Generative AI in Medicine Opportunities and Challenges - AI PM Community Session #34

Key Takeaways

talking specifically about my own Journey again I was coming from a background that is not so common for product managers basically I was a biomedical researcher what what really helped me about you know getting hands on and being that person before becoming actually getting the job for me it was all about okay this is the theory how can I come up with the way that I actually implement [Music] it hello everyone super excited to be here uh for some of you basically it's it's Saturday so thanks for spending your Saturday with me so this is maram RoR I'm a a product manager with Amazon AGI I'd like to talk a little bit about Genera Ai and healthare today opportunities and challenges a little bit about my background I came to the US 2012 to get my graduate degree in biomedical engineering life at the time looked like like being in the lab with pipet PhD um data Gathering dealing with clinical uh trials so never thought um I find my uh way where I am today but what changed my past at the time I started taking online courses in machine learning I remember 12 2013 taking andw um machine learning so I would go to work to operating rooms in the morning and taking online courses at night it was so fascinating that I naturally started kind of connecting what I learned last night and kind of figuring out okay this is happening in the field in operating room how can I take that learning and basically solve something for patient so that um kind of led to starting my uh first um Healthcare AI um startup um and from there my path open to product managment and I'm now product manager with Amazon AGI um just before we start just a quick note the views Express are my own and not the views of my employer so moving to the next so as setting the agenda what we are going to cover today is basically again talking about connecting learnings to practice it's always fascinating to me um and taking all the advancement research and theories and kind of finding the so what for it and applying it to to practice so hopefully today I um I'm going to try to talk about the rising opportunities in healthcare using large language models um regardless of which industry you're you're in or which llm product you're focusing on hopefully some of the learnings are going to be applicable for you as well and the second point I'm trying to accomplish for some of us product managers at some point or will kind of think about this question of whether um how and and when to specialize especially in the basically ml product management I'm going to give a glimpse of what it means to specialize as a ml product management in healthc care field a spoiler alert that's going to include reading a lot of um scientific papers you're going to see throughout my presentation uh I've cited papers from nature science Jama which are the basically um um scientific permanent scientific papers in in healthcare so includes a lot of uh uh reading about um research especially with with patient and clinical data H last but not least uh I'm going to shed some light on a startup landscape in healthcare AI um we know it's a interesting path for some of product managers to go into entrepreneurship and for those of you interested maybe I can try to give you some starter point to see whether Healthcare is for you or or not so what we will not cover we will not go into technical details of large language models um we uh please go see Community session number 33 on YouTube it was a really good one it's a recent one and also we are not going to cover fundamentals of a generative AI product management again another good session number 32 that you can watch on YouTube to be able to understand those terms but it's important to set the stage I'm going to quickly talk about this um most of you already know these but just making sure I think no um machine learning presentation becomes machine learning presentation without these circles so we have machine learning we have deep learning learning subset of machine learning and we have generated AI subset of deep learning machine learning has been around for um for since 1980 for ages it's not new of course the Advent of you know Computing more resources going into um or and more brain going into this field we are going to we are seeing more and more advancement in this field even large language models have been around um for for years now but with chat GPT great example of finding product Market fit um it basically marketed to individuals everybody regardless of background they were able to understand the value of it and being able to use it so that's why we saw more and more basically a spotlight on llms so quickly generative AI is about um um an artificial intelligence capable of generating either it's text image other data types um and most of them uh today are basically based on uh large language models and foundational models uh llms are deep learning models and that consume train on massive data sets we will talk about what massive um means again making sure we and stay um it's it's subjective what it means massive so we are going to uh touch on that uh to create um the a new combination of text that mimic natural language based on the training data foundational models large machine learning models pre-trained with the intention that they are going to be fine-tuned for a specific tasks so putting that aside talking about massive what that means uh we have um multiple llms and we are seeing every week more and more um coming to um existence as you see parameters count is going into the the T territory trillions but one thing to um call out is although we talk about size a lot when it comes to llms and size of the parameters number of parameters is just one factor a lot of other factors such as you know training data or fine-tuning protocols or architecture impacts the performance of the models for example llama very a very powerful one only has 65 billion parameters so um with with that in mind let's go to healthcare so I almost removed this slide but there is one point that I want to um it's interesting I want to talk about it so before 1920 there weren't any health records so not even papers from 1920 to 1960 um it was basically paper based medical record electronic health record um came into a spotlight 1960 and government stepped in 1970 at the time it was hotly debated topic super Innovative um and by 2016 96% of United States already had EHR so it's the um semi saturated in market today it's not Innovative it's kind of like a du it's it's there but it used to be used to be um Innovative and today um ehrs are enabler of llms which is innovation today so what make me thinking what is that ni thing post llm and applications that llms and all the use cases are going to be just enablers so some good good um question to sometimes uh think about so a little bit deeper just painting the overall picture of llms in healthcare when we talk about Healthcare we usually can think about it in three separate basically fields of clinical application research application and educational application this graph is from um nature I cited the paper here I highly recommend going reading this paper if you're interested to dive deeper into okay what are the llms the specific llms being used in which of these um categories but just just um summarizing um if we talk about limitations which are super important in medicine it's about lack of accuracy recency transparency and ethical concerns um again we have these issues in many different fields but imagine we are dealing with the lives of people so stake is higher although these limitations I'm expecting very soon seeing um llms implemented in practice of course in the beginning it's going to start with tasks um that are basically with the stakes being lower um not using personal data or a specialist knowledge is is not directly needed so a lot of these are being basically uh dictated by policy makers it's not just about the the performance or um performance of the models but these limitations for good reasons um um we we we should uh should guide us to put some guard Trails where and when developing um llms um in healthcare again that's a good uh um paper for those of you interested to dive deeper um going into application so um again a lot of different use cases exist um a lot of research is being done on different fields but the ones that I picked um I'm really fascinated passionate about the clinical aspect of it you know as product managers it's really important to have empathy for for users and it's just super easy to stay empathetic um um to individuals who are already sick on goinging pain so that is the field that I feel as a as a product manager you can really see the impact of the work here you're you're you're doing and it's at least for me super motivative so the use cases I picked are focused on basically clinical aspect of healthcare the first one uh is basically virtual nurses or virtual assistants talking directly with with end users with individuals um this is super important this is not a New Concept um at least at the time that I was going through pitch decks and and my startup and one out of three or two out of three was basically one of these but suddenly we are seeing uptic of this specific use case I'm going to talk about it later in the presentation um just because llms can directly impact um and bring value to this specific use case so today just only in in the US developed country 60 million adults are dealing with two or more chronic diseases and our re resources can only support 20% so what happens to the remaining 80% and imagine this is way bigger um when it comes to developing countries so one interesting research I cited here looked at 195 randomly drawn patient question on the social media expert went in blinded to whether chatbot responded or physician responded rated responses on two fronts of quality of rating quality and empathy and we see that at least in this research chatbot beat Physicians so definitely there is an opportunity to explore there the Second Use case I think is super important is um predicting Adverse Events um I cited one paper here um researchers at Queen University um by and used basically um 7 100,000 structure Radiology reports to successfully predicting cancer progression pars um again why I cited the paper down there in case you're interested to dive deeper but but another factor is that imagine stakeholders in the field of Health Healthcare um we have medical records but there are some nonmedical information that are super um important that are not being leverage used today for example we have a stakeholders such as speech therapists or social workers and one Speaking of Health Equity one a factor that impacts the well-being of or outcome of a procedure for for patients is and basically socioeconomic factor or how much support this person has if we let him go home who's going to take care of him and that has a huge basic has a huge question that we are not answering today imagining llms going into those non-medical recourse uh notes and being able to suggest that okay this person requires additional care for us to be able to kind of support um this the well-being longer term so again huge impact there two more use cases conversational AI Diagnostics this is compared to the first use case I talked about this is about um um conversation between the Physicians um healthcare provider and jackpot um one one pain Point um that's that's very highlighted is basically um um resource not being enough in healthcare um time saving is huge benefit for for Physicians so imagine basically and chatbot being able to assist doctors in diagnosis I put the example here I know that it's not easy to read it's it's not important basically it's a conversation between chat but in a very technical language with a doctor about basically certain um reports or um or images um the last one that I want to talk about is drug Discovery um just giving you some some idea of how a time consuming and or resource consuming developing one drug is Imagine It Takes a decade um years for sure to develop a new drug and it ranges the cost between three 40 million to 2.8 billion just for one drug um so time consuming resource consuming um I'm not suggesting that llms are going to come and cut the cost by significant now we have you know regulations in place we have Food and Drug Administration for some good reason we have a process implemented that I'm not expecting to change anytime soon but imagine just on initial steps of how it's worked today is kind of like try an error of putting 3D structures what if llms go in evaluate the likelihood of various amino acids interacting effectively with the drug molecule using vast amount of training data in molecular biology and pharmacology and at least in the beginning suggesting some some hints to scientists to at least move move faster and uh H speaking of those use cases and bringing to practice it's it's always interesting to me whenever I want to get a hint on what's happening in a certain industry I go and look at the start startups recent startups what's funded basically wherever money goes usually Innovation follows H so these are the basically map of 90 plus Healthcare startups to watch um in in a different category of Imaging Diagnostics drug Discovery and genomics fitness virtual assistant Hospital decision supports and and and the rest um majority number wise is in the category of Imaging and diag notics 29% raise a combined 1.5 billion average of 50 million each just keep in mind something is coming in the next slides that I want you to remember this number $50 million through multiple rounds of uh funding um so keep that in mind I will tell you why uh 60% of these companies are based in the United States um and um China and and UK are following and the reason I put this data is here because I feel a lot of especially in healthcare highly regulated policy making is going to be super important when it comes to how fast Innovation moves and we see China catching up because of its government AI Focus development strategy we we see UK basis startup um accelerating in at least number of healthcare startups um one reason I say that's super important again data is called British citizens share their Anonymous healthc care data with the British national health services so imagine that those startups can take those Anonymous data and apply it in different applications super super useful so I I told you to remember that 50 million number one a specific one that was super exciting super interesting for me to to go and read a little bit more about it hypocritic AI raised 50 million in seed funding only it's it's a huge number and when when you look at their crunch base um Page by the way crunch base is a great resource if if you want to go and just uh explore uh what what new startups are um coming to existence so just looking at it they are not even promising something huge they saying the focus is on non- diagnostics imagine do you remember like Watson Health IBM they were forced to sell at loss because of technical difficulty and now we seeing a a small one to 10 person startup raising 50 million uh in seed funding only and it's not even groundbreaking right I I I mentioned at least at the time that I was pitching for my own startup at least two one uh or two out of three was around this concept so what's really happening um this is startup is basically focusing on um um use case is going to be explaining benefits and billing reminder and preop questions and this is this is a interesting part delivering negative test results only meaning nothing is wrong so it comes basically the question of whether it's a matter of how model is empathetic or not or it's a matter of liability or just not pissing off the customer so we don't know those details yet but it's interesting how they are marketing um they're offering the good thing that makes me excited is unlike human beings um models can speak every language and imagine that how much value it can have for developing countries um the question for me to me at least it was underwhelming and the question arises that are we there yet and looking at this this um Benchmark table they shared and basically they compared uh their model with three other llms and the Improvement ranges from 2% to somewhere around 14% And different basically accuracy area but just looking at the the absolute numbers for example 76.5% in kind of Pharmacy um accuracy in this in um um in in in that area to me at least it was underwhelming so it's exciting to see maybe uh to entrepreneurs or future entrepreneurs out there I think this is just signaling that and there are the bar is not that high for us to step in again it's just an opinion so speaking of diagnosis um I think it's very important to um also look into to where big Tech um going um one a specific one that I found super exciting is Google Deep Mind um development of Amy articulate medical intelligence Explorer um basically trained on the transcripts of 100,000 physician patient dialogues 65 CL clinician written summary of ICU notes thousands of questions from United States and medical license examination and what they did they compared the performance on six front of diagnostic accuracy patient confidence in care perceived openness and honesty empathy escalation recommendation and management plan and they are compare basically they compared Amy with PCP which is primary care physician and on all fronts basically Amy is is beating um human again it's from their website side I feel that independent studies are definitely needed in here to kind of validate some of the findings um but I I dropped the example of conversation in here basically patient is saying I have a chest pain um and accompanied by discomfort in upper stomach um the um the AI responds with I'm sorry to hear you're experiencing that asking for more details and back and forth um you see some basically it was really interesting for me uh to start with you know acknowledging um um patient and then um asking for clarifying question and being really detailed in um what is needed to understand so this is something to watch out for um and and at least made me really really excited so we can just talk about all the excitement without talking about challenges a lot of challenges that you see me talking is is shared among different Industries um for for some of you out there um regardless of what industry you're you're in um most of these all of them apply um to different fields as well but again stake is higher when it comes to healthcare we're dealing with the health H well-being of the individuals so it's extra important to to look into these limitations challenges and being able to find a way to mitigate so the first one bias and fairness again for me to kind of deliver the message of why this is super important I cited the paper here a 2019 study found that an algorithm already implemented in hospitals that decided which patient needed care falsely concluded black patients are healthier than equally sick white patients so what happened there so in in when they were developing model um the output they focused on basically they measured that Sickness by how much um basically Insurance claim was submitted so higher insur Insurance claim and patient being sicker however we know that in socioeconomic like um underserved communities they don't have insurance or have access to healthc care providers so of course their insurance bill is going to be zero to To None So tying um assist Health based on um Insurance claim was basically leading why this this already implemented uh model was was hurtful um and research showed that they went in they switch replacing um their insurance claim data with patients biological data and that reduced the bias by 84% so again um um garbage in garbage out value in value out it's super important what data we are using and we need to push for explainability um in the next slide I'm quickly going to touch on what explainability means and challenges but not going uh deep into that I cited a paper for that for those of you interested um the the other one is basically making sure we're testing validating monitoring models for any um potential bias um for example one thing we can use is counterfactual fairness the idea is that okay if you if you replace and change the gender of this person to a different gender are you getting the same results super kind of intuitive so that's one example of how um models can be evaluated the next one is cost um we know llms huge models uh require significant cost and resource um and imagine this is of course easy for big Tech uh easier for big Tech but for low resource uh communities this is a huge uh basically burden um one way is leveraging cloud computing um cloud computing have been around for for ages basically it eliminates upfront investment in Computing by paying only for the portion that you use as you go the other ones is basically leveraging more efficient models leveraging foundational models we talked about them basically using all already not starting from a scratch but already um trained models and basically building on top of that domain adaptation and basically transfer learning is one way to look into there are a lot of research out there um touching based on on this I basically H cited one um from science um in here and cost sharings through collaboration I think compared to other Industries U Healthcare is um doing fairly good in collaborations basically a lot of collaborations um basically everybody is so used to collaborations um at least in my view in in healthcare so cost sharings through collaboration is one way to mitigate the cost of these models I talked about explainability quickly touching on this one another paper cited in here which was um very useful um explainability refers to the ability to explain or present the behavioral models in human understandable terms so when it comes to patients um in order to build trust with patients um when um will be basically suggesting okay here's the course of action or treatment we are taking it's very important to explain how we got there so that's super simple that's what we mean by explainability um so when it comes to llms com compares to traditional deep learning models the scale of llms in term of parameters we looked at and training data makes understanding and interpreting decision making process more more difficult and more and computer intensive continuing on challenges uh we already are um I I almost forgot the term hallucination until basically chat GPT and llms came into picture so we are all familiar with hallucination really sometime really hard to um um recognize how accurate the output is and imagine if this again stick way higher in in healthcare and in addition to that and the complexity of medical language and context can make it even more difficult for llms to cap capture the nuances so big challenge when it comes to healthcare some ways to U mitigate it basically leveraging reinforcement learning keeping the expert in the loop one interesting multidisciplinary collaboration I saw and I basically dropped um you will see on the right side of the presentation address to GitHub it's a framework for now it's focusing on um basically multidisciplinary collaboration so how it works is that okay we we are we the subject is about cancer we're Gathering basically um um experts in that field and they analyze and they write a report llm um summarizes report gathers all the experts H it's continues iterate on on the report until Anonymous um until consensus is reached and basically that's that's the idea um the other um Avenue we should take is dynamic training of course um in in healthcare especially things findings are uh things are moving fast in sense of you know findings new data emerging uh from from clinical data so continuous updating and training of a model is is the key and the one obvious one is data privacy protecting basically sensitive personal data is important just last year around I believe June open AI announced the leakage and leak of the information which was also including some Financial um information so those are those becomes super more important when it comes to you know H sensitive data and Health Data for patients so it's important to maybe we can Implement some apis to Independent more secure applications um those are the things to uh keep in mind so uh last but not least before we go to questions it was important for me to ask um Dolly what Dolly thinks about llms in healthcare you see this is the image generated um what was interesting to me is presence of maybe too many people and maybe this is signaling to us that uh the models are not replacing humans but it's too soon to tell so we will see in future this concludes the presentation would love to open it to everyone would love to make it a discussion um not one way so um any question if anyone have answer feel free to also raise hand and make make this one a conversation thank you thank you so much maram this was a great presentation really enjoyed it um and uh I'm going to start um letting questions come in so just like one quick note that I was going to say is um for those of you that are still interested in enrolling into our upcoming cohort that basically starts next week um the link to the cohort is here uh we're extending the deadline of the application to the end of the weekend so you can still have some time um to apply and I will be kind of letting people in as soon as we can just so that um you have a chance to get in into the program and prianka welcome to the cohort we got a lot of people we have two different sessions both Saturday morning and Tuesday evening so very exciting stuff and U people are receiving their Communications for course shortly so the first question is coming in from Pua that says what do you think um the startups in healthc care are um just see are funded despite all the regulatory hurdles got it so definitely so regulations are not new in healthcare um they have been around and regardless of that we have seen a greater startups being successful IPO exiting so it's definitely a consideration for individuals going in but there are many ways to basically mitigate the risk when it comes to healthcare startups one thing that I think can can be super helpful is engaging with experts from the gecko and building a group of of um um advisers from very beginning um the course of if if you're thinking about starting your own a startup it's super important to have them H and um basically have their in advice from day one also I think spending on engaging with um um law experts in the field or regularity experts in the field um is is an investment uh and those are the things that there are some nuances for example um um this startup that I talked about focusing on non-diagnosis is signaling that they are basically trying to H ease some of the regulations on that front so so many different levels that levels that in individuals can can pull to basically do risk being a healthcare startup out there thanks so much for that answer we got one more question from juta juta is actually one of our Co cohort alumni as well so it's great to see that a lot of the cohort alumni are also attending the Comm session so she asked Can you comment on the trials and testing the AI health care products um before production grade release so basically I guess you're trying to get a sense of like what sort of Trials and testing these AI Healthcare products are doing before they actually release their products into production that's a great question so the question is basically how what are the steps before they go to yeah like how do you do trials like how do you do testing given that these are like kind of AI based and I guess they're like kind of less reliant on human trials in a traditional way have you noticed any Trends or is it to judge I haven't seen any trend on on that front but I I I want to touch based on basically collaboration aspect of a a lot of of basically proving concept and kind of proving that um an idea in in healthcare AI spaces working requires actually uh doing U some data collection involving experts for example we saw that one research of comparing chatbot with a human performance those are the things that will be done through the collaboration and and and that data collection I haven't seen any Chang in trending how um research which is is being done just solely on um you know focusing on the AI aspect of it got it thanks so much for that um so I'm going to search to LinkedIn for a second so there's a question coming from LinkedIn asking um as a medical device startup um in AI how can I explore the pathways that are available for regulatory approvals and I guess this is a start this is a question that I can imagine a lot of people that are thinking about like doing an AI project in healthcare or kind of like thinking about like how do I deal with like all the regulatory challenges and I think you touched on it briefly by mentioning uh it's a good idea to have experts in the team like people that are already familiar with that but maybe you can expand on that a little bit that would be great yeah perfect I definitely think that's an investment I know that at least in the beginning of you know start of life cycle we don't have money to to spend we we are basically Ultra Frugal but I believe that it is an investment especially since a lot of regulations because this AI llms kind of like dropped H and and I'm expecting a lot of changes soon happening on regulations front um so I I wouldn't even make a comment on that but I urge everyone to engage with experts in in a regular regulatory space um from the day one of even thinking about about your [Music] startup got it thanks so much um so a couple more questions so chali uh Chell is also attending our next cohort so first of all welcome to the cohort and second of all um thank you for your questions and question is what kind of metrics are generally tracked for llms in healthcare it's a good question um so I think it it really depends on the design of the study and and outcome but I would say for example to the one the ones that we looked at um um accuracy is is one um and basically it really depends on what what they are they are measuring um but what I suggest is basically a lot of those um papers that I cited have details of okay here here was the study design here is the approach taken and here is basically what was the success met um each of these basically uh algorithms or human we kind of compared against and I I would say that would be the best next step for um for individuals to deep uh dive deeper into um those studies got it um I'll actually ask one follow-up question on that on the accuracy side especially in the world of Health Care um what's the best way to actually track and measure accuracy like what sort of metrics like how would you actually go about measuring whether or not you have high versus you know low accuracy got it so I think it's it's very similar to um kind of evaluation of ml models in in general but one example um imagine that the model is about you know uh predicting if the tumor exist or not like binary 01 so so basically um one way is we are basically we have labeled data um and we measure the accuracy based on okay how the model is U performing in basically um predicting 01 whether two more exist or not got it okay that makes sense thanks so much um so the next question is from shasa um he's asking data privacy is a big deal in healthcare so what kind of data are being used for basically training or like test data good question uh so um for example for MRI images um all redact so no information no um personally identifiable information is feed into models um for some of these chat Bots those are even not um um even identifiable information from the gecko I was was reading an article and one of the things kind of preparing data in a preparation data they found a way to automatically redact or remove personally identifiable um information from the recourse that was also really interesting for me they they um basically um U removed patient information from 70,000 reports automatically they didn't go into detail of how they did it but that's basically one way of preparing and kind of reducing the risk when it comes to um privacy of data got it thanks for that so we got more questions coming in so this question is from Alice on LinkedIn really good question and uh I have part of the answer I'll provide after yours but um so the question is hi Mariam I'm a PM in attech startups I'm currently upscaling my MLA skills and one enter the medical domain I've been known to have strong soft skills and Technical knowledge but I don't have programming or AI ml background um so what's the best way to Pivot and um he says that her concern is not so much about the domain pivot part she thinks that she can kind of learn that part but like how does she go for product roles um in AI without former education in data or ml or AI Vision do you want to take that one I feel that you're yeah sure of course area have you heard of this court cohort at PM exercises called AI product management cohort we've honestly like designed that exactly for um the audience that's asking this question that you get to get your hands dirty even though you might not have a lot of technical background and um the program is Mentor really give you that foundational knowledge you also get to build a product during the cohort so um and we'll provide you with the guidance that you need for that so we think that hopefully that would be a good solution but my general opinion to be honest in anything that you want to do is like be that individual before you get hired for that role so if you're trying to do the other way you're trying to become a PM in attech um go build the product in attech and release it so that you can actually learn through the whole experience and you can demo it and present it and sometimes you can do it through a cohort sometimes you can do it yourself but that's the best way to pass it back to you I would like to plus one on on that one and talking specifically about my own Journey again I was coming from a background that is is not so common for product managers basically I was a um biomedical researchers researcher what what really helped me um to your point vision is about you know getting hands on and being that person before becoming actually getting the job for me it was all about okay this this is the theory how can I come up with the way that I actually implement it um so Hands-On learning super um super important and being able to actually develop something become a builder before become a product manager is is really key so it's good to hear um Vision that the cohort individuals are going through basically Project based learning good to hear thank you um so lot of question is still coming in so this is an interesting one so and this is something that I've also seen to be honest even myself as a user especially given the hallucinations that we're seeing like sometimes wondering can I rely on the answer that is provided to me so the question is what are some of the steps that a PM can potentially take to have end users trust using the AI Healthcare products like basically the question is more like from an adoption perspective uh what can you do to um help businesses Drive adoption of the users and kind of trust that their AI responses are reliable all right that's that's a really good one um I think it's it applies to every uh industry and every um basically product management um a work stream or project but especially important in healthcare I think um being it able to um generate Rich data um so again collaborating with Scientists to come up with some research design really evaluating the model and taking those data and translating it in a super easy way for individuals to understand is is one way to really say that okay it's not that just us telling you this is a good good thing you should start using but this is the data and here is we basically we proving it with data rather than us telling you what to do so I think that that is the key again collecting enough data um um building confidence um using um back to with data and being able to translate those uh information in an easy to understand or applicable way for customers to to start trusting you um um with with the product especially in this highrisk area got it thanks so much for that answer um I think we finally caught up on all the questions let me see if there's anything else that's coming in LinkedIn sometimes has a little bit of a delay let me just like take one final look yeah we're good um so I wanted to you know thank you one more time thank you so much uh for being here and really appreciate it um I just wanted to kind of take um one or two minutes to invite if any of the previous cohort participants are interested uh in sharing their experiences um with the cohort I would really appreciate it I can see um there are a few people here um maybe either Jan if you're interested or jna or three I can see you Moen you're also here as well anybody who's interested please just go ahead share your thoughts I'm kind of putting you guys on the spot if you don't feel comfortable it's okay too but uh if you are go ahead and share your thoughts okay I can see perfect thank you I'm happy to share some thoughts um so my background was more on the design side uh the ux and digital design Building Products um so I'm not an engineer and I came into the course um you know a little apprehensive about how technical it would be but I will say that um it was a great experience we started from from the ground up and uh you know beon and and mahash were very patient in answering any questions we had and the Project based learning was really great and I feel like I now have a really uh solid foundation and working on my own uh side project as well and and a lot of what I learned is being applied so I highly recommend it awesome thank you so much that very nice of you and um I'm really glad that like in our community now a year into it like we have people coming from an industry sharing their unique perspectives um we welcome people from different backgrounds um you know engineering non-engineering design medical um Finance um highly technical consumer Tech so this is really a place where you can kind of just like come in every Saturday and learn so I want to just like kind of leave it at that and thank you so much Mariam again for coming in and Jan thank you so much for the shout out really appreciate it we hope to have you Mariam here again at some point and I will stay in touch but um with that said um I think we can wrap it up is there anything you wanted to kind of share at the end Mariam no thanks for having me I enjoyed this session and best of luck everybody in your Journeys awesome thank you so question go ahead go ahead sure go ahead yeah I was wondering you know for startups right how do they get access to Patient data and medical data right because that's the key to developing whatever you want exactly and and that's why I called out UK as startups having a basically upper hand because of access to um patient information available um publicly as a kind of public data set um if we don't have that again collaboration with um medical institutions University is is a pth that a lot of startups are taking uh I have a question actually it's a very interesting discussion so far uh so um are they also using synthetic data for this kind of um cases yeah any and like how they are gener in because we have been working some time back with x-rays so it becomes very difficult to get those volume yep yep exactly and I didn't touch on it in my presentation but uh generative AI um one of the um very important benefits um individuals are realizing in healthcare is actually in U generating synthetic data uh to your point um um one of the pinpoints in data in healthare is lack of data and access to data so um um definitely uh generative AI is being leveraged for generating synthetic data and we're seeing more and more and the trend toward basically more research being done on on data space too sh thanks awesome um I get I got one more question from somebody posting so maybe given we still have like seven more minutes I'll keep going is that okay yeah sure the question is what are the major challenges that you faced as a AI startup in healthcare oh gosh that's a really good one um so we're talking about you know 2015 um it it was it it AI wasn't this um kind of mainstream at the time um of course um they were around for for ages but um it was wasn't like today that everybody was was talking about it so one of the major challenges that I felt at least at the time um is basically whether the industry is prepared to embrace the that that technology and kind of openness to that technology was something that okay do you need to take extra step to educate your customers whether they need it in the first place or they realize the need and you don't need to spend time you know um convincing them that they need it um so we starting basically on a higher higher ground so today I think it might be easier because a lot of basically um um depending on what your startup is but individuals institutions um are realizing importance of AI at the time the biggest challenge I had um was basically convincing um at the time or main uh um customer segment was uh pharmaceutical companies focusing on clinical trial data management is basically the biggest challenge was convincing them that they they needed in the in the first place so I would say that was the biggest challenge and uh doing a startup is also a challenge on itself in so many different areas so kudos to you for focusing on that and I think the other experience that I've had in the startup is like you learned so much that it's almost like an intense um exercise like after that everything else seems pretty easy so highly recommend it to anybody who's like thinking about breaking in to join a startup or do their own startup it might seem like a you know High failure rate but it's really I think you should say it's High Learning rate you have very very good learnings during those few years so um it's almost like it yeah go ahead it's it's interesting what you mentioned on the learning aspect it's so spot on because I started my recommendation saying you know engage with those experts from day one because that was exactly one of the learnings I I learned back um that basic it's it's super important to not be frugal when it comes to basically um knowledge expertise so it's interesting you touched on that point definitely um intense way of learning um and and and growing for sure I also recommend highly recommend entrepreneurship thank you thank you for that okay great I think uh we're pretty much right on time couple minutes early um but thank you so much everyone for joining in I think we're gonna um start wrapping the call and a lot of people are thank youing you Mariam as well and Linkedin also a lot of people are sending very positive messages so thank you so much again for coming and uh for those of you um that are interested next week we're going to be here again with one of the um alumni of our AI product manager cohort um Mahesh not the Mahesh that you guys all know the other Mahesh who is a part of our cohort 4 or three I believe um he's going to be um going over one of the projects that he's been working on so that's very interesting um I think we're going to share more information on it on LinkedIn the event is coming up on LinkedIn over the next day or two um make sure that you attend the next one I think it's going to be very valuable for everybody so thanks so much again and uh have a good weekend bye