AI in Health Care: What’s Here Now and What’s Next-An Expert Panel Discussion Featuring Neel Trivedi, MD; Ross Filice, MD; and Torbjørn Furuseth, MD

[00:00:07] John Marshall, MD: Hey, everybody out there in internet land. This is John Marshall for Oncology Unscripted.

Question: is artificial intelligence in your world today? I know it’s in our world, and what we thought we should do is really drill down on AI 2026 and talk to some people who are really living it day in and day out, and try to understand what the impact of AI is, both good and bad, in our world of health care, clinical research, and the future of medicine in general. And I am lucky enough to know people who are really, really smart about this and who were also willing to join me for a sort of quick overview interview discussion around AI and health care. And I’ve got 3 of my good friends online, as I suggested: Dr Neel Trivedi, Dr Ross Filice, and Dr Torbjørn Furuseth from a little bit around the world, if you will, who’ve joined me. So, Neel, let’s start with you.

[00:01:04] Neel Trivedi, MD: Thanks, John. So, my name’s Neel Trivedi. I’m a medical oncologist specializing in GI malignancies at MD Anderson Cooper, which is in New Jersey. So, happy to be here.

[00:01:13] John Marshall, MD: And an Hoya graduate, let’s not forget.

[00:01:16] Neel Trivedi, MD: Right.

[00:01:16] John Marshall, MD: Come on. We’ve got to have that Georgetown plug in there. Dr Ross, Alisa, and I spend a lot of our days together, back and forth on Teams and on Zoom, etc. So, Ross, introduce yourself.

[00:01:27] Ross Filice, MD: Yeah, I’m Ross Filice. I’m an abdominal radiologist here at Georgetown with John, and I also do a fair amount of informatics and technology work for our practice.

[00:01:36] John Marshall, MD: We’re very grateful for all of that hard work, and I’m lucky enough to be in a position to work with some other folks around the world who are thinking about the application of AI for the future. So, Dr Furuseth, you introduce yourself.

[00:01:51] Torbjørn Furuseth, MD: Yeah. Hi, everyone. I’m Torbjørn Furuseth, CEO and co-founder of DoMore Diagnostics. And we use AI to do a deep analysis of the tumor tissues so we can predict outcomes and personalize treatment for cancer patients. I’m a medical doctor from Norway.

[00:02:08] John Marshall, MD: That’s where you’re sitting right now. So, it’s after dinnertime for you while the rest of us are a little earlier in the day. Everyone’s taken time out of their busy day to join us. So, let’s get right down to it. And Neel, let’s start with you. You’re a busy oncologist day in and day out. Tell me a little bit about how you’re folding in what we call AI into just your everyday practice.

[00:02:28] Neel Trivedi, MD: Sure. The main way I’m using AI in my practice is the DAX Copilot. That’s an AI scribe that’s really been built into a lot of the electronic medical records. I was one of the early adopters at my last hospital, I think back in 2024, using it. I was just interested in trying something new that might help my practice, and it’s been great, honestly. It saves a lot of time. Before I used this scribe, I was spending a lot of time doing notes at home. I was basically in the patient room focusing more on the computer than talking to the patient, and having the scribe just listening on your cell phone in the room allows me to make eye contact with the patient much more, be much more present, and honestly enjoy my job probably more too. Because I didn’t go into medicine to document notes. I really went into medicine to treat people and really make them feel better, not just with medications, but with just human interaction.

[00:03:17] John Marshall, MD: So, you use this, right? You—

[00:03:19] Neel Trivedi, MD: Mm-hmm.

[00:03:20] John Marshall, MD: It listens, and then later on there’s a note. How do you do the physical exam? Do you say it out loud, or do you add it later? What do you do?

[00:03:27] Neel Trivedi, MD: You can. So, I’ll say the pertinent things, if there’s something different, especially if I’m describing a rash or something like that. I’ll say it out loud, but for the most part, again, my physical exam, I can just type it up afterwards. That’s really quick. It’s more with the HPI. So, in oncology, we talk to patients for a long time and really get a thorough history, and remembering every little detail of that is difficult. So, I think it’s been helping me take a better history too.

[00:03:50] John Marshall, MD: So, you think the quality of that note’s pretty good? It’s justifying the level of service and the billing and all of that. Before I let you go on to Ross, aren’t you looking stuff up? I mean, some funky gene mutation or you’re not quite sure the side effect of a drug? Are you using AI and searches and things like that?

[00:04:10] Neel Trivedi, MD: A little bit. OpenEvidence has been helpful, I think, for kind of basic lit searches. But I still kind of like to do things the old-fashioned way. I think as experts, we really need to know the science behind what we’re doing. So, showing your work sometimes can be difficult with AI. So, I’ll do it for a quick question, but for the most part, my searches are still Google Scholar and all those kinds of things: PubMed and stuff like that. But OpenEvidence can give you sources pretty quickly, so it’s helpful too.

[00:04:32] John Marshall, MD: And not much pushback from patients? Are they awkward? Do you have to get permission? How are you handling it?

[00:04:38] Neel Trivedi, MD: I start every visit saying, “Is it okay if I record?” even if I’ve seen the patient for a year or 2. So, they’re used to it. I’ve had maybe a handful of patients decline it, which is perfectly fine, but for the majority of people, I think everybody appreciates it because I’m, like I said, much more present in the visit. I was talking to my mom before I came on here, and she said a lot of her doctors are using it, and she likes it. Right? I think it allows them to really make eye contact, listen to her, take a better history. And so, I think for patient satisfaction too, it’s really important.

[00:05:04] John Marshall, MD: That’s awesome. Ross, let me switch to you. I think of what you do for a living as taking digital data and interpreting it. And so here we are, digital to human, back to information. So, talk a little bit about the evolution of AI in your field, in radiology in general.

[00:05:24] Ross Filice, MD: Sure. We use AI quite a bit. It was one of the first fields of medicine where AI was really applied, and that’s for 2 big reasons. One is that we are so digital, you know, sort of across the board. And because images were a natural place for AI to be applied as a transition from other spaces into medicine, we do use it a lot. It does help us. I think it makes us a little bit better in certain circumstances and makes us a little safer, and it certainly improves our care overall a little bit. But it’s interesting, you know, 10 years ago or so is when people started making quotes that radiologists would be completely replaced. And it was sort of obvious, to be honest, at the time that we wouldn’t, at least using what is current AI technology. And don’t get me wrong, I think it’s very valuable, but it hasn’t made a huge change in our ability to see and read more studies, right? It really hasn’t made us that much more efficient. It makes us a little bit better in a few niche places, but really hasn’t changed the way we practice largely. And honestly today, you know, John, as you know, we have more studies than we can read. We’re behind on our lists, and so we’re kind of desperate for AI to actually replace at least part of us, right? It’d be great if I could become twice as efficient. So, it’ll be a very exciting time, I think, in the near future as some of the future and new technologies in AI, you know, similar systems to some of the LLMs that we’re accustomed to, like ChatGPT or Claude, may be applied to imaging and may actually turn us into much more efficient radiologists. Because really, that’s kind of what we need in our practice and really across medicine, where we just don’t have enough of us.

[00:07:00] John Marshall, MD: I can’t help, as the oldest member of this panel here, remembering the day when you had film and people would load them into the big turn thing, whatever it was, where you could see a scan in panels. The whole thing would move up and down. Now it’s like, “Oh, let me find that. Oh, I can find the one that’s from Hopkins. I’ll just load it over here, and I’ll put it over here.” And we have it all at our fingertips now. It’s just an amazing evolution, revolution of that. So much so that our patients now also expect almost instantaneous results. And so we’re feeling that when a patient’s had their scan and then they’re waiting on that. And I know you feel it. I know we feel it. Back to your point about the digital stuff, you and I have talked over the years about, well, could it pick up a PE real quick, or could it tell the difference between that 8-mm nodule in the lung that’s cancer versus not cancer? Is there some evolution, revolution in that where that’s making your job easier, making those “cannot rule out” statements easier to say?

[00:08:11] Ross Filice, MD: To some degree. The PE example is a very good one. That one, the AI companies have done a really fantastic job of. We have it. We run all of our CT scans through our AI PE algorithm, and it’s very, very good at picking it up, so we really kind of rely on it. But it does make mistakes still. So, we haven’t gotten to the point where we can not scrutinize the images ourselves, unfortunately. So, it is very good. But not only does it need to be supervised, there still are thousands of other things we look for in a particular scan. So, that’s what I meant when I said it’s very good at niche situations. PEs are great, but that’s only solving a tiny fraction of all the things we need to look at on the imaging systems. And there’s really not a holistic AI evaluation system available today that can really look at an entire CT scan and sort of render all the diagnoses that we might be looking for.

[00:09:04] John Marshall, MD: I had a patient I sent for an MRI head because I was sure she had brain mets, right? And so, the nice resident called me and said, “Dr Marshall, you were wrong. The patient didn’t have brain mets, but we noticed 1-cm tonsil nodule that we’re pretty sure was on the [unclear] that you just did.” And by the way, I think AI would never pick that up, right? That’s a human figuring that out.

[00:09:25] Ross Filice, MD: Yeah. There are some algorithms too that will do a decent job of distinguishing between benign and malignant lesions. We have a thyroid system that does that. So, companies and researchers are looking at evaluations like that, where it’s not just, “Is it there or not?” but, “Is it dangerous or not?” But those are far from perfect. Not that we humans are perfect either, but they’re far from perfect. And again, it’s in pretty niche situations, you know, only in a few select diagnoses.

[00:09:53] John Marshall, MD: Before we kind of get to the future in digital pathology and all that, I want to switch gears there in a second. Is there stuff, Ross, that you don’t like about AI, that you feel like it’s a weight on you, or it puts you at more risk, or do you pretty much like everywhere it’s going?

[00:10:10] Ross Filice, MD: You know, largely, I think it’s great. I think it’s the future. I think it’s going to be applied across our worlds and work and pleasure and everything else pretty broadly, right? So, I think overall it’s really good, and we welcome it. And like I said, we need the help, frankly. You know, there are some potential negatives you have to be careful of, but I think overall, the benefits really outweigh those.

[00:10:31] John Marshall, MD: Thank you for sharing all of that. Dr Furuseth, I’ve got to get to you next because I see this as the future. When I come in and out of the country now, I have my passport in my hand, but no one ever asks for it because all they do is they take a picture of me. I stand in front of a camera, and it knows exactly who I am. I go to the guy or the gal at the desk in the uniform, and they’re like, “John Marshall, you’re in. Welcome home.” Right? So, I never open the document. They don’t have to prove anything. They know who I am. And right now, for cancer, for example, Neel and I are sending off all these specimens to get these huge terabytes of next-generation data that we then use to apply to a patient’s treatment. The work that you’ve been doing is more like my passport analogy, where you can take a picture of the tumor and know its characteristics and know how it will behave or not. So, maybe just a little bit of background for our audience about what it is your guys are doing, because everybody’s in on this. The digital path is one of the ways of the future, and AI enables it. So, give us a little bit of background on that first.

[00:11:44] Torbjørn Furuseth, MD: Yeah, sure. So, with the breakthrough AI and the large language model structure, it is now possible to process the images to a large extent that Ross is also using. And as Ross talks about, it’s very good at niche and specialized tasks, not for a general assessment of everything, where humans are still much better. But what we have done, and which is basically a new modality within precision oncology, is to train algorithms based on the histology image from the cancer versus a certain specific outcome. So, recurrence, survival, with or without treatment, so that you can train a super-specialized algorithm that really recognizes the pattern of a good, sort of benign, indolent tumor and a more aggressive tumor that the patient will probably die from.

[00:12:44] John Marshall, MD: Yeah. And so, when I think about this, the ability to take an H&E-stained image, scan it, put it into some AI algorithm, and within minutes, back to me, you know, stuff that right now is taking 1, 2, 3 weeks of turnaround time and the like. And arguably, when I think about it, the reason I’m so excited by it is that I don’t really know what it’s looking at because it’s looking at the entire tumor microenvironment. When I do a lot of next-generation sequencing, I’m only looking at the tumor and not necessarily the stroma. So, am I right in that thinking that it’s getting a bigger picture of the host and the tumor together?

[00:13:34] Torbjørn Furuseth, MD: Yeah, exactly. So, it’s basically the phenotype of the cancer, not only the genotype, and the whole assessment with the stroma, heterogeneity of the tumor, the microenvironment, all those factors that we know are important for the aggressiveness of the tumor and potential benefit of treatment and so on. So, it really adds. It’s a new modality that adds to the molecular analysis.

[00:14:01] John Marshall, MD: I’ve been really interested in the space because you and I obviously have known each other for a little while about this, and you and I are both eager for this to become more mainstream as a technology. But, you know, it’s only now that we’re sort of realizing the impact of next-generation sequencing and molecular profiling on therapies that we recognize that maybe digital path is a bridge, or at least a complement, to that. Do you see a time when I can look under the microscope with a digital path analysis and say MSI versus MSS, or a BRAF mutation or a ROS1 mutation or rearrangement? Are we going to get to that point?

[00:14:44] Torbjørn Furuseth, MD: I think we are at that point. There are algorithms that are able to exclude some mutations. I think to actually confirm it, you still would need the molecular analysis for quite some time. But we can, for example—there are algorithms that can say for half of the patients, these are clearly not MSI high, for example. These are microsatellite stable. For the remaining half, you need to do the molecular testing or the PCR.

[00:15:10] John Marshall, MD: So, it is evolving to that end. Ross, do you see that happening on your side with a lung nodule, for example, that you can say it’s negative or positive? Or do you think we’ll ever combine those 2 efforts, where you’ve got digital radiology imaging with some sort of clinical outcome?

[00:15:29] Ross Filice, MD: Yeah, people are trying to do it. You know, the field of radiomics, where you’re trying to get these features in imaging findings or tumors or nodules or lesions that may not be apparent to us as human beings, but may exist in that data, to try and, yeah, distinguish out whether things are malignant or benign or whether they’ll progress, etc. But nothing that we’re really using clinically today.

[00:15:51] John Marshall, MD: And then, Neel, back to you. I mean, if we get reports that look like that, right? So, we have to then, boots on the ground in front of the patient, decide: do we take this data and say, “Yep, that’s what I’m going to do,” or are we more traditional? We can’t rely on that yet. Where’s the sort of skepticism in our ability to take some of this fancy analysis and apply it to the patient?

[00:16:16] Neel Trivedi, MD: I think, as doctors, we tend to be skeptics. A lot of research is proving things wrong or proving things right. And like Ross had said, there are still a lot of mistakes in what AI does in a lot of different ways. My example is it’ll take whatever the patient or doctor says as truth. So, if you don’t proofread your note, you’ll have some weird stuff that ends up in there. So, like Ross was saying, I think we really need a doctor to proofread a lot of this stuff and see what’s usable and what’s not. But I think in these small niche areas right now, it can really be a big help, especially if you’re able to take an H&E slide and say, “Hey, there’s a high likelihood of a mutation here,” or not. I mean, you can spare somebody a biopsy, or you can get a biopsy and change their therapy. So, it can really help people.

[00:16:57] John Marshall, MD: Yeah, I mean, OpenEvidence, we’re all using this kind of thing a lot. And I don’t know if I’m right about this or not, but I tend to see what that spits back from science analysis and sort of trust it better than if I ask what Donald Trump is thinking, because that won’t be right. And I question that one, where the science one, I tend to give a little bit more validity. But it does come back to this: we have to recognize that was a machine that created that, and we have to decide whether it’s correct or not. So, maybe going around the horn backwards, Torbjørn, when you AI something to look it up from a science perspective, what’s your skepticism at that level? Do you believe it more than not? Where are you with that?

[00:17:43] Torbjørn Furuseth, MD: I’m a strong believer in the potential, and of course, there are some challenges as well. But I think what is important is that you also review or analyze how this AI is developed and validated because you need to have a separate training set, and you need to validate in a completely external, other dataset, and to be sure that it’s really generalizable to your lab, to your scanner, to your patient population, and so on. So, that is important. But when you see the data in thousands of patients from validations, that’s quite powerful.

[00:18:19] John Marshall, MD: Yeah. Ross, what’s your thought? What’s your level of, when you’re doing something for your own knowledge, your own trust of what spits back at you?

[00:18:29] Ross Filice, MD: Yeah, I mean, I think we’re largely skeptics of what AI is producing today. It’ll be interesting to see where that goes in the future. We, at the moment, sort of expect perfection or have to have complete supervision of all these algorithms, but that may not be the goal in the future because, of course, us as humans and doctors are not perfect either, right? Like me reading CTs is certainly not a level of perfect specificity and sensitivity. And so, will we ever get to the place where we as a society accept AI or machines to operate at a very good but not perfect level, which may produce all sorts of benefits in terms of efficiency and everything else? But we’ll need to accept some level of error, and I don’t know when or if that will happen.

[00:19:09] John Marshall, MD: Yeah. And Neel, you kind of referred to this earlier about needing to have the base knowledge in order to then reflect on what you’re seeing so you can kind of see, you know, is that really real or not, or where that’s coming from. Some of the senior members of the division who will go nameless—we were talking about this the other day. You’re a generation off of me. The newest people are a generation off of you. And they’re using this a lot. And one of my concerns is that the dependence on that, medical education and education in general, will remove some of that base knowledge so that you can be skeptical. Do you have any reflections or thoughts about that? As somebody who’s younger than I am, but you see the new ones coming along.

[00:19:57] Neel Trivedi, MD: I think a lot of the similar concerns came up when we were getting Google, right? These internet searches to find things out. And even the boards now are going to an open-book kind of format, right? Where it’s a more real-life version, where you can look things up very quickly. I think your concerns about losing that basic understanding are real, but there’s always going to be drawbacks to any advance in technology. So, hopefully, fingers crossed, everybody uses this responsibly, but we’ll see.

[00:20:24] John Marshall, MD: One last thing for the clinicians is that our patients have access to this too. They have access to their scans, their reports. They can look up stuff as well. Do you think that’s helping or hurting our cause? Neel, let me start with you.

[00:20:40] Neel Trivedi, MD: I think it’s helping. So, I can give you an example. I have a lot of patients who look up their diagnosis, their treatment on ChatGPT or another AI model, and it’s much better than talking to Dr Google that we used to have before AI was there, where patients would get a lot of misinformation out there. But the AI tends to be pretty good. So, I think it allows patients to advocate for themselves, educate themselves, and it’s actually been pretty helpful. So, I’ll actually recommend sometimes, if people have more questions, to use that instead of just Googling things.

[00:21:06] John Marshall, MD: And Ross, are you feeling that at all from a patient level?

[00:21:09] Ross Filice, MD: You know, we have less patient contact, of course, right? So, I’m certain that I have less experience, but it is interesting. There are places you can upload images to get diagnoses, and there have been services out there where you can upload chest x-rays and get a second AI opinion and so forth. So, you know, I think it’s the future. People are going to use this stuff, and I think we need to learn how to adapt to it and work with our patients as partners on it.

[00:21:34] John Marshall, MD: Neel, let me pick on you. Patients are worried about privacy. You said you’re going to ask them permission. It’s going up to the cloud, your discussion. How do you know that the cloud is not also listening and going to then tell Blue Cross or the patient’s employer that they have cancer?

[00:21:50] Neel Trivedi, MD: So, thankfully, it’s all HIPAA compliant, the AI scribes that we use. So, that’s reassuring. We always do a consent form that they sign and then also verbally confirm that they’re okay with us recording. But I think it’s a concern, especially in a world where cybersecurity is not 100%. But this is, I think, a concern for every facet of life these days. So, it is just something that we have to be respectful of.

[00:22:11] John Marshall, MD: You guys were awesome. Thank you so much. I know everybody’s busy. I’m going to let you go back to work and do your thing. But as you all have said, AI is here to stay. It’s going to get better and better, hopefully make our jobs more effective and efficient. But I’m also hearing that humans aren’t done yet, that we will need to be involved in the health care process. And honestly, as a guy who may need you one day, I’m glad you all are out there to take care of us all. So, thank you each for joining us today on Oncology Unscripted.

<<End 22:51>>

This transcript was originally generated by AI and lightly edited for clarity.