76. Transforming Healthcare with AI

 I'm Dr. Myrdalis Diaz-Ramirez. Welcome to Episode #76!

Today we have a wonderful guest, Dr. Rachel Draelos from CyDoc and Glass Box! We are going to be talking about topics such as the current use of artificial intelligence in medicine, how it can potentially be used in the future, and Dr. Draelos' personal experiences with artificial intelligence and machine learning. Are you ready? Let's go!

Engage with Dr. Draelos' resources here:

Glassbox: https://glassboxmedicine.com/
Linkedin: https://www.linkedin.com/in/rachel-draelos-md-phd-2754434a/
CyDoc Website: https://www.cydoc.ai/

Follow Dr. Myrdalis Diaz at these links: 

• Website: drmyrdalisdiaz.com
• Podcast: Design Your Physician Life
• Linkedin: drmyrdalis
• Facebook: myrdalisdiaz
• Instagram: drmyrdalisdiaz

Transcript

76-Draelos-Transforming Healthcare with AI
===

[00:00:00] Myrdalis Diaz-Ramirez: Hey guys, welcome to this episode of design your physician life. We have today, Dr. Rachel Dralis and welcome. Thank you for being 

[00:00:57] Rachel Draelos: here today. Thank you so much for [00:01:00] having me. 

[00:01:01] Myrdalis Diaz-Ramirez: So, for those of you who don't know, she is the CEO and founder of PsyDoc, which is a health tech startup creating an AI powered electronic medical record.

[00:01:11] Myrdalis Diaz-Ramirez: So, we all want to listen about this, to hear about this, and learn a lot about it. PsyDoc's first product is a smart patient intake form that generates medical notes and allows... Doctors to finish notes 3 times faster. Who doesn't want to do that? Say that he's launching a pilot program later this year in which participating practice receive a 2 month free trial.

[00:01:30] Myrdalis Diaz-Ramirez: So if you want that, please go and sign up, look it up in our notes here for the show. For her undergraduate education, she went. 1 year early with honors from she graduated 1 year early with honors from Cornell University. She had a BA in biological sciences. Then she became the 1st person to graduate from Duke University with an MD and PhD in computer science and her research focuses on artificial intelligence methods.

[00:01:56] Myrdalis Diaz-Ramirez: Development for medical applications. She has published [00:02:00] research in machine learning, computer vision, and natural language processing for AI applications that range from identifying abnormalities in chest CT scans to medical concepts, tagging in notes, mutation, pathology, prediction from genetic data, and hospital admission prediction from EMR data.

[00:02:17] Myrdalis Diaz-Ramirez: Dr. Dralos is also an avid science communicator and writer of the blog, Glass. Box medicine, which covers topics in healthcare and AI. And I'm super, super, super excited to have you here today. 

[00:02:30] Rachel Draelos: Thank you so much. And thank you for the introduction. Very happy to be here. 

[00:02:35] Myrdalis Diaz-Ramirez: Well, you know, Would you have thought when you went into into your MD PhD that you would be living in this era that we are living in 2023 regarding artificial intelligence?

[00:02:46] Rachel Draelos: I definitely would not have predicted it. When I started the MD PhD program, I was actually thinking I was gonna end up doing more computational genetics type of work and through a series of events at the beginning of [00:03:00] graduate school, I ended up working in ai. And even when I started my PhD in AI, I would not have predicted how widespread and popular AI has become.

[00:03:09] Rachel Draelos: You 

[00:03:09] Myrdalis Diaz-Ramirez: know, I have chills right now because I'm, I'm a little bit of a, a geek myself in college. I taught C and Pascal. And then I was in an era. Probably earlier than your era where we were at MIT for a summer when email was created and we were like in the computer lab, they're sending messages to each other and we could not believe our eyes like, okay, come here, sit here and then we would like write an email right there and then look to the computer next door.

[00:03:40] Myrdalis Diaz-Ramirez: It's like, did you get it? Did you get it? It's like, yes. Like we're all excited, like we couldn't believe that we were receiving and sending emails and we're already programming, you know, before that. And then my life took a different turn. I, I, I, right now I don't remember how to program anything on C or Pascal, but we always had that you know, a technological [00:04:00] curiosity, let's call it like that.

[00:04:01] Myrdalis Diaz-Ramirez: And in your case, it's paying off right now. It's amazing. 

[00:04:05] Rachel Draelos: Well, thank you. It's definitely a really exciting time to be an AI. And I, because I'm also really passionate about healthcare, I think that the intersection is it's a great place to be right now. 


[00:04:15] Dr. Draelos' Story
---

[00:04:15] Myrdalis Diaz-Ramirez: So tell me before we go into details about, you know, how your life looks right now, who chose to be a doctor?

[00:04:22] Myrdalis Diaz-Ramirez: Was it like little Rachel or teen Rachel or who was that? 

[00:04:28] Rachel Draelos: That's a great question. It, I think it was an evolution over time. So, in my immediate family there aren't any physicians and actually out of, you know, aunts and uncles and grandparents there, there also aren't any physicians. So, it was something where I knew if I was going to go to medical school, that would be sort of different.

[00:04:45] Rachel Draelos: But I had. I've always been really interested in medicine as a, you know, human, I guess positive human interaction and I always had found it appealing idea of improving human health, since you know with [00:05:00] without our health, you know it's hard to do anything else. I, I think the real final decision came part way through college during college.

[00:05:08] Rachel Draelos: I had, you know, you get exposed to so many new things. And so it felt like, you know, every other week I had a new career idea, but by the, by the end of college, I had really solidified that I definitely wanted to go into medicine. So I would say that was how the process went. 

[00:05:23] Myrdalis Diaz-Ramirez: And how did you match that with computer science?

[00:05:28] Myrdalis Diaz-Ramirez: Yeah, so that's 

[00:05:28] Rachel Draelos: also a bit of an interesting story. I, funnily enough, was pretty resistant to computer science for most of my childhood and teen years. I grew up in Redmond in Washington, where Microsoft has really bloomed. And because I grew up surrounded by People in the tech industry, I thought, well, this is just what everybody does.

[00:05:49] Rachel Draelos: Everybody's in computer science. Everybody's in software engineering. I should do something different. And I actually resisted even trying it until college. And my dad had been nagging me for years. You know, you really need to take a computer science [00:06:00] class. I really think you'll like it. I finally listened to him and was very surprised that I loved it.

[00:06:05] Rachel Draelos: I thought it was really interesting, really fun. And so by the time I graduated from college, I knew I wanted to combine medicine with computation in my career. Was he like 

[00:06:17] Myrdalis Diaz-Ramirez: a computer guy? 

[00:06:18] Rachel Draelos: Yeah, he was a software engineer, or he is the software engineer. 

[00:06:21] Myrdalis Diaz-Ramirez: Okay, that explains that part. Yes. So, how, you know, you went to Duke and you were the first person to graduate from from this.

[00:06:32] Myrdalis Diaz-Ramirez: How was the process there was there like any, any particular steps that you had to take that was not normal you know like between the medical school and the pH the computer science department was there any. Conflict was just, you know what, this was easy, I'll do this, I'll do that, and then just boom, get it done.

[00:06:47] Myrdalis Diaz-Ramirez: How was that 

[00:06:49] Rachel Draelos: process? The process ended up working out really well. I had applied to Duke through the formal MD PhD program, and when I did my medical school application cycle, I [00:07:00] actually was only applying to MD PhD programs at the time, since I knew I wanted to do a PhD. And Duke was a really unique program in that They would actually allow a PhD in a computational field.

[00:07:12] Rachel Draelos: At the time that I applied, there were a lot of MD PhD programs that had a very specific list of what kinds of PhDs you were allowed to do. And it would be things like biological sciences or pharmacology, things housed within the medical school. So Duke, Duke was very unusual in that they said you can do a PhD in any department that will take you.

[00:07:30] Rachel Draelos: So when it was time for me to start my PhD, I basically walked to the administrative offices of the computer science department and introduced myself, basically said, Hi, I'm an MD PhD student. I'm supposed to start my PhD this year. Can I join your department? They were like, Who are you? What are you doing?

[00:07:46] Rachel Draelos: So I was eventually able to, to I explained to them how the program worked and convinced them that I should join their department and then they allowed me to start my PhD there and I just, you know, went through the computer science requirements, just like [00:08:00] any other regular grad student because there wasn't any overlap at all between the CS PhD program and the medical school requirements.

[00:08:06] Rachel Draelos: So I just basically did the usual computer science PhD and then at the end of that went back and finished med school. 

[00:08:13] Myrdalis Diaz-Ramirez: So how was the order? Like you did the, the, the basic science for medicine, then went ahead and did your PhD for how long? 

[00:08:22] Rachel Draelos: My PhD was five years. And since I was at Duke, my, my path was a little bit different.

[00:08:27] Rachel Draelos: Most MD PhD programs, you have two years of MD lecture, and then you do your PhD, and then you do two years of your clerkships at the end. But since Duke already has a built in research year with their med school, it worked out really well because I basically did one year of lecture, one year of clerkships.

[00:08:43] Rachel Draelos: Then I did five years of my PhD and then I came back and I did my last year of medical school. So it worked out to be eight years even though my PhD took 

[00:08:49] Myrdalis Diaz-Ramirez: five. Oh my goodness gracious. Well, that's something. So congratulations on doing that. I'm pretty sure that you must be celebrating that you took that, you made that decision.[00:09:00] 

[00:09:00] Myrdalis Diaz-Ramirez: Nowadays that was, that's awesome. Let's move on into actually, you know, I'd like to know before we move on into what you're doing these days. is your thesis. What is your thesis on? Sure. So 

[00:09:13] Rachel Draelos: my main thesis project was using machine learning to interpret CT scans. And I focused on a couple of different subcategories of that problem.

[00:09:23] Rachel Draelos: So one angle of it was just creating a data set that could be used to train AI. And I made the rad chest CT data set, which has 36, 000 CT scans in it. I also developed a natural language processing framework that could take the radiology reports and extract basically yes or no labels for all these different abnormalities of interest.

[00:09:42] Rachel Draelos: And then I trained an actual machine learning model on the CT volumes to predict the abnormalities. From the from the CT volumes. I became really interested in explainable AI partway through my PhD. So I then went down that route and ended up developing a new explainability method that essentially will highlight the [00:10:00] pieces of the CT scan that were used to make each of the predictions.

[00:10:02] Rachel Draelos: And that's really important because as I learned during my PhD, sometimes. AI models can have really good performance for the wrong reasons. There is this, there's this great paper where they looked at a chest x ray classifier that was supposed to identify pneumonia and ended up discovering that the model was actually using things like the positioning and shape of the medical or medical tokens for sidedness.

[00:10:26] Rachel Draelos: Or post processing artifacts to, for the hospital, because for whatever reason, in their data set, the hospitals had really different pneumonia prevalences. So you have to be really careful with these models since sometimes they behave in weird ways. And that's why explainability is really important because then you get extra insight into, you know, was it, was the model actually looking at the lung nodule when it said that there was a lung nodule, or was it looking at something else?

[00:10:49] Myrdalis Diaz-Ramirez: That's super interesting because you don't think about these things, you know, as a user, right, you'll say, okay, this. This thing will do that and that's fine. But, that's a very important field that [00:11:00] is even more important for physicians to be involved in right when they're when we're doing , these AI applications, whatever it is right now.

[00:11:09] Myrdalis Diaz-Ramirez: And we'll talk about the impact that that's going to have, for example, in radiology as a whole, from your point of view. But. As physicians, you have somebody who's necessarily not involved in this new technology because it sounds like fun and it sounds like it's a good thing that can be done. Doesn't mean that's necessarily going to be accurate and it's going to give us the best results are based as you say, you know, in the correct basis of thought of train of thought, you know, like how , the algorithm is made.


[00:11:37] Future AI Use & What Radiologists Are Saying
---

[00:11:37] Myrdalis Diaz-Ramirez: So now that you know, that was some years ago when you did that thesis and right now we have quite a few applications that are working on radiology specifically. Have you and and I see that you've worked in so many other areas of AI. Have you had an opportunity to talk to the. daily radiologists, you know, the [00:12:00] one who really sees the patient, some of them, how are, how has been your interaction in terms of, you know, I've seen this happening.

[00:12:07] Myrdalis Diaz-Ramirez: This is what I think will happen. And then what they're thinking is going to happen. 

[00:12:13] Rachel Draelos: Yeah, that's a great question. I think because radiology is so image focused and there have been, there's been so much progress in image related AI recently. It's definitely a hot topic. I think people have pretty different opinions on it.

[00:12:28] Rachel Draelos: So there's, there's some people who think, you know, there's, you know, it's going to take a really long time for AI to have an impact and there's other people are saying, you know, it's right now or it's tomorrow. I, I think that it's definitely going to have some impact. Thank you. One thing I did gain a much deeper appreciation for during my PhD is just how difficult it is to actually build models that could be clinically trustworthy and have good performance even on super rare findings.

[00:12:53] Rachel Draelos: Because you need to, you end up needing to have these really massive data sets in order for any model to be able to see rare [00:13:00] findings frequently enough. And even then, There, AI has enough of a tendency to pick up on correlations that you just have to do so much validation in order to be able to trust those models.

[00:13:10] Rachel Draelos: So I hope that with everything that people are working on for radiology, that there's just a lot of radiologist involvement to make sure that as people are interested in this area, it. Fits with what you would want from a clinical perspective and you don't just end up with AI doing weird things because it, you know, it's, it's, I think everybody should approach AI as don't trust it until you've proven that it deserves to be 

[00:13:36] Myrdalis Diaz-Ramirez: trusted.

[00:13:38] Myrdalis Diaz-Ramirez: And from your point of view, you know, you've said something very important. So let's do a call to action right now to those radiologists who are listening to this. What you have to do is get involved, right? Don't be scared and don't deny it because it's present. We just have to be involved. And as you said, you can there is room for physicians to go there and really be involved in the [00:14:00] development of these tools and the regulations of these tools.

[00:14:02] Myrdalis Diaz-Ramirez: So that's the 1st call to action right there in your case. How do you see the, the tool being used in the future? Suppose that, you know what, all that validation is done, the explainability has, you know, everything's done and it's working really well. Suppose that that happens. How do you see that impacting radiology specifically?

[00:14:26] Myrdalis Diaz-Ramirez: So, in a world where it reads super well? 

[00:14:30] Rachel Draelos: Yeah, well, personally, I think there Should always be a human in the loop at, you know, at least for for the upcoming future, just because I do think it will take quite a bit of time and effort before we would have a I systems that would be so trustworthy. We could Just sort of leave them alone.

[00:14:54] Rachel Draelos: I think there'll be a long time that will need people involved. And I mean, if we had [00:15:00] a system that if we had an AI system that was extremely good at identifying a very wide range of abnormal findings and images. That in that case, you might not need quite as many radiologists, but based on my experience, I think that the current AI systems are not yet equipped to deal with unusual things.

[00:15:23] Rachel Draelos: And that's something where, you know, it's a it's a strength of human intelligence that we can identify outliers and we can identify weird things and and think about them and reason about them in ways that a lot of these models. aren't doing. So it'll be interesting to see how it plays out. I mean, I, I don't want to make any predictions of, you know, in, in this many years, this many radiologists will be affected by AI or anything like that, but I do think there will be an effect.

[00:15:49] Rachel Draelos: And I, and also to echo your point, I think it's super important for radiologists to get involved in the development of these systems now just because it is happening and physicians really do need to [00:16:00] play a role in this. And I'll also say machine learning people. Generally speaking, they love it when doctors want to collaborate with them because they don't really know much about how medicine works and they may not even understand the images that they're looking at necessarily.

[00:16:14] Rachel Draelos: So having a radiologist involved can make a really big difference in how the models are developed and how the project goes. So yeah, it's definitely a great idea to get involved. 

[00:16:24] Myrdalis Diaz-Ramirez: You know, it's funny because from our point of view, we might be Like even threatened by the presence of somebody who knows so much about computers as physicians.

[00:16:32] Myrdalis Diaz-Ramirez: And then it's the you know, it's, it goes both ways, the admiration and we shouldn't be scared of each other. We should just strive to cooperate and collaborate with, with each other and finding that space for the good of our humanity, for our patients. So that's, that's a cool point of view that you share it from both sides.

[00:16:51] Myrdalis Diaz-Ramirez: It's, you know, when you say that 


[00:16:54] AI vs. Machine Learning
---

[00:16:54] Myrdalis Diaz-Ramirez: Can you tell us the difference between artificial intelligence and machine learning? 

[00:16:58] Rachel Draelos: Absolutely. [00:17:00] So artificial intelligence is a really broad umbrella term, and it refers to any system that is in a computer that displays intelligent behavior. So within artificial intelligence, you have a subcategory that's machine learning, but then you also have different types of artificial intelligence, like expert systems, where an expert system in the simplest form could be something like If then rules that were created by an expert in an area to produce intelligent seeming behavior.

[00:17:26] Rachel Draelos: So AI is a huge umbrella. Nowadays, when most people talk about AI, most of the time they're talking about some new machine learning model that's come out. So machine learning specifically is a type of AI where. The computer learned without being explicitly programmed, and it's learning from data, meaning you have to choose your data very carefully because the type of data that you feed in is going to really affect what the model learns and what kind of relationships it learns within machine learning.

[00:17:53] Rachel Draelos: There's a bunch of different types. So something like chat GPT, which is all the rage [00:18:00] these days. That's a generative machine learning model generative, just meaning that it's producing in, in the case of chat GPT text. So it's generating text. You can also have other types of models, like a classifier that if we're taking Texas example, a classifier could basically tag text with certain labels.

[00:18:16] Rachel Draelos: So using a medical example, you can have a classifier that would take a radiology report and say, you know, nodule present or absent based on the text stuff like that. So there's, there's many different. purposes that machine learning models can have, but they all share that they learn from whatever data 

[00:18:31] Myrdalis Diaz-Ramirez: you feed them.

[00:18:33] Myrdalis Diaz-Ramirez: Thank you very much. I know that, you know, many people will have that question. There is, there's the answer. And 


[00:18:38] Dr. Draelos' AI Interests (?)
---

[00:18:38] Myrdalis Diaz-Ramirez: right now you talk about chat GPT and, and all these other like generative AI. Do you have a favorite couple of things that you might be just. That you're playing or using with in your work that is not necessarily that is not necessarily your company right now.

[00:18:54] Rachel Draelos: So I think for me, I'm, I'm very [00:19:00] interested in. Maybe this isn't a very popular viewpoint, but I'm very interested in everything that is wrong. A lot of popular models these days and trying to get more insights into it, just because I think that it's really important to be. Conservative around deployment of AI.

[00:19:16] Rachel Draelos: So if we take chat GPT as an example touchy PT is this, you know, it's an incredible technology and anyone can interact with it and, you know, have this this impressive conversations with it. It's not perfect, though, so it has been trained to sound very confident but that's not necessarily related to how accurate it is that's very good at very confidently.

[00:19:37] Rachel Draelos: Making stuff up. And it also has a lot of inbuilt biases because it was trained on essentially the internet. So whatever biases you encounter on the internet, that's in the model. You don't generally see that because they've locked it down behind the moderator API so that anything that's overtly disturbing doesn't end up making it to the end user.

[00:19:52] Rachel Draelos: But I think something I'm really interested in is mitigating bias in AI models. And also on the explainability side, [00:20:00] trying to get as much insight as possible into why they're producing the output that they're producing. Just because in any real world application, I think you need as much understanding as you can get about the models you're using to make sure you don't have unintended consequences.

[00:20:13] Myrdalis Diaz-Ramirez: That's awesome. You know, those are topics and you say like, it's not popular, but it's so vital. For these to work so vital for people to understand where these models are coming from. And yes, we use them and we have fun with them and we are really integrating them into our, into our daily workflow, but we don't necessarily know those things that you just ask.

[00:20:36] Myrdalis Diaz-Ramirez: Right? You know, like the explainability aspects and all these things, which are so important. One of the things that we see, you know, my husband, myself from the generation that we come is that we were creators. Of technology in our generation, and then the newer generation has been mainly users and as users, there's a lot of just taking in whatever is.

[00:20:59] Myrdalis Diaz-Ramirez: [00:21:00] Coming without necessarily asking about where it comes from, or or if it's really the right thing to do, or how can I fix it? You know, it's I'm given this. Let's just do it or how can I investigate, you know, can I do this or or be more creative on that point of view? 


[00:21:17] Challenges in Trusting AI
---

[00:21:17] Myrdalis Diaz-Ramirez: So talking about the generations, you have a particular train, right? You have your interests and stuff, but people in your generation and the younger generations might be more users than creators, you know, this, and just accepting things as they come. Have you seen that to be a challenge in the way that you're relating with people that you're, you're working with or associating with?

[00:21:42] Myrdalis Diaz-Ramirez: I think it's a really good 

[00:21:42] Rachel Draelos: question. I'm not sure that I personally have seen generational divide necessarily, but something that I have seen which you alluded to is many people that I've talked to have this. innate trust of AI, almost like because now we've reached a point [00:22:00] where these gender language models sound like humans, it's almost like we give them the benefit of the doubt.

[00:22:05] Rachel Draelos: We want to trust them the way that we want to trust other humans. But the downside there is that They're not human and they don't think that the way humans do they don't have the kind of reasoning capacity that people do or the embodied experience or any of it. So I think the best way to approach AI is to actually not trust it by default and always ask questions.

[00:22:26] Rachel Draelos: Actually just just last night, I had this thought HR. Departments around the country. They're going to try to do resume screening with chat GPT. And it's something I probably am going to do at some point in the next couple of weeks is generate a resume data set and then have, have it automatically evaluate.

[00:22:44] Rachel Draelos: And just last night I did like a very quick test where I took two little blurbs about, you know, such and such universities, such and such extracurricular activities. And the only things I swapped out were the gender Implication of the name and men's soccer for women's soccer [00:23:00] or something like that.

[00:23:00] Rachel Draelos: And I asked chat TV to rate the candidates and the one that sounded male got 100 and the identical one that sounded female got 70 out of 100. So I, I'm going to try to, you know, flesh, flesh this out a bit more over the next few weeks and maybe put, put up some preprint or something, but it's just, it's good to always ask questions because.

[00:23:21] Rachel Draelos: There can be weird things happening behind the scenes that are not immediately transparent. 

[00:23:26] Myrdalis Diaz-Ramirez: That's great. You know, we really have to be curious and where things are coming from. And just that little exercise, that's the same thing. That's one thing that I understand that it doesn't have is reproducibility, right?

[00:23:36] Myrdalis Diaz-Ramirez: We cannot really reproduce this exactly the same under the same circumstances. Like you asked it something, and then it can give you three different answers, three different times. And it's, it's not. The same thing all the time. So it's something to be aware of. So before we get, I'm like, I'm just to get to Psyduck.


[00:23:54] Dr. Draelos' Involvement in Special Areas
---

[00:23:54] Myrdalis Diaz-Ramirez: You did work in a few more things. And I'm very excited about because you had the hospital [00:24:00] admission prediction from the EMR, which you worked on. And you also worked on genetics, on genetic data. So tell me about those two projects. And then we'll move on to these days with Psyduck. Of 

[00:24:13] Rachel Draelos: course. So with the hospital admission prediction, I was involved in projects on that topic in two ways.

[00:24:19] Rachel Draelos: So one of them was with Dr. Michael Carey, who's at Duke, and it was specifically focused on hip fracture readmission prediction, and also hip fracture, and that was using EHR data as an input, and then also using EHR data as input. your labels in order to train a supervised model that would predict these hip fracture outcomes.

[00:24:40] Rachel Draelos: And that, and that was a really I think a good example of one way that EHR data can be leveraged for something that is, you know, it's of interest to knowing different factors that might contribute towards. Hip fracture risk. So one of the models we used was a logistic regression model and that type of model can tell you which of your inputs it'll directly tell you which of your inputs was related to your output.

[00:24:59] Rachel Draelos: So [00:25:00] you can get some insight into factors that might affect hip fracture readmission in that particular case. I also worked on an admission prediction model towards the beginning of my PhD. That one didn't end up becoming a Paper that's out there because it was part of the operations side of Duke rather than research, but that was with the group related to Medicare Assured Savings Program data set also from the EHR and insurance claims, and that was with hospital admission prediction in general, so not associated with a specific condition.

[00:25:30] Rachel Draelos: And then on the genetic side, I worked with Dr. Andrew Landstrom on a project to predict whether a particular mutation would cause disease or not. Thank you. focused specifically on genes related to inherited my, my brain is blank on the word, but genes that can cause erbidias if you have mutations of them.

[00:25:50] Rachel Draelos: So, we worked on that project which was a model that would. Basically consume information about what the actual mutation was in terms [00:26:00] of the actual change in the DNA, but then also different annotations, such as what part of the protein it was it was coding for and so on. So those were both really fun applications of machine learning.

[00:26:12] Rachel Draelos: I enjoyed working on both 

[00:26:13] Myrdalis Diaz-Ramirez: those projects. You know, there's so much opportunity for physicians to get involved in this type of of technology. It's like, endless, you know, what you're saying, you've been through so many different stages of of patient care, right from the basic. Science to, like, the, the hospital admissions and all these things, it's like, there is opportunity.

[00:26:33] Myrdalis Diaz-Ramirez: So, guys, you have you have. places where you can work and develop and work with these teams and they're anxious. It seems that they're anxious also to have physicians get involved in these situations in of development.


[00:26:45] Dr. Draelos and CyDoc
---

[00:26:45] Myrdalis Diaz-Ramirez: So let's talk about PsyDoc. How did you get to PsyDoc? That's 

[00:26:53] Rachel Draelos: a great question.

[00:26:54] Rachel Draelos: So I, I founded PsyDoc. Back in 2018, I was in the middle of the MD PhD program at that [00:27:00] point and really founded the company out of both of the aspects of my training pathway that I had experienced at that point on the MD side. I had already done a chunk of my clerkships and was really struck by how the.

[00:27:14] Rachel Draelos: EHRs from a user experience standpoint leave a lot to be desired. They can be very clunky, very frustrating to use. It can, it's not very intuitive to figure out how to do certain things. And you can end up spending so much time just trying to figure out how to make the software do something rather than just doing it.

[00:27:30] Rachel Draelos: So, on the user experience side, I thought this is, this is a huge problem. Also I was really struck by the lack of interoperability and how all these patient records are siloed away at separate EHRs. They don't talk to each other. And you can end up with situations where this, you know, a patient comes in and you aren't able to know that much about them because the record is just scattered.

[00:27:49] Rachel Draelos: Or, you know, you have a giant box of paper records that is. Huge. And no one has time to go through all that paper. So I was thinking, this is crazy. You know, this is, these are all problems that have [00:28:00] technologically already been solved. So we need to have, you know, better UI. We need to have patient records that are seamlessly shared.

[00:28:05] Rachel Draelos: And then when I started grad school, I started working on all of these machine learning projects that are using medical data. In a lot of cases, medical data that came out of the EHR, whether that's medical images, medical notes, diagnoses, procedures, medications, labs, and All of that data is actually really messy.

[00:28:22] Rachel Draelos: So just to give you one very specific example, if you take PFTs and you're looking at a giant table of PFT lab results, the units are totally non standardized. So for example, sometimes they'll have like. Capital L representing leaders, and sometimes it'll be lowercase l i t representing leaders, and sometimes it'll be the word leaders written out.

[00:28:42] Rachel Draelos: And as a human, I mean, you can look at that and say, well, I mean, I know all that means leaders, but the way computers are, they don't know all that means leaders. So you have to go in and, and clean all that data and basically standardize it. So it's all, you know, for example, capital L or something, if you want to make sure it's all the same units and you, maybe it's not even [00:29:00] the same unit.

[00:29:00] Rachel Draelos: So this is another example with like a blood glucose. If you go if you just want to say, Oh, I want blood glucose as an input to my machine learning model. Well, it turns out that there are maybe 60 or 70 different descriptors that all mean a blood glucose. And the descriptor is different because maybe it came from a different lab, or maybe they changed the name of it at some point in the path.

[00:29:20] Rachel Draelos: So you end up having to have a person go through and be like, Oh yeah, that means blood glucose. That means blood glucose. That means blood glucose. And then the units may not all be the same. So you have to convert all the units so that they're all the same units across all the patients. And it's this kind of thing I was thinking.

[00:29:33] Rachel Draelos: You know, this is a crazy amount of data cleaning that's happening. That's a huge barrier to leveraging AI within the EHR because the data is just not in a state that it can be easily consumed by these models. And then finally, after you go through all this effort where you pull the data, you clean the data, you train the model, you validate the model, you have clinical input to make sure everything is, you know, working appropriately and it's medically relevant and all that.

[00:29:54] Rachel Draelos: There's no way to plug it back in to the EMR afterwards. So it ends up being its [00:30:00] own separate little thing on its own separate little island and, you know, maybe you get the output as like a Excel spreadsheet that you email around or something, but there's no meaningful integration back into the EHR. So, I founded PsyDoc because I had this, I have this vision of an EHR that has a great user interface that is really intuitive, really easy to use.

[00:30:18] Rachel Draelos: All the patient records are unified, integrated seamlessly so you don't have to deal with these weird situations of patchy records. Also it's, the data is really clean and organized, so it's easy to develop AI models and validate them. And then finally, there's a way to plug the models back in exactly at the point in the workflow where they're needed in order to save time, all as part of the same piece of software.

[00:30:39] Rachel Draelos: So that's what I found at PsiDoc, is to work towards that vision. 

[00:30:43] Myrdalis Diaz-Ramirez: Oh my goodness right now I like my head is spinning. I was just given for a review for medical record something that's a thousand four hundred and four pages. I'm like okay. How am I going to go and most of it's going to be, you know, if it's a hospital admission, you know how [00:31:00] that goes.

[00:31:00] Myrdalis Diaz-Ramirez: It's like tons of nothingness, right? Yeah. How do you navigate through through that in a period of time? That makes sense when you have a patient in front of you right there. Exactly. Can you plug this into something, like scan it and that, that something gives you back a summary of what's happening, you know?

[00:31:19] Myrdalis Diaz-Ramirez: Yeah, yeah. And I don't have to read the 1, 404 pages. Like, it's crazy. I have other things to do with my time, 

[00:31:28] Rachel Draelos: right? Yeah, the only way that would work is there was like something out of a cartoon where then you just the pages and then you magically absorb it. But I mean, that can't happen. So yeah, I mean, technology where you can scan it in and then do, you know, image to text and then text summarization.

[00:31:43] Rachel Draelos: I mean, that that kind of technology is out there. It's just not plugged in with medicine yet. But it's totally feasible. And, you know, even having a situation where it's really easy for those 1000 pages of records. To not even need to be printed out because they can just be sent seamlessly over the internet and [00:32:00] just appear in all the right fields in the EHR systems.

[00:32:03] Rachel Draelos: You don't even need to worry about printing it out and, you know, killing all those trees. I mean, 

[00:32:07] Myrdalis Diaz-Ramirez: even better. Yes. Yes. That's awesome. So tell us how Psyduck is doing these days for physicians. 

[00:32:16] Rachel Draelos: Yeah. So, since the HR vision is really a huge project, what we've started with is something that we're excited about people being able to use now and in a compatible way with whatever systems they have, just as our starting point for our company.

[00:32:31] Rachel Draelos: So what we have built is we've built. On AI expert system that can generate different questions that a physician might want to ask a patient depending on their reason for coming in for a visit. So what we, what we have is very outpatient focused at the moment. And if someone's coming in for diabetes follow up versus if they're coming in for say new headaches or something, there'll be a different set of questions that they'll have PsyDoc intake form.

[00:32:55] Rachel Draelos: And then we can collect. That data beforehand and summarize it in the format [00:33:00] of a note. So, that has two benefits. One is that the visit itself could be more streamlined because some information has been collected up front. And then also there's less need for typing because part of the documentation is already written before the visit has started.

[00:33:14] Rachel Draelos: Awesome. 

[00:33:15] Myrdalis Diaz-Ramirez: And then which situations are you seeing it being most successful right now? 

[00:33:20] Rachel Draelos: So right now we're focusing mostly on family medicine and obstetrics and gynecology clinics but we did build in a really easy way for us to add new questionnaire content so that this really could be used in essentially any outpatient setting, even, say, a subspecialist outpatient setting since we have the ability to add in a lot more content.

[00:33:40] Rachel Draelos: Like if you're a physician, you're, you know, the world expert in this one particular condition and you also have, like, Specific things that you would want to ask your patients. We have ways to make custom questionnaires that can do the text generation specifically for your practice. So 

[00:33:53] Myrdalis Diaz-Ramirez: imagine I'm a family physician, right?

[00:33:55] Myrdalis Diaz-Ramirez: And I'm going to use PsychDoc for my for my clinic. How would [00:34:00] that work from an user point of view? Yeah, 

[00:34:04] Rachel Draelos: so that is something that we have been exploring lots of options with lately talking to different clinics. And I'll, I'll just share one of the possibilities. So, so one possibility is when you create an account with PsyDoc, we can generate a QR code that could be just put up in a waiting room.

[00:34:22] Rachel Draelos: And then a patient can come in for their appointment and scan the QR code with their phone. And then the questions will Automatically show up. They can answer the questions and then we'll generate the note behind the scenes automatically. So, it'll get sent through our secure back end to be able to basically show up on in a, in a browser, secure browser window in, on the physician's computer.

[00:34:44] Rachel Draelos: And there'll be a little, little paragraphs about their specific reasons for visiting. So that's one of the workloads that we're we're, we're going to offer. 

[00:34:53] Myrdalis Diaz-Ramirez: When you're developing this, you know, and you have to prove your concept and you have to, as you said, [00:35:00] make sure that the model is working, giving exact information.

[00:35:03] Myrdalis Diaz-Ramirez: What's a good amount of cases that you have to process before you put out this available to, to clinicians? Yeah, that 

[00:35:13] Rachel Draelos: is a great question. So, since you know, this is a conversation geared towards visions, I will actually just openly confess, we are using an expert system and not machine learning precisely because of the issues around factuality that can arise with certain machine learning systems.

[00:35:29] Rachel Draelos: So PsyDoc is built off of this expert system that is designed with medical reasoning built in in a very clear way so that we don't have to worry about, you know, if, is it going to be factually true or is the is the text generation going to be right? We, we guarantee that everything that the patient puts in is accurately represented in the text that comes out.

[00:35:51] Rachel Draelos: And we're sure about that. For machine learning the question of, you know, how much data do you need in order to trust something? That really depends on a couple of things. So one, it [00:36:00] depends on how difficult is the problem you're trying to solve. Cause if you're trying to solve a really difficult problem, chances are you're going to need a lot more data than if you're trying to solve something that's really straightforward.

[00:36:08] Rachel Draelos: Another factor is how much variability you would expect in the data set. So is it something where everybody is going to have sort of very similar looking data, or is it something where you could have a lot of outliers or a lot of like really unique kind of. So I know that's kind of a wishy washy answer, but if I had to, you know, if I had to make up some numbers, I'll say that you know, for the, for the medical imaging CT scan project that I worked on during my PhD, that was the data set of 36, 000 images and had the model had quite high performance.

[00:36:41] Rachel Draelos: If you cut the data set down to around 2000 images, then the performance definitely went down. So just ballpark, it was like. Maybe from 90 99% with the full data set to maybe like 70% with the tiny data set. There have been other projects I've worked on where the examples for, so for some of the HR data where maybe there's [00:37:00] 50, 000 patients or 100, 000 patients.

[00:37:02] Rachel Draelos: But if you cut that down to say, maybe 10, 000, the model will still learn something. It just won't do quite as well. So, anyway, with PsyDoc, we're really excited about incorporating different types of AI. And we're starting with expert systems first, because that's the most trustworthy. 

[00:37:18] Myrdalis Diaz-Ramirez: That's amazing because at the end of the day, you still need thousands and thousands of patients to be able to validate those systems.

[00:37:23] Myrdalis Diaz-Ramirez: If it's, you know, for, for machine learning models. So it's, it's a lot, it's a lot of work that is being done.


[00:37:32] Dr. Draelos' Ideas on the Potential Future of AI in Medicine
---

[00:37:32] Myrdalis Diaz-Ramirez: Definitely. So let's move into talking a little bit about the future of AI. How do you see these in, in the medical practice? And I've had these conversations with a couple of people about. You know, having a system where the person comes and you guys as a physician and the patient, there's this conversation, right?

[00:37:52] Myrdalis Diaz-Ramirez: About the problem, the current situation, defining our soap model, and then the [00:38:00] machine giving you an assessment and not only giving assessment, giving you alternatives, like treatment plans for that particular situation. In the current climate, you know, like politically and and also development and scientific development and all these and accepting of of the communities.

[00:38:19] Myrdalis Diaz-Ramirez: How do you see how far do you see that that could happen where I come, I talked to somebody, and then the machine will be there available to give us the alternatives and being used in medicine. 

[00:38:36] Rachel Draelos: That's a great question. I, I hope that is a little further out just because I think the AI needs more work before it can safely be used for making you know, diagnoses in a, in a real world setting.

[00:38:49] Rachel Draelos: I think one thing that has a lot of potential is to use AI for information retrieval. So maybe instead of the AI system, so there, there's kind of different ways that if we're [00:39:00] talking about, say, language models, there's different ways that those can relate to knowledge. So one way is that. The actual model itself in the model weights that can capture knowledge, but that's based only on whatever data it was trained on, meaning that if the model was trained up to data from 2021, it doesn't know anything beyond 2021.

[00:39:18] Rachel Draelos: What I think is a more powerful use case would be using them for essentially. Better searching and retrieval. So if you had like a large language model that was plugged into up to date or something, and then it could easily you could type in, you know, whatever question it was, you had specifically and then it could pull a whole bunch of the articles and highlight exactly the place that you wanted to look.

[00:39:40] Rachel Draelos: I think some kind of retrieval system like that could be something that first of all, I think it's more trustworthy because ultimately what you're looking at is the content of the article that's created by you. People who are experts. You know, but also that that kind of technology that capability and I does exist and I think could be very useful in terms of [00:40:00] proposing alternatives.

[00:40:01] Rachel Draelos: I, you know, I think whenever you go the route of the model has to sort of pick a diagnosis, then that probably falls in the medical device bucket and you have to go through the whole FDA approval process before you can do something like that. But it's, you know, it's, it's good that anything making it.

[00:40:16] Rachel Draelos: Diagnosis would have to be approved as a medical device, since we definitely need to have rigorous standards to that. I, I think it's a ways out. My, my sense is that I, I think that AI is hopefully most likely gonna impact things like Respond to Asian messages, for example. So if any of you use like auto complete features in your email, I actually don't because I find them distracting.

[00:40:40] Rachel Draelos: But I know some people like autocomplete email. I think there's some interest in maybe having something like that for messages or even just something that's like better sorting or triage of messages. I've heard of Autocomplete. I. I mean, used for maybe triaging medical images to, you know, identifying things that need to be right [00:41:00] earlier.

[00:41:00] Rachel Draelos: I, one of my friends Dr. Lawrence go. He also graduated from Duke me, P. C. program, and he has a startup and 1 of their offerings is a system that can catch missed pulmonary. So, they work with a large radiology company and they can identify that were missed. Yeah, it's something where physicians, I think, need to convey.

[00:41:23] Rachel Draelos: What is it that they want AI to help with? And I hope that that can steer what AI ends up getting used for in medicine. And the more that physicians can speak up about, Hey, you know, I really want AI to help me with, you know, maybe my messages, or I really want AI to kind of. Be this background feature of double checking for errors or anything like that.

[00:41:43] Rachel Draelos: That will help guide AI so that it can be used in the most productive way in healthcare. 

[00:41:49] Myrdalis Diaz-Ramirez: Yeah, that's, that's important guys. Once again, call to action involvement, right? We have to be involved in, in, in making those decisions and choices. So something out there, [00:42:00] right? It obviously depends on on what you give the machine, right?

[00:42:04] Myrdalis Diaz-Ramirez: You have to give this input so that you can obtain good output. Like, in the example that you gave earlier, where, you know, they were choosing the right answer, but it was because of the wrong, the wrong reason. So that's that's something important. 


[00:42:20] Quantum Computing
---

[00:42:20] Myrdalis Diaz-Ramirez: However, there's also other components that are going to come to affect.

[00:42:25] Myrdalis Diaz-Ramirez: Probably the speed at which these solutions are coming out and 1 of them is quantum computing. Have you do you have any ideas about how that's going to affect the the processes that we have right now, in terms of production of all this technology and development. 

[00:42:44] Rachel Draelos: Yeah quantum computing is a really interesting area.

[00:42:47] Rachel Draelos: So there were a couple of the classes I took in grad school that touched on quantum computing. It wasn't the main focus of my PhD, so I don't have as deep a knowledge of quantum computing as I do of AI. Although I do know that there [00:43:00] are people who are specifically researching ways that quantum computing could be used for new types of AI.

[00:43:05] Rachel Draelos: I think with quantum computing, you know, if we ever got to the point where there were these very large quantum computers easily available, that would have huge ramifications be, you know, in medicine, but also beyond because things that we take for granted, like being able to securely send information over the internet we have to rethink how that's currently working because, you know, all of our current encryption could be broken by a really large quantum computer fairly easily.

[00:43:29] Rachel Draelos: So that would definitely be a, yeah. A very interesting time in society as a whole. If we, if we get to the point where we have easily available large quantum computers. 

[00:43:40] Myrdalis Diaz-Ramirez: Yeah, we should get one right at the beginning of it. 

[00:43:46] Rachel Draelos: Yeah, get to see everything on the internet, transparently.

[00:43:52] Rachel Draelos: You know, we're breaking all the encryption without anyone knowing. 


[00:43:54] Dr. Draelos' Final Tips for Physicians
---

[00:43:54] Myrdalis Diaz-Ramirez: Now that we're getting to the end of our conversation, we've, you know, we've explored your career and what you think that [00:44:00] AI is. Do you have any particular Even from an entrepreneurial point of view, because right now you're, you're not only a physician or a scientist, you're an entrepreneur who's building this side dog company.

[00:44:09] Myrdalis Diaz-Ramirez: Do you have any specific tips for physicians who are listening to us regarding any of these aspects, you know, from, from the aspect of science, from the aspect of being involved, from the aspect of entrepreneurship within a world with artificial intelligence? 

[00:44:24] Rachel Draelos: Yeah, I think one. First thought that comes to mind is if you're interested in getting involved in AI or if you're interested in entrepreneurship and starting your own company, I think you should just go for it.

[00:44:36] Rachel Draelos: And both AI and entrepreneurship are things that, that I think are best to learn from, you know, from the inside out as you're involved, as you're working on something that you're really excited about. So, you know, don't, don't hold back. If it's something that you want to do, you should go for it and you'll be able to learn.

[00:44:53] Rachel Draelos: Pick it up. And you know, you can surround yourself. Another thought comes to mind. You can surround yourself with people who, you know, know more about that topic [00:45:00] than you. So if you're just getting into AI, you can find people who are already in AI and learn from them. And if you're just getting into starting your own business, you can find other people who are starting their own businesses and learn from them.

[00:45:10] Rachel Draelos: And you know, people I think are very willing to share things that they know. And also there's tons of really great resources. There's. For both entrepreneurship and AI available online. In, in the AI space in particular, there's a ton of really great courses that you can dive into and they can, you know, start from, they all start from different levels.

[00:45:30] Rachel Draelos: So there's a lot of courses out there online that start from just assuming, you know, you're just getting into it. So, yeah, I would say, you know, main, main, my main thought would be just go for it. And also, you know, feel free to reach out to me 'cause I'm always happy to share different resources that I found helpful.

[00:45:46] Myrdalis Diaz-Ramirez: So if people want to reach out to you, where, who they how can they find you? Yeah 

[00:45:50] Rachel Draelos: I think the easiest way is to just email me. So, so there's a couple ways. So one way is I have a contact page that directly emails me on my blog, [00:46:00] which is glassboxmedicine. com. Another way is you can, I can just, send my email.

[00:46:05] Rachel Draelos: Maybe that's a bad idea, but it's my first name, Rachel, R A C H E L dot last name, Dralos, D R A L O S at psydoc. ai. So C Y D O C dot A I. And that's my email. You can just send me an email. Well, 

[00:46:19] Myrdalis Diaz-Ramirez: awesome. Thank you for being here. And what's CyDoc's website? 

[00:46:24] Rachel Draelos: Our website is cydoc. ai. So C Y D O C dot A I.

[00:46:30] Myrdalis Diaz-Ramirez: Thank you for being here. This has been such a wonderful conversation. I'm so happy to have met somebody like you where we can talk about, you know, things that were Amazing and unthought of at the time that you decided to go into this route of an MD PhD with, you know, computer science. So it's congratulations on that decision, you know, and on your company and everything that you're doing also to help physicians, which you're, you're helping us right now in [00:47:00] having better experience for, for us and for our patients.

[00:47:03] Myrdalis Diaz-Ramirez: So thanks for 

[00:47:04] Rachel Draelos: being here. Well, thank you so much for having me. It's been really a great conversation. I appreciate the opportunity. 

[00:47:11] Myrdalis Diaz-Ramirez: Thanks. 


[00:47:11] Outro
---

[00:47:11] .

[00:47:11] And that was it for today. I hope that you find as much inspiration. I'm hyped. 

[00:47:16] I hope you are too. If you want to actually get a hold of Dr. Rachel Dreylos, visit her at glassboxmedicine. com. Or through her company, cy.ai. That is ccy d c.ai. We're gonna keep bringing you all these topics that are so important in these times, 

[00:47:35] of, entrepreneurship and shifting with, 

[00:47:37] the help of 

[00:47:38] artificial intelligence. 

[00:47:39] Check out our website. 

[00:47:41] At max alert dot com M. A. X. A. L. L. U. R. E. and also our YouTube channel, where you can see all these previous episodes that we've had before about entrepreneurship for physicians, helping you to reach success in your life and regain and retain and remain in control of your physician life.

[00:47:58] We'll see you then have a great time. Thanks. Bye.