So, this is going to be kind of a big picture of, you know, we're all dermatologists, we're going to be faced with a lot of questions around AI starting today, yesterday, tomorrow, and every single one of us is going to have to think about it, and how do we do that? And I'm so glad you taught us so much about sort of the terminology and some of the active questions, and I'm going to try to take us through how do we integrate, you know, the science that Dr. Novoa told us about into our sort of everyday life in the context of AI and DERM. I do have some disclosures, but they're not relevant. I do not speak for any of the societies that I'm involved with. So where are we today? You know, there's one FDA-approved application that just came out in the last month for PCP's derma sensor. There was a prospective trial performed, but it was not randomized, and then we have a couple other skin lesion analyzers that were previously approved, Melafine and Nevisense, but actually none of these use just photos, so they all have proprietary detection devices, and so, you know, that's a little bit different from some of the stuff we're going to talk about today. For static classification tasks, so you know, when you look at a photo and when AI algorithms look at a photo and try to come up with a diagnosis, they have outperformed us since around 2017, but that's not what we do, right? You know, I don't look at a photo and say, I'm going to decide to biopsy or I'm going to decide to do something. I integrate that photo with a lot of other knowledge, and I think about my complete clinical workflow in the context of these technologies, and that's what we're going to talk about today. Okay, so I know we said the Q&A was later, but I actually want people to answer some questions right now, and if no one answers them, then I'm going to make my co-panelists do it. So, you know, just like raise your hand and tell me, you know, what would you like to automate in your clinical practice right now? Like what would be an example of something that you do not like doing that you wish a computer would do for you? Go ahead. Oh, the front desk, amazing. So there's so many tasks the front desk does, they're so overwhelmed all day long. Do they need help with certain things that could be done by computer scheduling, managing a wait list? You know, my front desk manually keeps a Word document of a wait list, okay? That should definitely be done by a computer. What other examples? Go ahead. Okay, I need you to actually shout. Oh, prior auth, great example. You know, now, you know, you have this set of documents on your hard drive, like every single time you prescribe tretinoin, you know, you go search for that letter that you already wrote and you rewrite it again. That's a perfect example. What else? Go ahead. Oh, my gosh, Accutane, sorry for shouting. I couldn't agree with you more. You know, and even like the things that they have to, even the follow-ups for Accutane, you know, there's a thousand things that this patient has to attest to every single month, like definitely automation could help us. Anything else? Go ahead. Oh, my gosh, triaging the inbox. This is such a good one, and I love that you said that because there's actually so much work in radiology on this. So you know, radiology, they have this inbox of all the cases that came in, and there's so many algorithms that now are saying like, okay, I automatically found a stroke. I'm going to move that to the top. I'm not going to tell you that there is a stroke, but I just want you to look at that scan before this other scan. And you know, some of my patients send me holiday cards. It's so nice, but that does not have to be at the top of my inbasket. You know, I probably could respond to that a week later. So that's a great point. You know, it's so interesting that all of you have come up with tasks, and none of them are the hottest things that have come out of DERM-AI research. And so that's one of the reasons I love being at this meeting. You know, most of the DERM-AI research to date has been around images. How do you analyze images? How do you diagnose skin cancer? How do you monitor lesions? And those things are important. I think they totally would help us, but they're not addressing the pain points of our practice, and that's where we can really have an influence on the direction of where this goes. So I actually had to put inbaskets here because I agree with you. The other thing I thought about was scribes. You know, documentation is such a challenging thing, and support for no documentation is really important. So I'm going to take us through, you know, when we think about all of these automation tasks, and we think, you know, I really want an AI model to help me with this task. How do I think about implementing that? Either buying something that does that for me, or developing one from scratch, or working with a computer scientist to develop one from scratch. What do I do? And hopefully I can integrate some of these suggestions. So the first thing we already did, this is one of my favorite papers on sort of the big picture, how do we do this, need finding. The most important thing is to address a true clinical need. How do we help our patients? How do we help ourselves? We can develop so many fancy engineering tools. I come from bioengineering, so I feel like I can say some things about engineers that are challenging, you know, and sometimes you develop an amazing tool, and you're kind of in search, you're a hammer in search of a nail, as opposed to being really coming from the need. Okay, so we talked about need finding, but I'll tell you again, true clinical need. Then the other thing is, you know, with these models, we need to determine where the model is going to fit in our actions. You know, I really loved that in-basket triage idea, because this is just going to triage your in-basket, and then that's going to determine the first thing you need, but you're still going to be the doctor who decides what to do with that information. And then an output-action pair is the same thing, like if you have a draft prior authorization, what are you going to do? What is the workflow in your clinic that's going to get that prior authorization reviewed, edited, sent, exactly there, and then what are the ideal workflows? So one of the challenges that we have is that our current systems do not allow seamless integration of AI models into them. You know, we don't have really easy access to labeled images. You know, if radiologists, you know, any CT stroke can be fed immediately to the CT stroke algorithm and sent back to us. Our photos are not labeled for their intent. Our in-basket messages don't usually have structured fields that say, you know, I'm worried about this lesion that an algorithm could easily pull. You know, we have to redesign our entire structure to actually be able to deploy those workflows. We're going to talk about that some more too. Okay, so you know, the fundamental thing about building an AI model is the data, and I think we're going to spend a lot of time on that today. It's a huge, huge open area in dermatology. How do we ethically obtain the data that we need to answer the questions that we have? And what do we do if we don't have that data? How do we assess the completeness of our data? And I'm going to talk a little bit about that too. This is one of my favorite comics about this. So, okay, so you have this pile of data, usually, right, this comes from your EMR, it's kind of like happenstance of whatever you have, or it's a public data set that is whatever the curator decided to make, and then, you know, yep, you just put this data into this pile of linear algebra, sorry, Rob, for oversimplifying all of what you said, what if the answers are wrong? Right? Okay, so maybe we just like mix it again and see what else happens. That's probably not the data-related answer that we need. You know, if the answers are wrong, we need to start thinking again about what is the data that we used, and how do we make it better? So I really want to make this open to two major use cases for us. One is, you know, you're in an academic center, and you want to work with someone, and you want to build your own model specific to your own use case, but I think most of us are going to actually experience that someone else has built this model, and someone wants you to integrate it into your practice, someone wants you to buy it, someone wants you to think about your return on investment, and so then we need to actually still ask those questions about data, even if we aren't the ones who did the data curation in the first place. So what data do you have access to? Multisource data is really important. We heard already that models will really, really greedily learn any information that they're given, so they will overfit to whatever situation they're trained on, and so multisource data from other places, other clinical settings, it's going to be really important to make sure that they don't greedily learn something that's wrong. If using someone else's model, we still need them to transparently share the data that was used for training, and, you know, we recently did a study of the apps on the App Store to see, you know, what kind of data transparency is there, and right now, there really isn't much. The models are not being required to tell you what data they used for training, anything about how that data was collected, curated, where it came from, and that's something that we as the end users need to demand. And then we also need to understand, you know, gold standard labels. I think for situations like in-basket triage, prior authorization, we really don't understand how we're going to even evaluate if the model is doing a good job, and that's something that, again, we as clinicians get to say, you know, if you make my in-basket five minutes faster or if I feel like I did a better job seeing the patients, maybe that's sufficient, but that doesn't exist yet, and that's where we really have a voice. We talked a lot about public data sets already. I can sort of skip to this, but if you are developing your own model, there are public data sets that exist. This is a skin cancer data set, and you can contribute to the field by curating data in your own clinical practice setting and making it public. We also have worked with this data nutrition project to simplify the labeling, so when we think about asking technology that is trying to be deployed in our clinic, you know, tell us about the data that you used for training, they might say, you know, that's proprietary and we really can't share that, and I get it. You know, if they used 10,000 images, I also don't want to sit through the 10,000 images. I want something simple that will help me understand if this is going to work for me, and so this is just an example of a nutrition label that might help us. Do we have a pointer? Oh, okay, amazing. Thank you so much. So here you have some examples of, like, how a data set might be able to be labeled, you know, the intended use cases. This is just for identifying melanoma in lesion images, so if someone is trying to tell you that they're going to do in-basket triage on this, you're like, okay, please don't use this data set for this purpose. It might tell you, you know, this is about humans. It includes subpopulations, so that might be really good for assessing a model for bias, and then this nutrition label, which I like, can tell you a little bit about whether it feels complete for that use case or if there's mitigations that you have to do around potential harms. Okay, so, you know, we figured out our need, right? We are the clinicians. We said this is what we really need in our clinical practice. We identified some, potentially some vendors. We asked them already about the data that they used. We're feeling really excited, like this feels like a really good match, so what other questions do we ask? So, we asked them about the model selection, and, you know, we already learned about some of the frameworks that they can use. I really don't expect us to be able to say, like, I prefer that you use reinforcement learning versus supervised learning. You know, that is a decision that the engineers have to make, but we do want them to match their use cases. But we do want them to match their use case when they develop their model to the use case that we need, and that's really important because otherwise they might not work as well as we think they're going to work. What are the risks versus benefits? So, this is the major question. How are we going to predict and measure the potential harms versus the potential benefits to us? And minimizing and using bias mitigation strategies, this is critical and it's a huge problem in dermatology, and we talked about this. Heterogeneous data. Oh, well, the models will always work the best if they are defined narrowly and use a very narrow data set, but that's not going to mimic our clinical practice, and that's something that, again, as advocates for our patients and as advocates for the healthcare system, we need to say, you know, our patients don't all look the same. They don't all have the same disease. They don't all have the same problems or questions, and we need to make sure that it works for all of them and that we're not being unfair when we're using these models. Okay, you guys are doing so great. You know, we're almost about to buy an AI model for our practice to help us with our inbox. So, what do we do after that? You know, we're not done, okay? We still need to think about, first, the structures. How are we going to deploy them into our practice? And, two, how are we going to make sure that they continue to work forever or at least for some time? So, we talked about this, too. What is the role of physicians? How are we going to actually use the model? How is it going to make our life better again? You know, it's so easy to deploy a system that requires us to do 5,000 more clicks or 10,000 more things, and we absolutely don't want that to be the situation that we do when we deploy this model. It needs to actually work in practice. So, we need to actually pilot test directly that that's working with the physician, and we need to monitor the performance over time. And, again, like, please do not let that be you personally reviewing the last 100 cases and making sure that it actually worked. This needs to be an automated way to say, over time, this is working 90% of the time, and now it's 80%, and when am I going to say this doesn't work anymore? It's a very common thing for AI models that there is drift in the data that they use. So, you know, your patient population may naturally age. That's something that will happen to all of us over time. And so, that questions that they ask in their inbox might be different. The prevalence of skin cancer might increase because our population is aging. So, it's just a natural example. And so, the question answering or the prior ops that are specifically around acne, and now none of our patients have acne anymore, it's going to change. And we need to make sure that we're automating that process because, again, we don't have any time. This is supposed to make our life better. And so, I really want us to keep an eye on that. You know, we recently talked about how you develop a checklist for image-based AI in dermatology. You all don't need to memorize this paper, but I do think there is a big role for us to work together to say, what are the requirements for our systems? How do we better advocate for our patients? How do we look at the apps on the App Store and say, look, there is no transparency around what data was used for training? Another big issue on the App Store and with other types of large language models is, how are they using our data? You know, if I upload a picture or I ask a question about a prior op, you know, I'm saying, like, write me a prior op for a 30-year-old with acne. You know, are they keeping that data? Are they going to use it for their own retraining? There's not a lot of requirements around the transparency of data use and privacy. And that's another place where we need to be really careful and really thoughtful. And that's it. And I'm so excited for the rest of this session and the discussions that we're going to have. Thank you so much. Thank you so much. Thank you.

AI integration

dermatology practices

workflow efficiency

transparency

collaboration