Latest / EHA Unplugged / Accelerating Clinical Trials with AI (with Dr Jan Moritz Middeke)

Transcript

Accelerating Clinical Trials with AI (with Dr Jan Moritz Middeke) · EHA Unplugged

• 0:03
  Greetings and welcome to EHA Unplug, the official podcast
• 0:06
  channel of the European Hematology Association EHA.
• 0:10
  Hello. I am Isabella Rivera, a medical
• 0:12
  writer for EHA. Today we're thrilled to have
• 0:16
  Doctor Jan Moritz Mideka, a distinguished hematologist and
• 0:20
  AI specialist. We're going to be talking today
• 0:24
  about AI to accelerate clinical trials.
• 0:27
  Welcome, Doctor Mideke, can you please introduce yourself?
• 0:31
  Hi, yes, thanks for having me. I'm I'm Moritz Malika from the
• 0:35
  University Hospital in Dresden. I'm working there as a
• 0:38
  hematologist and have a research group on AI in hematology.
• 0:43
  So we want to understand how you can accelerate clinical trials
• 0:49
  by generating data with AI. Can you please explain first
• 0:54
  what is synthetic data? Synthetic data is, as the name
• 0:58
  says, data which is made-up, and the idea is that we face a big
• 1:04
  problem with clinical trials in general, but especially in
• 1:10
  hematology where a lot of the diseases are rare diseases.
• 1:14
  For example, acute myeloid leukemia is a rare disease per
• 1:17
  see and we already know that there are several subgroups with
• 1:22
  targeted therapies. So the number of patients we can
• 1:26
  include in clinical trials are even lower.
• 1:30
  And therefore we need to think about new ways how we can
• 1:35
  accelerate how we perform clinical trials in order to get
• 1:41
  sufficient data to get the knowledge about novel therapies.
• 1:47
  So what types of synthetic data are you generating that could be
• 1:52
  used for clinical trials? So the idea is that you have
• 1:56
  synthetic data which is reassembling the original cohort
• 2:00
  in a a most perfect way, but without being able to get back
• 2:06
  to the original data. So you're generating images.
• 2:11
  This can be used used both for images but also for tabular
• 2:15
  data. And what we did in recent study
• 2:18
  and also other groups have shown is that you can reassemble real
• 2:23
  clinical trial data in terms of baseline data, but also in terms
• 2:28
  of outcome results based on cohorts of of real patients.
• 2:35
  And the synthetic data is reassembling all the information
• 2:39
  about the original patient, which is like age, sex, but also
• 2:44
  molecular changes and then the outcome after a given
• 2:49
  intervention. And how do you generate the
• 2:52
  synthetic data? So for this we used guns,
• 2:57
  Generative Adversarial Networks which is a known form of AI
• 3:04
  which has been used for image generation, quite successful
• 3:07
  already, but can also be used for tabular data.
• 3:12
  And in our specific work, we use two types of these, so and flow
• 3:16
  and C Top Gun. So these are types of deep
• 3:21
  learning and it's generative modelling.
• 3:23
  You have described 4 scenarios in which synthetic data
• 3:28
  generation can be useful and I would like to walk through this
• 3:33
  for the audience. So first, AI generated data
• 3:40
  could help to prevent privacy concerns about clinical trial
• 3:48
  date data of the patients. Exactly.
• 3:50
  So the generated synthetic data can be shared without any
• 3:55
  privacy issues. So we checked that there was no
• 3:59
  single patient in the synthetic cohort which was exactly
• 4:03
  matching the original patient. So there's no way that you can
• 4:08
  go back to the original patient, which of course is the case if
• 4:12
  you just have an anonymized data.
• 4:16
  So this makes it really easy. You can just upload these data
• 4:20
  and make them available also for other research group and see
• 4:24
  public. So how does this compare to the
• 4:27
  DI identification or anonymisation of patient data?
• 4:31
  So anonymisation of these data isn't 100% sure.
• 4:37
  And with the synthetic data it's 100% sure.
• 4:40
  So there's no way that you can go back to the original patient
• 4:43
  because see, there's no original data coming from one single
• 4:49
  patient. So it's basically 100% sure that
• 4:52
  you can't go back to the characteristics of one single
• 4:58
  patient. So this is a big step forward
• 5:01
  for this kind of data. Also, you would create new
• 5:04
  patients and this would lead to either cohort augmentation, so
• 5:10
  some synthetic patients among patients that are real or
• 5:15
  completely a substitution and you have a complete new cohort.
• 5:22
  How robust is this process? So for the generation, it's very
• 5:28
  robust. We showed and also other groups,
• 5:30
  Matteo de la Porta did this for MD's showed that it's
• 5:35
  reassembling the whole Crestrix of the patients forgiven
• 5:40
  disease. So it's not only age and sex in
• 5:43
  terms of the distribution, but it also comes down to the
• 5:47
  molecular changes, which is very important because we know that
• 5:51
  there are certain relationships between certain mutations.
• 5:54
  So for example, if you have an NPM 1 mutation, you are more
• 5:57
  likely to also have inflate 3 mutation.
• 6:00
  And these correlations are preserved in the synthetic data
• 6:04
  set. And this holds also true for
• 6:07
  correlation between age and certain cytogenetic
• 6:10
  abnormalities like complex karyotypes with with higher age
• 6:13
  and so on. So biologically important
• 6:17
  correlations are preserved within this data.
• 6:20
  And even more important, this also holds true for the
• 6:24
  correlation to the outcome data. So known factors for poor or
• 6:29
  good prognosis are also preserved in the synthetic data.
• 6:33
  So if you have a patient with some complex karyotype, he also
• 6:39
  has a worse prognosis in the synthetic group of patients.
• 6:43
  And this is super important because if we want to use them
• 6:49
  as a real control group, we have to make sure that this is also
• 6:53
  like in the real world. So there's also A use for
• 6:58
  exploratory analysis. So synthetically enlarged data
• 7:02
  sets. What is the interest of this?
• 7:05
  We know that there are several important subgroups when we stay
• 7:09
  in the case of acute myeloid leukemia, certain molecular
• 7:13
  subgroups which are recurrent but are rare.
• 7:16
  And so for these group of patients, it's of course very
• 7:20
  important to know how they fare with a certain therapy and with
• 7:26
  a synthetic generation of data, you can substitute or enlarge
• 7:32
  these rare subgroups. How reliable is this process?
• 7:38
  We have all heard about hallucinations in AI.
• 7:43
  Can this generative AI generate misleading data?
• 7:50
  Well, of course, you always have to show it for a given group,
• 7:55
  but for the examples I mentioned, this is very robust.
• 7:59
  So there are some small differences, but in general this
• 8:05
  is very robust in terms of how the baseline distribution of
• 8:11
  characteristics is, but also for the outcome data.
• 8:15
  However, especially when the outcome events get lower in the
• 8:20
  training group, this leads to a lower robustness of the model.
• 8:26
  So there is still open questions where we have to work on.
• 8:31
  And the last case you mentioned where this kind of synthetic
• 8:34
  data can be used is in model training or benchmarking.
• 8:38
  Can you explain this further? So the idea is that we also have
• 8:43
  to generate data for rare subgroups and for that we can
• 8:48
  use synthetic data to enlarge the group of patients basically.
• 8:53
  For in clinical trials this generated patients are called
• 8:59
  now synthetic patients. Can you summarize what the
• 9:02
  advantages to use this kind of synthetic patients?
• 9:06
  We have to point out that so far synthetic patients have not been
• 9:11
  used within clinical trials. And I believe that there's still
• 9:15
  a lot of work to do also from a regulatory aspect to show that
• 9:20
  this is really feasible. On the other hand, we know that
• 9:23
  the generation works and the problem is so big that we have
• 9:27
  to think of new ways how we can accelerate clinical testing.
• 9:32
  So for this example, we used only intensively treated acute
• 9:37
  myeloid leukemia patients almost with our targeted therapy.
• 9:42
  So this has to be shown with targeted therapies.
• 9:45
  This has to be shown with real life data and with certain
• 9:49
  intervention. And there's a lot of work to do
• 9:52
  before we can really say, OK, so we one group of patients within
• 9:57
  a clinical trial for in the end leading to an approval or
• 10:03
  identification of certain efficacy of a of a given drug.
• 10:08
  So really is work in progress, how closely does these cohorts
• 10:13
  match real world cohorts? OK, that's a difficult question
• 10:19
  because it hasn't been shown yet.
• 10:22
  So for the moment you don't know really how well and it's being
• 10:25
  tested. We know how well it matches for
• 10:29
  a given group of patients we studied or other groups studied
• 10:33
  and there it is working well. However, we have to, yeah, prove
• 10:37
  this for real world cohorts and and certain interventions.
• 10:43
  So the idea would be to use these cohorts mostly in the
• 10:47
  control group in a clinical trial at the beginning at least.
• 10:52
  Yes, I, I, I think this is the most realistic and and closest
• 10:56
  scenario that we use this as a control group, an enlarged
• 11:00
  control group, yes, with a certain intervention.
• 11:04
  Which would be interesting because then you don't have to
• 11:07
  assign patients to a control group.
• 11:10
  So you mentioned the regulatory considerations.
• 11:13
  We have some tools at the moment, like GDPR or HIPAA.
• 11:18
  Are they enough to address the potential issues that this kind
• 11:22
  of synthetic patients would bring?
• 11:30
  Well, I think this requires complete new thinking about the
• 11:35
  approval process and the, the way we look at clinical trials.
• 11:40
  We need a way to prove that this is really working.
• 11:43
  And and I envision this in the 1st place to go alongside
• 11:48
  clinical trials. And this is also what we are
• 11:50
  currently doing. So we're going back to clinical
• 11:53
  trials, we already did and we are going to see how many
• 11:59
  patients we can substitute by synthetic patients to get the
• 12:05
  results we already know. So for me this is an important
• 12:10
  step to show how we can incorporate A synthetic patients
• 12:16
  in real clinical trials. And I think this step has to be
• 12:20
  done in close contact with the authorities that we together
• 12:27
  define criteria where we say, OK, this is working and we can
• 12:32
  move on further. So if you had a crystal ball and
• 12:40
  could see what the future holds in five years time, do you think
• 12:44
  it's going to be used a lot? Yeah, what five years is close
• 12:47
  in in clinical testing, but I, I know that the way we did
• 12:54
  clinical trials in rare disease like acute myeloid leukemia in
• 12:58
  the past is over. So there won't be clinical
• 13:01
  trials with like 1000 patient randomized with a certain drug
• 13:07
  because in the meanwhile other drugs are coming up we we gain
• 13:12
  further knowledge. So this is getting harder and
• 13:16
  harder to perform these large trials.
• 13:18
  So we have to adopt new methods. And I think synthetic data is 1
• 13:23
  very promising way. But there are several steps
• 13:27
  before we can really say, OK, that this is working and
• 13:30
  enhancing our clinical testing. Are there other uses of this
• 13:35
  data that you are now developing for prognostic capabilities in
• 13:45
  hematological malignancies? So coming back to images,
• 13:52
  there's another interesting field where we can enhance the
• 13:56
  amount of images for rare subgroups of patients.
• 14:00
  So basically doing the same thing for small subgroups within
• 14:05
  clinical trials. We can also do this for images
• 14:08
  of of bone marrow smears, for example.
• 14:11
  This is already, it's been used. Well, I wouldn't say it's been
• 14:15
  used, but it's like we and others have shown that this is
• 14:19
  working and that you can generate synthetic images based
• 14:25
  on real bone marrow smears and that you can't really tell which
• 14:29
  is a synthetic or a real bone marrow smear picture.
• 14:34
  And these have been used to. Train.
• 14:37
  Well, it hasn't been used, but that's the idea.
• 14:40
  So the idea is that you can generate more data and then
• 14:44
  increase model development. So in the future you think that
• 14:52
  you will go beyond the control group, so you would use the
• 14:54
  synthetic data also in the test group.
• 14:59
  The improvement of the techniques is so fast that we
• 15:03
  can really like simulate what is going on within a single cancer
• 15:09
  cell with a certain intervention.
• 15:12
  And with that, I believe that in, I don't know how many years,
• 15:17
  but in the future we will be able to also simulate what will
• 15:21
  happen to a certain group of patients when we give the drug
• 15:27
  X. Yes, I'm I'm a biologist by
• 15:30
  training and I find difficult to imagine that you will be able to
• 15:37
  reproduce this complex system in its totality, but I look forward
• 15:43
  to it. It would be fantastic.
• 15:45
  Yeah, agree on that. But like five years back, you
• 15:50
  probably would have said that in predicting protein structure,
• 15:54
  which is now working almost perfectly.
• 15:58
  So yeah. I think it's really promising
• 16:02
  and I really hope this works because it would help enormously
• 16:07
  in accelerating clinical trials and treating the other patients.
• 16:13
  So thank you very much, Doctor Mideke, for sharing your
• 16:17
  insights and your experience with us.
• 16:20
  Thank you to everybody for listening.
• 16:22
  And if you like this episode, don't forget to like it and
• 16:26
  share it with your colleagues. And stay tuned for our next
• 16:30
  episode. Thank you.
• 16:31
  Thank you.