train a model to
• 12:28
  produce patients data sets that are like the real ones but not
• 12:34
  perfect copy of the real ones. So we've slide modifications and
• 12:38
  those new data sets are then called synthetic data sets.
• 12:43
  What are pros of this of this approach?
• 12:47
  You have better patient privacy as the real patient data are
• 12:53
  then not needed for sharing any more.
• 12:56
  You can share those slightly modified data, but we can also
• 13:00
  augment the data set, increase the number at least to some
• 13:03
  extent and to make some studies more powerful with this.
• 13:08
  That's an emerging field. We will have to see for what
• 13:11
  applications is really useful in clinical practice and where the
• 13:14
  limitation of this technique are.
• 13:16
  So for all our listeners who was also interested in synthetic
• 13:20
  data, we will have a specific podcast on that with Professor
• 13:24
  Matteo de la Porta. If you're interested, please
• 13:27
  keep an eye out for today episode by subscribing to the
• 13:31
  EHA Unplugged Podcast channel. So taking all the things that
• 13:36
  you've spoken about, along with the language models, being able
• 13:39
  to suggest reports and other communications, this is moving
• 13:43
  towards an AI doctor. What is your take on this?
• 13:48
  Well, we should not overestimate what AI is currently able to do.
• 13:54
  So those models are helpful in many instances.
• 13:58
  They can well help us get rid of the boring stuff to do and then
• 14:05
  focus on the complex things. Right now, AI will not be able
• 14:11
  to solve all the very complicated cases and really get
• 14:14
  up with much new information on a single case.
• 14:17
  This is really something where the human is still in the front.
• 14:22
  However, it is a matter of working with the eye, not
• 14:26
  against the eye, so it should really help us in our process
• 14:30
  and not replace what we are doing in this combination of
• 14:33
  human intelligence together with the eye is definitely the way to
• 14:37
  go for it. So what you're saying is that we
• 14:41
  shouldn't be afraid of AI, but rather we should embrace this
• 14:44
  change and move forward with it? Absolutely agree.
• 14:47
  So don't be afraid, try to learn about AI and here EHA can be
• 14:53
  very helpful with those training activities, trying to learn
• 14:58
  about their pros, the cons, the limitations, how you can use AI,
• 15:03
  how you can use AI in a way that was really helpful for you in
• 15:08
  your daily practice. And yeah, just you this
• 15:12
  opportunity, it's a great area where we're leaving right now
• 15:15
  and we should not only focus on negative sides, but really get
• 15:20
  into the doing with this. There was a wonderful quote at
• 15:23
  the end of your talk here at the EHA 2024 Congress.
• 15:27
  Would you be able to share it again with our listeners here?
• 15:30
  Yeah, the quote is from Eric Topple, one of the most famous
• 15:34
  physicians dealing with AI. And he said AI will not replace
• 15:39
  physicians, but physicians using AI will replace physicians not
• 15:44
  using AI. It's a very interesting quote,
• 15:49
  yes. Thank you very much, Professor
• 15:52
  Holman for this wonderful talk and also thank you for everyone
• 15:56
  listening. If you're interested in more
• 15:59
  podcasts, do stay tuned and check out our EHA Unplugged
• 16:04
  channel. Until then, have a nice one.