Healthcare Unfiltered Express
Healthcare Unfiltered Express curates quick-hit conversations between Dr. Chadi Nabhan and his guests on hot-button topics in the healthcare world today.
Healthcare Unfiltered Express
Episode 56: AI and Patient Care
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
On this Healthcare Unfiltered EXPRESS, Dr. Arturo Bonilla, CMO of Massive Bio and an AI oncologist expert who leads many AI efforts nationally, joined Chadi to provide his insights on where AI is being applied in patient care. He gives timely use case examples that will resonate with everyone.
AI is here to stay and we must know where and how it applies.
Please join the conversation, subscribe, like, comment, and write a review.
Arturo, welcome to Healthcare Unfiltered Express. I appreciate your time. Thanks, my friend.
SPEAKER_01Thanks as always to have me here.
SPEAKER_00So I see a lot at this, you know, uh on social media, on LinkedIn and so on, very heavily involved in AI. So this is your the perfect person guest for what I want to talk about, how physicians are using AI, but also how patients and families are using AI. But a little bit about you in terms of what you do, where you practice, and what got you interested in AI?
SPEAKER_01Yeah, uh, so yeah, for those who uh don't know much about me, um, you know, I'm um, you know, Arturo Lois Buni, I'm a medical oncologist. I focus on GI malignancies for several years. Um, I am the chief of HEMONC at St. Luke's University Head Network, which is a network of uh 17 hospitals now that uh operate in the Lehigh Valley. So I've been in Pennsylvania for uh you know a decade and a half now, um, serving the community. And um, because of um my interest in technology and and and out of frustration some uh to some extent on seeing how technology can really you know track my my food and and you know where an Uber can be, but they cannot even tell where you know how can I help my patient for uh you know a trial or for decision support with all these new tools coming up. That's what got me into AI. So um I used the old AI before, that the one was more like this, the symbolic AI we call it, like more like the decision support. Um but then uh with the advent of you know these you know the GPT technologies, um, with the the we call now neurosymbolic AI. So I'm using the best best of both worlds um to use um you know different uh you know different cases for in the example, for example, uh pathology, or how we can implement radiology models for our patients in the ERs, um uh or if we can use a decision support uh that is implemented with an uh electronic medical record that we have in place. So a lot of moving parts, very exciting time. Uh, and and to me is about being practical. Uh so uh and for those who are interested, that you know I have a um a Metscape um column monthly that's called um practical AI. So if anyone wants to wants me to write about something different that they're interested about, happy to have a conversation as well there.
SPEAKER_00That's wonderful. Congrats on everything you're doing. So I've asked you to come on the show, Arturo, because I want to give viewers and listeners both sides the side of the practitioner and the clinician, and then the side of the patient and the caregiver. I have interviewed uh a patient and a caregiver who gave me an insight into what she was doing and so on. But I want to see that from your lens. Let's start by you, as a physician, clinician, researcher, and provider. How are you using AI? What are you using, and how are you implementing that in your practice?
SPEAKER_01Yes, so I um uh I think the vast majority of the US physicians uh have some form of AI um expertise to some extent or experience. Uh, the first one is with uh ambient AI. So we ambient AI is just this um um artificial intelligence uh scribes where you use your cell phone and click a button and then they can record the conversation or listen to it, and then gives you a summary uh into the note. So you don't have to be spending time documenting or typing too many things, but you just rather spend time with the patient talking to them. Um I also use it within my electronic medical records, so we have Epic, but of course there's many other CERN and beyond that are doing kind of similar approaches where they're doing data summaries. So I leverage those data summaries to be able to know what's happened with the patient the last visit. If there's any emergency department visits, I ask questions and I do prompts. So they use certain techniques called RAG, uh, which is retrieval augmentation generation, which is just taking chunks of the data and giving you a summary of what's happened to the patient over a certain period of time. And they put a ground truth, which is basically giving you the uh specific reference, you may say, as if you are looking in a chapter of a book, but tells you where that CAT scan was, what uh where the results are.
SPEAKER_00So hold on, hold on. You're using this, so so you go and you upload like a progress note on a on a large language model and you search it. What do you how do you do that?
SPEAKER_01Yeah, so uh Epic has uh already an integrated LLM uh that is independent of the front-end models that we use. Uh, there's their own development, you know. You have anyone can do this technology these days, it's very commoditized. So they they had it into um it's called Cosmos, the whole thing. Um, but the you look at the LLM from um from Epic and you click summarize here for me. So the last story. So it just basically does this little loop and it gives you the summary based on a question. Say, I want to know what how what happened with the patient since the last basic with oncology. And it will pull up the information using the same structure as if you're doing you know chat GPT or Claude or Grok or Gemini and gives you that information. I use that not to copy paste directly because sometimes you need to verify it and we need to grade benchmark better, but it's very helpful for me to, hey, okay, the patient had a scan, I didn't know. I look into that and I can use that information to guide my notes. So I'm saving time and effort compared to what I used to have in the past, trying to search one by one all the different uh you know taps in in my EMR. Now I'm using that on a daily basis to make it easier for me to look for things.
SPEAKER_00Interesting. Okay. So um in addition to that, uh you mentioned you're still using like the radiology pathology type of thing and and so on, or is that still uh happening? Yeah, so um we uh Lisa St.
SPEAKER_01Luke's we have a good relationship with the Microsoft Foundry. So we're actually evaluating models that we deployed into the emergency department for not just phoncology for like fractures and uh finding lung nodules, and we can use that information to get the patient to see a pulmonologist, for example. Uh so detecting it from uh from random x-rays that are done in the emergency department or just done as part of regular workup. So those are evolving. We uh in and there are different use cases across. We have a actually a committee that is looking for what algorithms should be deployed and what are the benchmarks that had been set up so we can reliably use them in our population. Uh so because not all models are the same, as you know, for example, there's mammograms right now, uh models that have been very well proven in the in in Europe, uh, you know, like the Maasai trial that had like you know 80,000 patients total tested. Uh, but now that those are different populations in the US. So now there's a study uh called, I think, the panorama trial or something like that. No, panorama is for like uh no uh prison trial. The panorama is for pancreatic cancer, because close to me. Um and that is one is looking for uh the patient population in the US that had a different breast density and using mammograms for uh for assistance uh to basically do don't do double readout. You only need one readout. Uh so those are things that I'm I'm doing now on an ongoing basis. And uh and I also, of course, do the trial matching and you know, been doing it on the massive bio side on the independently, but I still use massive bio for my patients. Um I get them to sign up and then and then I can have the trial report for them faster. So um so those are multiple use cases I do, and of course, literature search. So there's a lot of different things happening in that space. So we have open evidence, we have Netscape AI, we have the um Doc GPT, the GPT one for Doximity, I think Docsimity GPT, and also up-to-date AI. So all of them are fighting for our ideas. Um, so I'm trying to you know leverage the best I can and get the best information from them.
SPEAKER_00Now let's shift to patients and family and caregivers. What have you seen, your patients, the families that you care for? How are you seeing them utilizing AI?
SPEAKER_01Yeah, so there I think it's a generational gap because of the use of technology. The digital divide is not just because of access, but also because of knowledge and and and um some form of digital literacy. So most of the AIUs I've seen have been by caregivers uh of you know, they're like on the younger side of the of the spectrum. Um, so they can use that for their parents or grandparents. Uh, they have these kind of data summarizations from I see them from Gemini, I seen some Chat GPT, I see I saw someone from Perplexity uh basically giving me uh a report and say, hey, what do you think about this? Or I look at the I'd look at the records and this is the summary. Uh so I think there's good efforts here, but the problem is that they haven't been fully homogenized. So uh some data sources are good, some others are not that good. And it's also depending on the questions. Uh so most patients and and caregivers have never been trained on prompt engineering, but neither physicians, you know, we have to learn that. Uh, but asking the questions because the models have this something called sycophancy, where they tend to give you an answer that you're seeking, right? They they kind of like always want to reaffirm you, like make you like the best friend. So that sycophantcy lets to give them answers um and kind of like a confirmation bias. So we we need to be probably stewards of that as well and explain to patients. I've been taking a little bit of education on that, uh, telling them, okay, this is how you should look for data if you're really looking for options, and what are the sources that you should look? Um, so um, and that's very emergent. You know, Chat GPT yesterday uh said that there was gonna be this uh, you know, uh Open AI health that is going to get your kind of medical records and start collecting them and doing use cases with that. Um, so that's exciting, but yeah, I think we need to do prospective data uh because we don't we we want we don't want replacement, of course. We I always follow the Morvik's paradox. You know, there's things that we can relinquish to AI, but there's a lot of things in oncology and beyond that really belong to us to take home because we need we know we need to make sure that the sources are okay.
SPEAKER_00Uh are you are you starting to utilize AI in the way you teach students, residents, fellows, and and so on?
SPEAKER_01Uh yes, um I um I used it for two major um um uh you know activities. The first one is to um do summarization of like conferences and stuff and things that you know I don't have the time to go to all the different things. Some people have their lives dedicated to that, which is amazing, uh, but I don't have the time for that. So so I use that to give me uh do a deep research and they give me a summary. And then I use that summary to create you know slides that help me to teach my fellows or or residents or myself. Um and and I use some data sources, like for example, uh, you know, ASCO GI happens, so we have the best of ASCO GI. I know the topics and the abstracts, so I can at least summarize them all and put them in front of me so I can be very nuanced on how we're doing this. I'm doing the AI updates on GI oncology as well. So I'm doing precisely that. Um it helped me to frame the information, and I can be much more effective in putting the slides that actually make sense. Um and and yeah, I think it's a matter of knowing the right sources. So to me, it is I'm very specific in how I do the prompts and what data I want to pull. If I have a full uh paper that I'm really interested, I'm asking, okay, I want to just distill these pages and give you the information that I need to do. Yeah.
SPEAKER_00So my last question to you, Arturo, is it's all nice and dandy when we talk about AI sometimes. But as we know, there's always the other side. So, what concerns you about AI? What are the things that maybe you have some reservations about or that would keep you cautious about AI? In medicine, of course.
SPEAKER_01Yeah, in medicine, um yeah, it's always the end use, right? That the AI can be used for good and and for not so good. Um, so there was a recent paper showing you that the same AI tool for decision support, if it's being used by a physician, gives you an answer, but if you use by a payer, it gives you a different one. So if if the physician is using it, it will say, Yeah, this is 100% appropriate for you to use. If it's used by a payer, it will say, Yeah, but you should try this other thing because it may be not be as good as cost-effective. Um, so uh it it the key component for me is to have the human in the loop. And in this case, the physician in the loop as the part of the validation of these, uh, all these algorithms. Uh, a lot of the things are happening in AI are have been done in a vacuum or done by non-physicians, mostly technologists, or the physicians are just like little annotators, but they're not the ones helping to decision support the whole process. Uh, so what I'm mostly concerned about is um missing out on the data that we physicians have to give to make it much more contextual and nuanced because that's how drives utilization. I will use something, only something that really applies to me. Uh, and I don't want to miss also uh uh a unique case or or a potential patient that um that may be part of that AI inclusion. So I don't want to miss, for example, in talking about mammograms, it really needs to be the uh applicable to the population that I treat. Uh, if I'm looking at an algorithm for you know expression of PDL1 on gastric cancer, they it has to be applicable to the you know Western population because we know the Asian population will be different, but they have much more prevalence of certain diseases. So all those models need to be fully generalizable. And as of now, we don't have appropriate benchmarks for any of those things. Zero. None of the models that you see deployed have any real bench benchmarking except for like, oh, it passed the USMLE test. Well, great. If I have an open book and all the time in the world, then I'll get 100%, or at least decent passing rates. That's the concept of MOC, right? So it's the same. Like, yeah, you can pass MOC. If you have five minutes and open book and any data source, the same happened with LLM. So to me, a benchmarking of passing the USMLE or the bar or questions that are going to be found somewhere and you have the answer already is not really meaningful. What is important is difficult questions that are nuanced, contextual to the patient, and really generalizable to what we do and the complex care that we deliver.
SPEAKER_00Arturo, this was amazing. Very, very nice, concise, and just give us like both sides of the of the stories. Anything else before I let you go?
SPEAKER_01Um, no, I I think this is one of the several conversations we're gonna keep having. So um I'm excited to see what AI can bring. I think it can bring a lot of things for good. And uh FDA, uh, you know, all the you know different entities are paying attention. So I'm I'm interested to see how that evolves and how we can be part of the conversation. Uh and yeah, no, the only tool in my home here is you know, massive bio uh clinical trials for everyone. That's what I want for everyone. So uh we're working with the American Society.
SPEAKER_00And I really think that deserves an entire podcast on its own.
SPEAKER_01But no, I appreciate that. Um so whoever wants to see it, there's ACS Acts, American Society Acts is now nationwide. We're supporting every single patient that has a cancer journey, so uh fully free. So please uh take advantage of that because it's a really good opportunity.
SPEAKER_00Doctor Arturo Bonina, thank you so much for coming on Healthcare on Filters Express.
SPEAKER_01Thank you so much for the invite.