Amer Zeidan:

Hello everyone. Thank you for joining us. This is a new session of VJHemOnc MDS sessions. My name is Amer Zeidan. I’m an Associate Professor of Medicine at Yale University and it’s my pleasure today to have a dedicated episode talking about artificial intelligence and machine learning algorithms in hematologic malignancies, but in particular, in MDS. And it’s a pleasure to have two experts who are very experienced in this area who are going to share their thoughts about the current status and the future directions in how artificial intelligence will be incorporated in management of especially myeloid malignancies and MDS. We have Dr Anne Sophie Kubasch, who is a hematologist at the University of Leipzig. She specializes in the management of patients with myeloid malignancies with a focus on artificial intelligence, as well as Dr Aziz Nazha, who’s an adjunct staff at the Thomas Jefferson University, and previously at Cleveland Clinic. And Dr Nazha has also done extensive work on use of machine learning, especially in algorithms related to prognostication within MDS. Thank you both for joining me today, and I look forward to an interesting conversation. Maybe we can start by general discussion about what really artificial intelligence, or machine learning mean and are they the same or are they different? So maybe we can start Dr Nazha and Dr Kubasch can jump in as needed.

Aziz Nazha:

Thank you very much Amer, and thanks for this wonderful opportunity to be with all of you today. I think sometimes that those terminology get confused a little bit, so I’ll try to simplify them. What do we mean by artificial intelligence? So artificial intelligence: making a machine think and do things like a human without explicitly programming the machine. If a human can drive a car, can I make the machine drive the car without me telling the machine each single step around this journey?

Now under this big umbrella of artificial intelligence, there is machine learning. Machine learning is teaching algorithms and computers with data. There are two types of machine learning: one is supervised learning, and supervised learning is where we know the answer. Let’s say I have a dataset that has columns in it and I’m trying to build a model to predict response or no response to chemotherapy. In this example, I know the answer, and then I build the algorithm and then hide the answer from the algorithm and say, “Hey, is this a responder or non-responder?” The algorithm looks at the pattern and the data and comes up with the answer. That’s supervised learning. The other type of machine learning is unsupervised learning, and unsupervised learning is where we don’t know the answer. We ask the algorithm to put together a relationship in patterns in the data and come up with the answer to us. An example of that in hematology and in cancer research is clustering, so RNA cluster, clustering the RNA-seq two different groups that becomes… Because we don’t know those groups, but we ask the algorithm to put those groups together. And there are some semi-supervised learning, which is kind of in between.

Aziz Nazha:

Now the third part is deep learning and deep learning is a subset of machine learning, where the algorithm is using a neural network. And these neural networks are mathematical models of the neurons on our brain. So in our brain, you have neurons that are connected to each other, and those neurons fire up when we do things in a different place of the brain. If we try to mathematically model that, that becomes the neural network and the more layers we have, that becomes deep neural network.

Amer Zeidan:

Thank you so much for such a detailed answer. Anne Sophie, this concept I think, is starting to get into medicine, but many physicians, including hematologists, oncologists, are still not sure how this is going to apply in day-to-day care. Big picture, before we go into the small details of how this is going to work. Do you anticipate that these algorithms are actually going to make a major change in how we practice medicine day to day, and maybe in the next decade or so?

Anne Sophie Kubasch:

Thank you for this very important question. So from my standpoint, I think there will be a huge impact for our daily practice because our data are getting bigger and bigger. Every day, we collect more data into our registry datasets, and therefore I think, not only in the field of hematology, in the broad field of medicine, it’s the same case that diagnostics are getting more specific, more molecular data, more cytogenetic data, more clinical data is collected. And therefore, I think the application of AI will be more and more important to use this data, to find more prognostic and more stratification factors for our patients and to apply personalized medicine. And therefore, I think it will be a huge impact to apply AI within the next years.

Amer Zeidan:

Thank you.

Amer Zeidan:

Aziz, before we go into hematologic malignancies and MDS, and AML in detail, are there currently examples of where these technologies are actually in active use currently within any branch of medicine, whether it’s in terms of prognostication, or in terms of therapeutic decisions, or is this still within the investigational realm?

Aziz Nazha:

Excellent question. I think overall the journey for AI in healthcare is a little bit of an early journey yet compared to the other industries. However, we see some applications that start surface and make it to the clinic. The hardest part is, can we demonstrate the effectiveness of these algorithms in the clinic? If you think about radiology, for example, I stopped recently counting, but there are probably around 350 FDA cleared algorithms. And this is important to differentiate. So FDA cleared, meaning that the algorithm could be deployed, but then you always need a human to be part of supervising the algorithm. So these are not FDA approved, they are FDA cleared. Some of these algorithms are already implemented in some hospitals, and some hospitals saw benefit and some hospitals did not, so that’s one area. We’ve seen, for example, last year the approval of the first FDA cleared algorithm to detect prostate cancer on pathology slides. Again, it’s FDA cleared. So we see a lot of those algorithms are going to start coming to the clinic in the next couple of years. The question becomes, how do we implement them? How do we monitor them? How do we [inaudible 00:07:20] physicians how to interact with these algorithms? And do these algorithms at the end of the day provide value? And that’s to be determined in the next few years.

Anne Sophie Kubasch:

Just one point from the perspective of Germany or EU. So, we have the same problem with, of course, a lot of research is just in a field of AI, but there are a few products on the way to get the medical product, but only a few just went into the clinics. We have a mammography system here in Germany which is already licensed as a medical product, but this is a huge problem for most of the systems because it takes such a long time and after approval, of course, it’s not easy to get into the clinic and to find the physicians to work with it. So it’s maybe something we can also discuss how we can make this easier for the physicians to use these new systems.

Amer Zeidan:

Yeah. Are there ongoing efforts like this in terms of…just to make sure I understand, are you saying that the picture can be taken and then artificial algorithm mechanism will read it if it’s abnormal or normal without having a special radiologist review it?

Anne Sophie Kubasch:

Right. It works like this. So the system is evaluating the mammography and at the end, the physician of course gets the result and there’s a possibility to save time because of course the radiologists only have a deeper look into the pictures with the problem or with [inaudible 00:08:57] cancer.

Aziz Nazha:

This is a good example Amer of how these algorithms, the clinical application. So there have been multiple algorithms actually to try to read the mammogram and say this is the [inaudible 00:09:13] degree. And then also if there is any spot on the algorithm that has suspicion. Now in Europe, for example, let’s say you take UK for somebody to read a mammogram, you need two radiologists to say this is malignant, or this is suspicious of malignancy. So this is why it’s important and then I can maybe have the algorithm read one, and then the radiologist read the other one. The challenge in the United States becomes, am I aiding the radiologist or am I just replacing the radiologist in some instance? And today those algorithms cannot read the image on their own, so the radiologists still have to read the image and agree with the algorithm. This is why it becomes an operational issue because if I’m not replacing the radiologist, now how am I making it easier for them to read the image, recognizing there are certain true positive and true negative, like any imaging analysis. So this is where defer in the type of the use case that you have and some of the challenges when you take those algorithms to the clinic.

Amer Zeidan:

I think this is a very important point. I’m trying to make the analogy to these Tesla cars where they are self-driving, but you still need to have a driver sitting behind the wheel. So is that more because there are instances where the algorithm might fail or is that more from a regulatory point of view that the regulatory authorities want to have someone who is responsible at the end of the day because they cannot go after a machine if there is an error. So is this how it’s currently being thought of? Are there any studies about how often do you get inaccurate or differences in the read between the machine and person who supervises it?

Aziz Nazha:

So I would love to hear what Anne would say, but I think it’s multiple levels to this. So it’s multiple issues, meaning how do we evaluate the accuracy of these models? Because certainly they’re not accurate, not all the time, right? And how do you benchmark the accuracy? So the level of, for example, accuracy of a model is a very bad indicator of the model evaluation. Let’s say you have a CT scan of a bleed in the brain and 3% of the images will have a bleed because as you know [inaudible 00:11:56] and then the rest will not. So if you have an algorithm saying, “no, no, no” all the time, you’re 97% accurate, but your algorithm is clinically useless, right? It’s not useful. Then it comes to, “okay, now I got to use AUC”, the area under the curve, which a lot of people use, which is also another not really right matrix to use, because once you have skewed data like this, meaning 3% events, your AUC will be high, but then what we call precision and recall might be lower. I’ll give you an example, the bleed brain. There have been some algorithms where the precision is 30 or 40. What does that mean? If every 10 images that the algorithm say there is a bleed, four correct, six wrong, even though the area under the curve is much higher. That’s clinically problematic because if I’m interacting, I’m six times wrong out of four times, meaning if I flip a coin that becomes even better. So the first question is how do we evaluate these models and do they really help? The other, from the regulatory standpoint and from the patient standpoint and then from insurance standpoint. So today will the insurance pay for this algorithm if they are paying for radiologists and paying for the algorithm, and the answer? No. So if the insurance is not going to pay for it, nobody is going to adopt it. So that’s another layer. And then the important layer, there has been a lot of research now put in there: patients will refuse, actually. So a lot of patients have now concern that the algorithm will read their picture, the algorithm will discriminate against them. So there has been a lot of work toward that. So there are different layers that add to the challenge of just taking those algorithms and putting them in the hospital.

Anne Sophie Kubasch:

I totally agree. And I think as you pointed out, not only the patients, also some physicians at the beginning of such an implementation are a bit feared about the wrong decision and about the ethics. So what happens if they rely on the system and they trust it and they decide for this therapy decision in our case, and at the end, it turns out that it’s not a right decision or maybe another decision would be more matched to this patient case? So something in this field is of course not easy to discuss. And therefore, I think we should of course talk with the ethics departments, talk with our patients, talk with our physicians. And of course it takes time maybe to implement everything into the clinics and to get used to it, I think.

Amer Zeidan:

Those are good points. I do think that actually having the physicians embrace this is going to be a challenge. Traditionally I think all new changes are always difficult to implement. When electronic medical records started coming in until now people are having all kinds of complaints about them and how to do them in a good way. But I think one starts coming to the point of where you are making medical decisions based on the algorithm. There’s probably another layer of hesitancy. As I’m sure both of you know when you had a CAT scan showing progression many of us are not just relying on the report, they actually like to look at the CAT scan themselves. Even when you think about prognostic scores, I think one of the hesitancies is when people can calculate the risk score themselves using a calculator, many people feel more comfortable rather than just putting a million variables or having a machine spit out that this is a prognostic risk of this patient because there’s always this intrinsic concern about, “oh, if something is wrong” or what happens if not everything was being taken into care. So some of that might take just some good evidence and time to be implemented, but what do you think, Aziz, is needed to move these more into become mainstream? You mentioned the regulatory, you mentioned the insurance, but in terms of the data generated about their usefulness and in terms of the patient and physician acceptance. In those two particular areas, what do you think is needed?

Aziz Nazha:

Absolutely. So it’s really going to go back to the use case and what is the problem I’m trying to solve. And then how do we work instead of physician versus machine, it’s a physician plus machine. So the concept of me as a physician and hematologist, I’ll be better hematologist if I have better tools that equip me to take care of the patient. So I think that’s the most important part because at the beginning of this, a lot of companies and people start talking about replacing pathologists, replacing physicians, and that’s not the right context. It’s the question, how do I make the physicians more efficient in their work and provide value? So if we focus on the value, that becomes the number one question. What do I mean by value? There are a lot of models they build just to build a model, but the clinical value and impact is lacking. This is why you see a lot of these algorithms don’t make it to the clinic. And then one of the things that was useful to us. We always operate on “all models are wrong, but some are useful”. So it doesn’t mean that the model has to be correct 100% of the time, the question is the outcome of the model, how’s it going to help me as a physician? And also how do we benchmark it? So what do you benchmark? The model. Give you an example. When we built our personalized survival model, the accuracy or the C index of the model was 0.75. And people say for survival, we’re not talking about 90. We’re not talking about 80. We’re talking about 0.75 and then say, well, it’s low. Well, what is my standard today? And if my standard today is IPSS and IPSS-R, on the same data set that standards 0.64, the C index. That’s what I’m using today. So now I’m getting 10% better. Yes. I’m not getting perfect, but I’m getting 10% better. That’s number one, number two, when we put of the model out there, how do you use it in the clinic? So we make sure how to define higher risk and lower risk. And actually we demonstrated that when we use our model the outcome of a clinical trial could have been a changed. So now, when you start talking about value, when you start showing people value and then educate them how to use the model, eventually they will use the model.

Amer Zeidan:

We’ll be talking about MDS in a little bit, but I want to give Anne the chance, from a German system perspective, I think Germany is probably the closest to the US in terms of insurance patterns, because there’s a mix of a large national coverage, but also a number of private insurance systems. How do you see the intake of new technologies like this within the health system? Are German doctors more open to change? Or the patients? How do you see this?

Anne Sophie Kubasch:

So I think it depends on the kind of doctor, on the kind of hospital on this age. So in our young working group here in Leipzig, we have motivated doctors who want to apply AI, who want to apply new digital health solutions. So I think if the physician is motivated and also the patient, also there, it depends on this age, but we have 40 or 50 year old patients with MDS who just ask me for a digital solution to get the quality of life assessment. So I think it totally depends both on a physician and on a patient side and if both are coming together and are interested in this new solution, I think there will be no problem. We of course, also have the problem with insurance companies not funding most of the systems and therefore, as I already told you, it’s a long process for the application as a medical product and after approval we have a strong chance that the healthcare system will reimburse most of the solutions, but it takes time and only a few of these solutions are already licensed here.

Amer Zeidan:

So I think both of you agree that age is a very important factor for the physicians, I guess that applies to everything technology wise. I think younger people are certainly more open. We can see that also the electronic medical records and how many of the younger generation physicians actually are very comfortable doing the whole notes and everything even from their smartphones, which is quite impressive. So maybe we can start digging more specifically into MDS and if there is also somewhere that you will want to emphasize in AML as well, please feel free to chip in. And I think the major two areas where machine learning and artificial intelligence has been studied in MDS, and please correct me if there is more, is in the area of prognostication and the area of therapeutic decision. I don’t think on the area of diagnosis, which I think actually is a major area of unmet need in MDS because especially in lower risk MDS, when there is no excess blasts, when there is this dysplasia. And Aziz brought up the idea of there have been all these rumors or people saying machines are going to replace pathologists and they are going to read the bone marrow, and I actually did see it in a couple of meetings where algorithms were used to read the bone marrows for lower risk MDS, but maybe we can start there and then start talking about prognosis and then therapeutic decisions. Aziz are any other areas that you think it’s going to be important?

Aziz Nazha:

I think you’re absolutely right. So if you dissect it about diagnosis, prognosis, and then treatment selection slash clinical trials. If we start with diagnosis, as you mentioned Amer, it’s a big challenge, at least in MDS. There have been studies and a lot of studies have shown that about 25 to 30% of patients seen at tertiary centers, the diagnosis get changed. Meaning patients told that they have MDS, but they don’t vice versa. The problem is identifying dysplasia on the slides becomes difficult. So one of the projects we have done, one way is you can look at the pathology slides and say, “okay, this is MDS versus other myeloid malignancy”. The challenge with that that we face when we scan that is you have to label. So you have to teach the computer. So think about it like teaching a pathology resident about the blast. Now you have to label all those and try to teach the algorithm, which is really hard to do. And you could pull up three pathologists and each one of them in terms of blast or dysplasia, they’re going to disagree. So we took it a step farther actually and this was a collaboration between US and Europe, where we said, can we use CBC and genomic data to say this is MDS versus other myeloid malignancies? So we built a dataset from Cleveland Clinic at the time and University of Pavia in Italy and we validated the model externally from data from Munich Leukemia Laboratory. What we showed that we build an algorithm that can take CBC and differential in small sets of genomic data and can tell you, this is MDS versus other myeloid malignancies like MDS/MPN, ICUS, CCUS and others in a 96% accuracy without doing a bone marrow biopsy. So it’s not 100%. Now how would you do use that in the clinic? One way you could say, “okay, I can use the algorithm to screen those patients”. And then with the help of pathologists, you could bring up the confidence of the pathologist because sometimes you get a report on a lower risk MDS, it could be lower risk MDS, it could be… You’re not getting a definition answer. So the question, can I improve that? So I think there is an opportunity still in diagnostics. And then the flip side to that is can I use the image to predict molecular subtypes? So one of the projects we were working on, this is not published, is can I look at the pathologist slide and say, this patient has tp53 mutation and what that will be and what the accuracy will be. So I don’t have to wait for the whole genome sequencing panel.

Amer Zeidan:

About in Germany Anne Sophie, are there efforts to try to use these algorithms for diagnostic purposes within MDS, AML or other myeloid malignancies?

Anne Sophie Kubasch:

So within our group, we are working not on the diagnostic field. We are working on the therapy decision field. So since around one and a half years, we are building an AI-based therapy decision tool where we try to integrate as much data as we can collect in three years’ time. And this therapy decision tool should be for patients with MDS, myeloma, and AML. And hopefully after this three-year funding period, we hope that for every patient, with every molecular feature, with every biotherapy entered into the system, by the physician, this is our approach, hopefully it’ll get a result based on case-based evaluation. And at the end, hopefully we will get a personalized treatment decision trying to find the best fitting therapy for this individual case. This is something we are working on. It’s already not published because we are still working on this field and we are trying to match the data sets, but this is something we will see within the next years, hopefully it’ll work. And it will also get therapy decision more faster, and maybe more accurate. This is our aim.

Amer Zeidan:

This is actually an exciting area. I think one of the things with MDS has been always that we did not have that many treatment options. I remember when we had many papers come looking at how do you select for azacitidine and decitabine sensitivity and at the end of the day, the answer was always, well, these are the only drugs we have for high-risk MDS so why would it matter. So I think this might be something that might change now that we have more drugs that are being tested and hopefully new drugs will be approved, but I suspect some of that will actually apply very well in AML and multiple myeloma since there is a large number of new drugs and we are actually starting to face clinical dilemmas in terms of how do you choose the initial therapy? How do you sequence it? Because there are so many different drugs and not enough trials to answer all the clinically important questions. So I think these algorithms are certainly going to be important. Aziz, you’ve done some prognostication work using the machine learning. So before you go into the details, I just want you to address this bigger picture issue, you touched on it a little bit and is that people are saying that the incremental value might not be that high. And why should we spend all this time and effort and implement these algorithms if the incremental value of doing all of that is going to be minimal? Similar arguments were made about the genetics at least in MDS, the blast count, for example, used to be a very major proxy of prognosis but then we realized that karyotypes are actually more important, and then molecular data started coming in and dilemmas have been how do you integrate all of this? So how do you see the incremental value of putting all of this together and then using machine learning to integrate?

Aziz Nazha:

Absolutely. That’s an excellent question. And then I think if we start with what’s the problem we’re trying to solve and then how we could solve the problem. So what we realize is if we look at our current standard systems that we use to risk stratify patients, we typically, as you know, in MDS we say higher risk and lower risk. And the treatment algorithms for those patients are different, right? So we tend to be more aggressive with high-risk patients. We tend to be more like observation, less aggressive with lower risk patients. And that’s a very important point because what it says, if I get the prognosis right, I get the treatment right and if I get the prognosis wrong, I get the treatment wrong. In other words if I label the patient as a higher risk, but the disease is behaving like a lower risk, then I’m over-treating that patient and vice versa. It turns out if we use our IPSS and IPSS-R system, in multiple publications we have shown that about 20 to 25% of those patients say higher risk. In fact, they are lower risk and then vice versa. So about those 25 to 30%, they’re getting the wrong treatment. So the question becomes, can I be more specific about my prognosis? The other conversation that actually led to building the personalized model is that a lot of patients start realizing that, okay, look, you’re telling me I’m high-risk, but you’re giving me general numbers. What is my number, my specific number? So we start seeing even patients in the high-risk group, they have a lot of heterogeneity, they behave completely different. So that’s raising a question, can we build a model that can give you personalized prediction that is specific for a given patient, and would that improve the current standard that we have? Which led to the project that we pulled together, international dataset to build a machine learning model that uses CBC demographic data and some genomic data, and then is going to give you this survival probability that’s specific for a given patient.

What we demonstrate, couple of things. To go back, okay, now we build the model, how do we evaluate the model clinically? And how do we show value of the model, compare what we have, because there’s not that much value. The whole purpose is not to use machine learning to use machine learning, is can I provide value doing machine learning, not doing traditional statistics? So what we showed, number one, as I said before, that the C index, the accuracy of the model significantly improved compared to what we have today, again, the standard IPSS and IPSS-R. But also what we showed that we took the data into data from prospective clinical trial. That prospective clinical trial was randomizing AZA to AZA versus lenalidomide, AZA versus vorinostat. The trial was negative. When we took that data, it was a small patient cohort, 75 patients who have the clinical and genomic data that qualify to be applied to our model. Interestingly, as you know, this trial was designed for higher risk MDS, and what we found 67% of those patients were actually lower risk by our model, not higher risk. Now, why this was important, because if you take that patient who has higher risk by our model and the IPSS, the combination of AZA lenalidomide almost doubled their survival compared to AZA alone. It wasn’t statistically significant because the numbers are small, right? Where the opposite happened on the lower risk, the lower risk actually get harmed. Their survival with AZA was about 24 months, with the combination was about 16 months. So you can see here, we’re not using the right patients to the right treatment at the right time. And this is why, will that model change how we look at clinical trials today and how do we actually risk stratify them based on the new model? So that was kind of the added benefit compared to just using traditional statistics and traditional modeling.

Amer Zeidan:

So, Anne Sophie, Aziz started touching on these therapeutic decisions and how potentially they could be influenced by these algorithms. Maybe you can tell us more about how are you going about designing this therapeutic decision tool that you are using? I think you mentioned it’s for AML, MDS and multiple myeloma.

Anne Sophie Kubasch:

So we named our tool KAIT. KAIT will be a decision tool for physicians, especially hematologists in outpatient centers, but also in inpatient centers in big university hospitals like here in Germany. And we think that something like KAIT will be useful in a [inaudible] setting, where of course we discuss our patients and there could be an evaluation of KAIT guiding the treatment physician by these AI case-based evaluation. On the other end side, we can think that we can support hematologists in the outpatient center using this KAIT system also to get the right treatment decision for their patients. And how we built the model. Since the last one and a half year, our main focus was on data harmonization. We talked to a lot of patient registries to be tried also to include a lot of clinical trial data, also matching these datasets into our system. And at the end the most important case is that our registry data should contain their response evaluation. This is something a lot of registry data is missing. We know their treatment line. We maybe also know the duration of treatment, but in a lot of registry data, there was no evaluation of treatment response and therefore it was not easy to collect this data, but now we are in a point where we think we are ready also to train our models. And hopefully after one and a half years, then we’ll be there, the final phase of our project, it will be applied to the physicians and will be in a pilot testing phase, applied into the tumor boards and in the outpatient centers. This is our aim and working hard on it.

Amer Zeidan:

So just to make sure I understand how this works, this is more like a living artificial intelligence device? I think of this again, as a human brain, everything you do, you are learning off of your previous experience and adjusting your next behavior. So in this algorithm, you are trying to predict the response and you are collecting the data prospectively after the patient, did they respond or did they not respond to the treatment? And then do you feed it back to the machine so that it keeps adjusting or increasing the applicability of the next prediction?

Anne Sophie Kubasch:

Right. This is completely right. So the system is learning from these case-based evaluations so more patients are entered into the system. It will be more accurate. And therefore, we think after the rollout phase where physicians are working with the KAIT system it will be getting better and better. This is our hope.

Amer Zeidan:

Perfect. So I think in the last few minutes, I want to touch on one point because I think there is a lot of active discussion about this in the field. And maybe I’ll use your tool Aziz, the one that you developed in the Cleveland Clinic with the MDS clinical research consortium with the machine learning for prognosis of MDS. And right now there was another tool, which I don’t think used machine learning, that came a large database from Europe. And then most recently we had molecular IPSS and this was not like a regular statistical modeling, but they used a large number of basically variables, but they did not use machine learning or artificial intelligence. And then you have these two camps, which one should be used? And I suspect the same thing is going to happen with therapeutic algorithms. And so how do you see what is the best way of having people have a consistent way of approaching this rather than people using this algorithm and other people using that algorithm?

Aziz Nazha:

Absolutely. So this is really important because the last thing we want to do is to have multiple algorithms and people applying multiple models, we want one unified model that everybody agrees on. Also, you brought up a very important point. How are these models easily to be used in the clinic? Because one of the European tools, it has 60 variables or more, the IPSS-R molecular has also a lot of variables and a lot of people don’t want to just sit and click and spend 15 minutes putting variables. We were very careful when we designed our model to come up with the least amount of variables that will give me the highest accuracy or the highest impact. To keep in mind that adaptability to a model is as important as the accuracy of the model. And again, if you have 60 variables, nobody can adapt. It’s going to be really hard to adapt to these models. So I think there will be some research to try to compare the model. But part of it not just compare the accuracy because they might be the same level of accuracy, C index or whatever evaluation, also how do I use them in the clinic and how easy it is for me to use them and would the physician use them? Because we all recognize that we need to move from the IPSS and IPSS-R, for multiple deficiencies, we talked about the question, how do we move to the next level?

Amer Zeidan:

Following on the same theme Anne Sophie, I think maybe we can use your tool as an example for these therapeutic decisions. And I think physicians, if you have 50 variables that you have to enter to come up with which treatment I’m going to use, and let’s assume for the sake of discussion here that we have multiple treatment choices for higher risk MDS. Are you designing your tool or are there attempts to design the tool where it extracts a lot of these variables automatically from the medical record? For example, the path report probably has 20 different variables between the blast count, the degree of dysplasia and each line, some fibrosis, other things. Does it extract it automatically and then gives a decision or do you have to actually manually enter every single variable? Because I do agree with Aziz that even if this is a most accurate tool, it’s going to be very difficult to implement from a practical point of view for wider use.

Anne Sophie Kubasch:

Yeah, you’re totally right. This is an important fact. I would describe it as a vision. So our vision is of course, that we can extract it automatically from our hospital system, from our digital systems we use for our patient documentation in our inpatient centers. And of course also the outpatient center, but maybe it’s the same problem not only in Germany, also in the US that every hospital and every outpatient center is using different kinds of documentation systems. So it’s not easy to get along with it, and therefore we think we can connect KAIT to our local system in Leipzig, this will be no problem, but it will not be easy to connect it to other university hospitals because every hospital it’s so different in using their programs. And therefore we think it’s a vision for the next years. But I think within the first year of using it into the daily practice, it will not be easy. And therefore we still need manual extraction of this data.

Amer Zeidan:

I think that’s a good point, but even when you have many hospitals in the US use Epic, it’s one of the most used systems, but even when you use the same system there are certain data elements, for example, the path report. I think you have to even develop algorithms to how to extract the data because it comes in a PDF. It’s not like: fibrosis yes or no, or it’s not done in a way that you can extract data easily. And I think that’s a big challenge because I think to make these things where you need algorithms that actually read these reports and are able to automatically extract the relevant data without a person reading the report and then entering the data, which I think is another layer of [crosstalk 00:41:52]

Anne Sophie Kubasch:

Absolutely.

Amer Zeidan:

…that’s added.

Amer Zeidan:

So I think we’re coming to the end of this discussion. I just want to give you the opportunity to close with any additional thoughts that you might have, maybe I’ll start with you Anne Sophie.

Anne Sophie Kubasch:

Thank you so much for this very interesting discussion. So I think maybe as a future outlook, we saw that the future can be bright in the application of AI in medicine and especially in the field of hematology, but maybe this is also a learning point from our development of KAIT, that a good registry and good patient annotation will be much more important for the application of AI in medicine. And therefore, I think maybe this is also learning for other physicians watching our video that it is very important to collect patient data on a regular basis. And then at the end, hopefully we can use these training data sets for the application of AI.

Aziz Nazha:

Awesome. Thanks Anne and thanks Amer for the opportunity to be with you in this lovely discussion. I always end any talk I give is the electric light did not come from continuous improvement of candles. So if we really want to improve the outcome for our MDS patients, and we certainly need to, because we have not moved the needle that much even in the last decade or so, we really need to embrace the novel technologies. And I’m not saying it’s going to be the answer. I’m not saying it’s going to be the cure, but really we have to take giant big steps, bold steps to change how we think about diagnosis, prognosis and selection of therapy, and start to embrace more digital technologies, as AI one of those technologies, and hopefully to leverage that technology, to get us to where we need to go faster and better.

Amer Zeidan:

You’re absolutely right. And I think these concepts of incremental change and evolution versus revolution and disruptive technology, I think all of these are going to be really important because a change in medicine has historically been somewhat gradual so major disruptions take time, but I hope it’s going to be happening in a way that will make life much better for patients and for physicians, clearly. Thank you so much. And this takes us to the end of this session of VJHemOnc MDS sessions and we will catch you next time. Thank you so much.

Anne Sophie Kubasch:

Thank you.

Aziz Nazha:

Thank you.