Hello and welcome to today’s VJHemOnc podcast. We are a global open access video journal bringing you the latest in hematology and hematological oncology. This month is Blood Cancer Awareness Month and this is the final episode of our four-part special podcast series. Today’s episode focuses on the role of artificial intelligence in personalising the management of blood cancer. First, you will hear from Carsten Niemann, who delves into the role of artificial intelligence and shares a treatment infection model for chronic lymphocytic leukemia that has been deployed in his institution.

Carsten Niemann

Good day, I’m Dr Carsten utoft Niemann. I’m a hematologist and physician working out of Copenhagen University Hospital, in Copenhagen, Denmark. And I’m working half-time in the clinic seeing patients and half-time leading a research group working on providing the right treatment to the right patient at the right time point. So that’s essentially, for me, what data-driven medicine or artificial intelligence in medicine or in health is about, making sure that what we consider medical art is going from just being medical art to being data-driven medical art, meaning that we look into all the data, that we have available for our patients now in an electronic health record and use pattern recognition, or you could call it machine learning, or you could call it artificial intelligence, or just data-driven approaches to all the data available to make sure that we identify the pattern for the patient needing a specific treatment at a specific time point. So that’s what health AI or medical AI is about for me. Then next, obviously, you would ask, how do we use medical AI or pattern recognition today in the clinic?

I guess most of you will, from time to time, use prognostic indices. They’re essentially analyzing quite few variables. It could be like five, maybe a maximum of six or eight variables for a patient. It could be for a patient with chronic lymphocytic leukemia, the CLL-IPI, that consists of a certain score calculated by points for TP53 mutations, IDHV, mutational status, deletion 17P, beta-2-microglobulin, age of the patient, clinical stage. So it’s very simple, and it’s rule-based. Rule-based means that we have done analysis, we have looked into how much should be the weight for each individual factor or variable. And then we calculate a score, and that score informs us about the risk for that specific patient. And we have that across, I guess, all hematological malignancies. Often we try to adjust those. We try to improve on top of these prognostic indices. So that was the rule-based part.

Then we have the data-driven part. We use some of these new mathematical approaches or data science approaches where we can take all the data. And often all the data means very few data still, if we have to enter manually all the data. Because for some of these models, if you want to make it robust, it will take a lot of different variables. We have one model I’ll come back to where we have 80 different variables, but ask your nurse or your data manager to enter the 80 different variables from the medical record, find them, identify them, enter them into the system, and a lot of them will actually be time series where you want all the historical blood results. So try to convince one of your colleagues to do that. But that’s what we need. So that’s why we need automatization and we need integration of these data-driven approaches where we have more complex data. We actually take a look from a very high helicopter perspective on the medical record for the specific patient on all different types of data. It could be genetic data. It could be routine laboratory results. It could even be patient-reported outcomes, what we call PRO. And it could sometimes also be text analysis, or it could be vital information, it could be medication history or diagnostic history. And all of this we feed into one model. And that works. We have seen it from image analysis. We have seen it from large language models. That is a really efficient way to predict what would be the next outcome. And that’s essentially what we want to do. We want to predict what would be the next outcome for this patient. And in a perfect world, we also want to predict if we change something. If we provide a different treatment, a different management for the disease for this specific patient can we actually change the outcome for this patient. So when I see the patient in my clinic what I want from artificial intelligence is to tell me to help me with decision support. To tell me that this patient in front of me has a high risk of infection or this patient in front of me will have a very high certainty not needing treatment within the next year. And that’s usually what we do in terms of medical art when we bring together all our knowledge from our training as physicians our specialization as hematologists from all the research meetings we attend and the research we participate in, all the literature we read we try to figure out what is the most likely outcome for the patient in front of us and how do we best improve it and that’s where we want medical AI to help us with this pattern recognition. Because having thousands and thousands of different variables for each patient, it’s getting harder and harder for me to actually have the overview and identify this pattern.

And looking at that, we have succeeded to implement one medical artificial intelligence device at our clinic for CLL, but we actually have several others that we do not really think about as medical AI from a day-to-day perspective. So remember last time you looked at an ECG for one of your patients. It actually says on the top what is the interpretation from the machine itself, whether it’s a sinus rhythm or whether you have a specific arrhythmia. It even calculates the QTC and the adjusted QTC for you on the ECG. That’s actually artificial intelligence it’s mostly rule based don’t worry about that but it could also to some degree be data driven. And we are starting for some of the image analysis also to get an automatized interpretation. But these are actually quite simple models because they’re looking at one type of data at the time and they’re providing a provisional approach to a diagnosis or to interpretation of the image or the ECG reading. Also a lot of our health systems now provide speech recognition, helping us writing our notes, helping us writing our pathology notes or pathology diagnosis. So we are using it but we are not using it for direct decision support where we try to integrate multiple different kinds of data. That’s where we are going to take the next step because what we’re using it for now is actually something that we as physicians can do. It may just do it faster in terms of writing our notes from speech recognition and providing automatized first step analysis of ECG readings or of x-rays or even pathology reports or demoscopy. But what we need now is to bring the medical AI to something where it can help us with things we cannot do as good as physicians.

And that’s what we tried to do with the CLL treatment infection model. So first we started with what is the clinical unmet need. A colleague of mine, he was a PhD student at the time, he started to look from a strictly epidemiological observational perspective. What is the cause of disease for patients with CLL from diagnosis going onwards? What is the risk of having severe infections before or after starting treatment for CLL, or no events at all. So in a competing risk scenario, looking at time from diagnosis to event, having three possible events, it could be a serious infection, it could be receiving CLL treatment, or dying without having received CLL treatment and without having had a serious infection. And quite to our surprise, he identified that it was more frequent that a patient had a severe infection. It’s defined as having a blood culture drawn because according to Danish guidelines, you would only draw a blood culture if you have a patient with a serious infection. So it’s having a blood culture drawn, receiving CLL treatment or dying. It’s more frequent to have the serious infection. It’s a quarter of the patients diagnosed with CLL who would have a serious infection within five years. It’s a little less than a quarter who would receive CLL treatment as a first event and quite few dying without any of these events. And the one-month mortality after having a serious infection was 10%. That’s really dangerous for patients before receiving treatment for CLL to have a serious infection. And we didn’t know that because most of them had these serious infections outside the hematology clinic because they did not need, at that time point, treatment for the CLL. We did not consider them like high-importance hematological patients, because we thought they were like the indolent CLL patients, not needing treatment for the disease, but they had an immune dysfunction due to CLL, and they were the ones suffering from these severe infections, and 10% of them actually dying within a month of the infection.

That was, it started me to work that focused on data-driven medicine, because we wanted to change that natural history of immune dysfunction in CLL. And we know from several studies, it does not make sense just to start preemptive treatment early on for all patients with CLL. We need to identify the ones with the immune dysfunction because we cannot improve outcome just looking at regular high-risk patients by CLL-IPI and treating targeted agents. It does not improve the outcome. So, we needed to develop the machine learning tool or the health AI, and that ended up with the CLL treatment infection model. We started out with 1800 variables, the model restricted it to just using 80 variables, and we came up with what’s called an ensemble machine learner using 28 different machine learners, that’s essentially mathematical equations, bringing them all together and and doing an assessment of the risk for individual patients, providing the individual risk and information whether the algorithm is confident in the prediction for that specific patient, and including explainability in terms of identifying the risk factors, not the general risk factors for the population of patients. But the specific risk factors for that patient. And we came up just among the first, I think, 200 patients with more than 130 different combinations of just the top three risk factors for the patients. So that’s why it’s a pattern. That’s why it’s not just a rule-based model, because it differs for different patients. But having these mathematical equations, we can actually identify for individual patients whether they’re high or low risk, whether we have a high confidence in the prediction for them, and what is the explainability, what is the reason this specific patient is a high risk patient. So that works.

We did validate it in different patient populations, both internal validation and external validation in a German cohort as well. And we have implemented it in a prospective clinical trial, international trial, where colleagues in Sweden and the Netherlands and Denmark include patients first for the pre-screening. And then if we identify the high-risk patients, we offer them to participate in a randomized phase 2 trial where we randomize between observation, which would be standard of care, of three cycles, three months treatment, targeted agents, combining a BTK inhibitor and a BCL2 inhibitor. That’s a trial called the PreVent-ACaLL trial because we use a acalabrutinib and venetoclax in the trial. We’re still waiting for the results. It’s been the worst trial ever I’ve been running. I’ve trials. We started one month before the COVID pandemic hit Denmark, and it’s dealing with infections. So that was not the best time point to start a trial dealing with infections. And we’ve had a lot of problems also both starting up the trial at new hospitals during the COVID pandemic, but also patients refusing to go on a trial when algorithms tell them that even though they do not feel sick, they have a high risk. So there’s a lot of communication about how to implement machine learning or data-driven research, how to test that, how to explain to your patient that this is actually how we are going to develop clinical treatment, clinical management in the future. So a lot of things to do. But also imagine that we have now had, I think, 400 patients through the pre-screening. That’s 400 hours of entering these data, the 80 different variables and a lot of them as time series for the trial management. And I had a hard time doing that.

So in parallel with the clinical trial, we set out also to automatize and integrate the prediction into our electronic health record system. In half of Denmark, the eastern part of Denmark where I work, we have an EPIC-based electronic health record system. So I think it’s covering maybe a third of all US hospitals as well, and also a lot of hospitals in the rest of the world. But as you may be aware, it’s a commercial electronic health record system, and it’s not open source. But to implement in a transparent way and deploy in a transparent way, medical artificial intelligence device, I believe strongly we need to be able to look through it. So we insisted to make the CLL treatment infection model open source, transparent, available code. And we were now going to integrate that into the EPIC system, the EPIC-based electronic health record system. So we needed to use the transfer from the Chronicle database. That’s what we look at as physicians at the front end of the EPIC-based EHR systems. And then use copy that is made almost in real time now to Caboodle, which is a backup, but where you can add extra data and you can work, you can look at the data for your EHR system. And we had written the CLL treatment infection model in a Python code, and we automatized it, joined it there, did a lot of validation, and we actually published a paper in NPJ digital medicine very recently on the deployment process and we try to make that a cookbook recipe for you on how you can actually yourself use different algorithms and deploy them into your electronic health record system, taking you step by step through the process on what you need to deal with to do this, and to do this deployment in a way where you can also monitor the performance of the model and the input variables and any drift in the input put variables to your model. And we did that in a way where it’s fully automatized every time you have a new patient with CLL or you have a patient with CLL with new data each night, because for economic reasons, we only run it each night, but you could actually do it every minute, every time you get new data. Then it triggers the model to run, and you get the prediction, and it presents us a view that we feed back into the EPIC-based system. So we keep all the raw data and all the prediction information in the Caboodle database where we have full control, but to make sure we can still provide the view in the context of the patient that the treating physician would see in the front end of EPIC, we feed that view back into EPIC. But it’s a dead view and EPIC, you cannot change it. We’ll feed in a new view if the prediction for that patient is updated. So it’s also an example how you can actually integrate a commercial closed software system for your EHR with an open transparent medical artificial intelligence device with an ongoing monitoring setup.

And that’s really where I think we’ll see the medical artificial intelligence field or pattern recognition or data-driven medical art go in the next 10 years. We’ll have tens and tens and maybe hundreds and hundreds of different decision support tools, data-driven decision support tools deployed into our systems, but we’ll move towards what we’ve seen in statistics, in epidemiology. When I trained as a medical student, as a young physician, most of what we used would be GraphPad Prism, or Stata, or SAS, other commercial statistical programs. But now, everyone would write their scripts in R, and we will share among us packages in R, which provide much better and faster and more nicely looking graphs and predictions and just regular statistics. And by doing open, transparent deployment of AI, medical AI to what we’ll achieve, and we can do that, link it to the commercial systems, the commercial electronic health record systems. So if we look ahead, we have a lot of the childhood disease problems now. We really struggling how to do it. It has taken us two years to deploy it into our electronic health record system. But now we are getting to know this process and we can speed up the process, but we need ways to monitor the different algorithms. And it’s not efficient to do that by regular clinical trials, because essentially what we provide is decision support tool helping physicians to do data driven medical art. And to do that, we need what is called stepped wedge cluster randomized deployment or implementation. So, what is this strange wording? That’s that rather than randomizing the patients for clinical trial or for different devices, when you do the deployment in different systems, you will select the different sites to start the deployment in a random order. In play, we can control and adjust for the time factor when we monitor how we improve or do not improve treatment for patients by deploying a new medical artificial intelligence device. And as it’s only offering decision support to the physicians, it’s not outright deciding the treatment for the patient. It’s still safe. We still keep the human in the loop. But it’s a way where we can monitor the deployment of these lower risk devices, but still making sure we get the information, we see how it actually impacts the treatment of our patients.

So, yes, we are going to see data-driven hematology. Yes, we’re going to see data-driven medical art with a lot of different decision support tools. We need to deploy them, implement them in a transparent way using open source code, including a lot of ways of monitoring how it impacts the outcome for our patients. And we need to design smart ways to randomize when the site would deploy it or implement the new tool to make sure we can get trustworthy assessment of the outcome for patients. And that may continue the improvement of care for our patients, making sure that the right patient gets the right treatment at the right time point.

Next, Gareth Morgan will share his insights on the role of AI in blood cancer management. He delves into the role of AI in the classification of myeloma and the prediction of patients who will need transplantation. He also discusses the use of AI to better understand genomic data.

Gareth Morgan

Hi, my name is Gareth Morgan. I’m a professor of hematology at New York University in New York City. We’ve been using AI or trying to explore the potential of AI to improve the clinical management of patients. And I think it’s going to be extraordinarily powerful and allow us to see patterns that you really can’t see with the naked eye and I think those patterns are going to be used to both predict patients response and hopefully if we get enough information to predict, the correct treatment for individual patients so I’m very excited about its use in the clinical arena.

And one of the exciting things about it is that it can take advantage of all of the data that you have available and use that, you know, entirety of the knowledge of a patient and use that to predict how you should treat them. And so, you know, an example of this, of what we’ve been doing is to take the standard clinical staging systems in myeloma, which is the ISS. And there’s a revised ISS, which just manually puts the kind of additional prognostic factors from cytogenetics and mutations, and you can improve that revised ISS. But still, it’s an imperfect tool. So what we tried to do was take all of that information, put it all into the largest amount of data that we could collect. And so fortunately, we’ve been generating ongoing studies from the UK over many years where we had lots of molecular data and clinical data and long-term follow-up. Then integrating that with some data sets generated here in the US. And then with like over 2,000 patients in the database, then use AI to look at that data and make it address patients. Good quality clinical questions. So could you generate a classifier that was better than other classifiers? And the answer was yes. But I think the really exciting thing about the data was that you could almost predict patients who needed autologous stem cell transplant from those who don’t need autologous stem cell transplant. So it’s clear that within the entirety of the myeloma population, some of the good risk patients with standard treatment or good quality, non-aggressive treatment that we have available now, they can live for years and years and years. And why expose them to the toxicity of autologous stem cell transplantation? So it turns out that the AI was able to recognize those populations and arguably it would save a lot of people from the toxicity of autologous stem cell transplantation and still allow them to have good quality long-term survival. And I think that whole process is now being augmented by the development of bispecifics and CAR-T therapy, which is also impacting on the general usage of autologous stem cell transplantation.

So I think the future we’ll see is using AI to help make specific decisions for individual patients. So I’m very excited about integrating AI into the clinical environment. The other area where AI is useful is, again, seeing patterns that humans can’t distinguish. So if we want to develop new prognostic factors, I think we have to move away simply from looking at clinical data and point mutations and the structure of the DNA and its tertiary organization then becomes extraordinarily important because it gives you a new level of data that we haven’t been able to look at before. And so the plan in that setting is to take single cell ATAC-seq, whole genome sequencing. Put those two together and then use the AI to predict the structure of the DNA, which you couldn’t do unless you did this technique called Hi-C. But Hi-C is sort of very difficult to apply to these very small clinical samples. So with really rather affordable and accessible data, we should be able to predict what’s going on, again, with individual patients, how their DNA is folded, and then put that into the clinical AI systems as we get more information.

So what I see is a future full of AI where we learn how to exploit the pattern recognition that it can do and then use those patterns to improve clinical outcome and clinical interventions in a predictable and usable way in the clinic because it’s fine having these technologies but if you can’t get them to patients, it’s a bit disappointing for patients. And what we really need to do is improve patient outcomes beyond where we’ve got to already.

Thank you so much for listening to today’s podcast. We hope you enjoyed. You can check out the other episodes of our Blood Cancer Awareness Month series on VJHemOnc.com. Be sure to follow us on Twitter @VJHemOnc and subscribe to our VJHemOnc podcasts on Spotify, Apple and Podbean. Until next time!