How AI Is Shaping Your Cancer Care

Lisa Hatfield: 

Artificial intelligence is not only changing how patients search for and understand health information, it is also beginning to reshape how cancer is researched, diagnosed, documented, and treated. Hello, I’m your host, Lisa Hatfield. In this DECODE podcast, we’re looking beyond the chatbot to explore how AI is being used across cancer care, from clinical research and earlier detection, to AI scribes, patient data, privacy, and the future of medical decision making. Please remember to download the resource guide for key takeaways and resources to help you better understand how AI is shaping your cancer care.

I’m joined by two experts who bring very different but complementary perspectives to this conversation. Dr. C. Ola Landgren is director of the Sylvester Myeloma Institute and a renowned physician-scientist exploring how AI can help advance precision cancer care and more individualized treatment decisions.

We also have Kacy Hatfield, a PhD researcher at Arizona State University’s Heart AI Lab, where she studies interpretable AI and how people can better understand and trust these systems in high-stakes settings.

Kacy is also my daughter, which gives her a unique perspective at the intersection of AI and the patient experience. Dr. Landgren and Kacy, thank you again both very much for being here.

Kacy Hatfield: 

Thank you.

Dr. C. Ola Landgren: 

Thank you very much for having me.

Lisa Hatfield: 

To begin with, I’d like to look at where AI may be having some of its greatest impact already in cancer research. So, Dr. Landgren, how is AI being used in cancer research today, and where do you see the greatest potential for it to accelerate new treatments and deepen our understanding of different cancers?

Dr. C. Ola Landgren: 

I think AI is used in many different ways in cancer research these days. Anything from more practical things, you could use it as a tool to provide patients with available clinical trials, so you could have it as a tool where you enter information about an individual patient, and the AI could help and say, this patient would actually be potentially eligible for this and this and this clinical trial, and could even suggest and say, given what the patient previously was treated with, this may be the best option.

So these are the types of processes that normally we would have trained either research nurses or physicians or other healthcare providers doing. So AI could be a tool to help the team members that could also be resource nurses and doctors, and get through more patient data and more trials quicker. So I think that’s something that we are using in our clinics, and many teams are doing that.

We’re also using it as a tool for the actual research, and just some high-point examples. You could develop, say, a intervention in the lab to try to treat in a certain situation, and there are so many different ways of doing it, so you can modify molecules and structures in different ways. So the AI could sort of seek out which are the different options, and could test, based on information we have about the different alternatives, which may be the best option. So you could seek sort of optimization for things you try to develop in the lab.

And lastly, one thing that we have spent a lot of time working on is also prediction models.

Dr. C. Ola Landgren: 

So, you can think of clinical trials, you treat, say, 100 patients with the same therapy, and then you say, what’s the median duration of the benefit? The median duration could be, say, 48 months. So, that’s how we historically have looked at data. But the truth of the matter is that no one is average, no one is median.The median is the summary of the whole group.

There will be patients that have much better outcome. That could, unfortunately, be patients that have much weaker outcome. So the average is just in the middle of this.

So if you knew a little bit more about those patients that did really, really well, and now you have a new patient looking at this therapy, if you could see what you have in common with, you could say that given what I know about my disease biology, my clinical overall status, my prior therapy, and so forth, looking at those patients that look similar to me, like a digital twin equivalent, what would be my predicted outcome? And that’s a very different way of approaching data.

So we’re going away from the median towards the individualized outcome. And I think this is going to revolutionize a lot of the work that we are doing. We have done a lot of work already.

We have developed models where we can predict outcome. We can also use these tools to say, given what we know about the biology, if you look to see patients we have treated in the past, which therapies work the best for this particular type of situation. And the model would say, based on all these cases, this seems to be the best option. And you can look at the data. So, I think we will have a lot of tools going forward for decision making and for better understanding of how we can advance the field.

Lisa Hatfield: 

Okay, thank you for that. So another question I have as a patient is, when it comes to diagnosis. Sometimes diagnoses are delayed. Patients have a very common symptom. I had back pain for a couple of years. I always went for my wellness exams. My diagnosis was delayed by about 2 years. Do you foresee AI helping prevent or reduce delayed diagnoses in cancer patients?

Dr. C. Ola Landgren: 

Absolutely, I do think so. And I think clinical medicine, going to med school, really, it’s about pattern recognition. So you’re trained to look for a combination of different factors that are off. But the truth of the matter is that most diseases have overlapping abnormalities with other conditions. And some things can also be normal. A lot of people could have back pain because they were carrying heavy bags, so they did exercise, or they did other things that trigger that back pain.

So some of the normal and some of the abnormal will always be overlapping, and that’s the challenge with clinical medicine, and I think a good clinical doctor is someone who is really able to put together all those pieces and set the correct diagnosis. Some diagnoses are very difficult. Myeloma is a difficult diagnosis. I think the AI could help and say, maybe to someone who sees a patient with certain symptoms, that here’s a list of the most likely. Have you ruled this out? And maybe a blood test could speed up the diagnosis.

Lisa Hatfield: 

Okay, thank you for that. Yeah, I had a great doctor, he’s retired now, and just, I wasn’t the normal candidate to have myeloma, so it took a little while for the diagnosis. So, I often wonder, well, if AI had been at play then, and maybe picked out some obscure…I don’t know, blood test or something out of the x-ray I had, maybe that would have led to a sooner diagnosis, thanks for explaining that.

So, from the research lab back to the, to the clinic, Dr. Landgren, are you using AI tools like AI scribes to document your clinical visits? And if so, how has that changed the way that you interact with your patients?

Dr. C. Ola Landgren: 

So, AI scribes is another thing I didn’t mention, so thank you for bringing that up. So, we do have, at our institution, at the University of Miami, we have access to AI scribe tools. So, we download a software on our cell phones, a little app, and we link that with our electronic medical record system.

And we can ask the patient if we go in the room, is it okay if we use this system? And if the patient says yes, you can start the app, and you have your phone on your desk, and you can have a regular conversation with the patient. So, how are you feeling today? Anything that has happened since we saw each other last time, any new symptoms? And you can examine the patient, and you can say out loud when you examine. I just listened to the heart, it’s regular rhythm, there are no abnormal findings, etc, etc.

And then you could have a discussion about what the next steps would be. We look through all the labs, there are no abnormalities, everything is within the normal range, there are these and those things checked, and everything looks normal. So let’s do a follow-up in, say, 2 or 3 months with blood tests, and we stay on the same treatment. The AI would then summarize this per a formula for how the notes can be written. And I would say it’s quite good.

But I’m very specialized, as you know. I’m a super specialist in myeloma and plasma cell disorders, so…the way it puts it together is not exactly the way I like it. So, I like very detailed information for sophisticated decision making. So, I have trained a team for years. I’ve been a doctor for over 30 years, so the way my team is trained is to put together information for optimal decision making. But I guess we can train the model to be more like the way we like to put together data.

But the model right now is more general, so it probably works across the board in a better way than it does for very specialized doctors. But I’m sure these models are getting better and better in the future.

Lisa Hatfield: 

Yeah.

One thing I like about the use of AI scribes in the clinic setting, in the exam room, my oncologists are great talking to me face-to-face, but since the implementation of electronic medical records, I have one nurse practitioner that’s… it’s no longer face-to-face, it’s face to back of head, so she’s using the AI scribe now, and we’re now talking face-to-face again, and I really like that.

I just think it’s better communication, so… I like the use of that. But then to talk about privacy, so how… is that information that’s collected, is that… does that just go to your healthcare system, and is it HIPAA protected? What is the privacy situation with that information that’s collected?

Dr. C. Ola Landgren: 

I’m not sure I’m an expert on exactly where the information goes, but from our user perspective, we have an app where the conversation is being recorded, and then it’s transcribed into a note into the medical record system, so that’s as safe as any medical record system. But where the recording is, that’s something I’m actually not sure about.

Lisa Hatfield: 

Okay, okay. And then Kacy, so a patient is uploading something sensitive, like a biopsy report or their lab results, into a public AI platform, not the same thing that the hospital’s using to collect the information. What should that patient understand about where that information may go, and how it may be stored, and what privacy protections are applicable there? Is it protected?

Kacy Hatfield: 

So, no, on public AI platforms, it doesn’t have to be HIPAA compliant, like anything used in healthcare, and so it’s important to make it as de-identifiable as possible and anonymize it as much as possible. Something like a biopsy report, once you upload it to a public AI platform, you can assume you’ve lost some amount of control over it. It may be stored to help train the model, and deleting it from your chat doesn’t mean it’s deleted from the server that it’s likely stored on.

We really don’t know where these servers are, or how much of our data is stored, and so I’d treat it like posting something semi-publicly, and you can often check the privacy policy of these AI platforms, and try and opt out of training, if you would prefer, but the main goal is to de-identify it as much as possible.

Lisa Hatfield: 

Okay, so if I take a screenshot of my lab results, do I just crop it and cut off my name and date of birth? Is that good enough?

Kacy Hatfield: 

Cropping is a good way. It’s a gray area in terms of user discretion and privacy. If you’re comfortable with it getting uploaded, as long as your name and any patient number or identifiable information is removed. And you have a… generally, paid accounts are a little bit more private than the public AI platform accounts, but still treat it as posting it semi-publicly.

Lisa Hatfield: 

Okay, and this is kind of an obscure question, but two people have asked this question in the past. Who owns that data? Can we retract it once we’ve put it into the public platform? Is that possible?

Kacy Hatfield: 

As far as right now, you cannot retract data. The right to agency of one’s own data is a heavily debated topic in AI and is very interesting, but in terms of who owns it, it is still debated whether you have agency over your own collected data.

Lisa Hatfield: 

Okay. All right, thank you. Well, I have a question for Dr. Landgren that I’m actually really excited to ask about. So, Dr. Landgren, can you give us an example of an AI project that you and your team are working on, and walk us through how patient data are used, and where does the data come from, and how are privacy and confidentiality protected, and what do you see the outcome of that project being?

Dr. C. Ola Landgren: 

We have spent a lot of time working on prediction models, so that’s a big topic in our AI group right now. So we have gathered information on the genomic features of the disease beyond FISH and cytogenetics. We have used whole genome sequencing and whole exome sequencing data. We have also gathered clinical information on individual patients. And then we have also, for each of these patients, collected the treatment data. And then we have the follow-up information. We have the progression-free survival and the overall survival information for patients.

We have thousands of cases like this, and as I mentioned, you can study the median, but you can also use these AI models that we are developing to identify patients that are biologically more similar. So you focus on cases that are very, very similar, and if you introduce a new case to the database and say, tell me, for patients that look exactly like this what will be the optimal treatment?

So, that’s something we have spent a lot of time working on. And the idea is to see if we can optimize outcome by using therapies, where there is evidence from the database that that is the optimal therapy. You’re also taking it to the next level. My hypothesis for a long time has been, when you look in a microscope and you see plasma cells, you call that myeloma. When you do FISH and cytogenetics, you see that there are many different subtypes. If you do sequencing, whole genome sequencing, you can actually see between 3,000 and 5,000 mutations per patient.

So my thinking has been for a long time, when you just look in the microscope, it looks the same, but when you start doing these genetic tests, it doesn’t look the same at all. It’s probably different diseases.

So what if you had better vision? Better vision than a human being? If you could look at pictures of the tumor, and you could make sense of it mathematically, could you actually see different diseases? So we worked on that for a long time.

And we have developed a system called CORAL, where we can evaluate pictures from a single biopsy. And we are now at the point where we, with one picture of a biopsy, or just a regular bone marrow biopsy, we can, with, say, 90% probability, have the computer saying what genetic subtype this disease is. So you don’t need to spend thousands of dollars doing expensive testing. You can just take a picture of it. If you do high resolution, and you run it through the algorithms we have developed, you can nail the subtype. Not 100%, but better even than the regular test. The regular tests are not 100% either, they are less than 90%. So, you can, just with a picture, come to the conclusion. And then we have integrated this with exactly what I was talking about, the treatment and the other clinical variables. And we have also expanded these models to pull information from the internet from other trials. So, you can start… you can start to model the added benefit of drugs to the therapies that are published.

And you can predict what would have happened if you did in a certain way. So, I think this is mind-blowing, and I think you could think of this as a tool to better guide decision-making for individual patients. And I think you could also think about this as a tool to even emulate the control side of a clinical trial. So, right now, we do randomized trials, but if you knew exactly what control arm you’re gonna use, you could take that information from a computer-based system, and you could have virtual controls.

I think this is going to, of course, cause a lot of drama on many levels. The regulators probably wouldn’t like that, so you could also do 10% to 20% of the controls coming from digital controls. So, the benefit would be that it would be quicker to do the study, it would be cheaper to do the study. So you don’t have to toss everything out, so you could do parts of your control from virtual controls, and we have worked on that, and we have very powerful tools that are very, very, very exact.

Lisa Hatfield: 

That is fascinating. So are you saying that there’s a chance that, a cancer patient may be able to go in. Plug in their patient profile into this data set that you have, and this patient might not have to be on a drug for 3 or 4 months, because you’ve already identified that that is not the best therapeutic treatment for that patient? Because right now, sometimes we cancer patients, we have to go on… it’s trial and error. We were on it for 3 or 4 months with all the associated toxicities and side effects. Is that… what you’re kind of saying is that this is a possibility, that cancer patients may have a more very customized, individualized treatment plan?

Dr. C. Ola Landgren: 

That is exactly what I’m saying. I’m not saying that we have that for every single drug, but to give some more practical answer back to you. For example, if you give, say, an immunotherapy, such as daratumumab, a CD38-targeted antibody.

That antibody will, to a high degree, depend on natural killer cells or T cells. So the antibody would bind to the myeloma cell, and then it would attract these other cells. So, again, if you come back to what I said about the picture. If you could train, and we have already done this, if you train your program to look for certain features in the biopsy, you can tell whether there are a lot of T cells and natural killer cells that are able to do the job with the drug.

So instead of trying the drug for 2 or 3 months, and it doesn’t work, you could take a picture, and you could say with 90% probability this will work. That’s much better. And I think that’s the way the field is going.

Lisa Hatfield: 

All right, that’s… as a patient, that’s really hopeful. I’m happy to hear that, that that’s being studied, and you’re researching that.

And so, Dr. Landgren, we’re gonna kind of shift over to, biases in data and equity, so one concern about AI is that the data used to develop these models may not fully represent every patient population. Is there a way that researchers can make sure AI tools work well for patients across different racial and ethnic backgrounds or geographic areas and healthcare settings, including people who may not receive care at a major academic center?

Dr. C. Ola Landgren: 

I think those are extremely important questions you’re raising, so it’s hard, probably, to overcome all those barriers in sort of one step. So what we have tried to do is that I mentioned about the images project we have worked on. We set it up at our own institution, but we then expanded it from the East Coast to the West Coast in America, so that was a way to, sort of improve the generalizability. We also established collaboration with groups in South America and Asia, and also in Africa. So, we have worked with groups to get access to data that way.

And that has also allowed us to see how quality of the data can vary, because if you do it at one institution, you can always optimize everything you do, but when you start working across different groups, even within the United States, the quality will not be the same, and people will do things differently.

So I think that is a hurdle that is not unique to AI. It’s just, like, generalizability. Clinical care sort of has the same limitations. Doctors and nurses are trained differently. Institutions are set up in different ways. So, it’s not a new, new challenge, if you want. I think then you ask about, like, ethnic and racial groups, so many drugs can work the same across ethnic and racial groups, but some drugs seem to work a little bit differently. That’s an area where there is not a whole lot of studies done. It would take more testing, probably. That is not only necessarily AI, it could be also biological testing to look for, say, how a drug is metabolized and how it interferes with different types of biological systems in the body. So, you’re asking a lot of very difficult and important questions.

Lisa Hatfield: 

Thank you. Kacy, I’m just going to ask you a really quick question. We talk about those racial and ethnic background differences. Historically, in clinical trials, there are some populations that are underrepresented, so is there a way… I don’t understand anything about how programs are coded. Is there a way to code in and tell AI we have this bias, we don’t have this representation, can you overcome it, or do we just collect data from this point forward that has that clinical trials are now trying to incorporate these underrepresented populations? Is there a way to code that out of it? You know, to code out this bias, or is it… is it just what it is? It’s historical data, so you can’t code it out. Does that make sense?

Kacy Hatfield: 

Yes, it makes sense. Bias is an entire field of AI, determining not only biases that maybe we can’t identify right away, but are working to identify, but also how do we mitigate, we call it post hoc or after-training biases. So it really depends on the use case, whether training augmentation or training data augmentation, bringing in or balancing data sets to make sure all applicable categories are of the same representation are shown to the model or not. So it depends on application, and there is an entire field working towards it. So there are many different biases, and depending on the bias, it has different solutions. in balancing the historical bias to the future predictability.

Lisa Hatfield: 

Okay, I kind of put you on the spot with that question. I have so many more questions, but I know we’re out of time, and I know Dr. Landgren probably has a million things to do.

So, I have one last question, really quick, Dr. Landgren and Kacy, as AI becomes more embedded in cancer care, what is one question that you think every patient should feel empowered to ask about how AI is being used in their care? So, Dr. Landgren, do you want to start with that?

Dr. C. Ola Landgren: 

I’m not sure I know exactly. Maybe one answer would be that a patient could ask, how could AI help me and guide on the optimal next treatment decision to get an overview, and that could sort of stimulate more of a conversation with a doctor. So, I think that would be one thing I could think of.

Lisa Hatfield: 

Okay, thank you, and then Kacy, how about you? What is one question that every patient should feel empowered to ask?

Kacy Hatfield: 

I would also maybe say they should feel empowered to ask how it is currently being used in the clinical setting, in clinical visits that they attend, to know exactly where their data is going, and…have, at least be knowledgeable about where their data is being used in terms of uploaded to AI.

Lisa Hatfield: 

I have so many more questions. I know we could talk about ethics, that’s kind of your area of expertise, Kacy, but we don’t have time. We ran out of time, but Dr. Landgren and Kacy, thank you so much for helping us better understand how AI may influence research, diagnosis, privacy, and decision-making across cancer care.

And to our patients and care partners, perhaps the most important takeaway is this. AI is an increasingly powerful tool, but understanding how it is developed, how your information is used, and where human judgment still matters can help you engage with these technologies more confidently. I’m Lisa Hatfield. Thank you for being with us.