AI in Healthcare: Securing what's coming

great awesome well thank you so much for the opportunity to present at b-sides i'm thrilled to be here my name is emma wilson and i'm a security consultant at accenture i've had many very exciting projects over the last few years of my career from being a security analyst implementing antivirus and firewalls to being a systems auditor and most recently i've had the privilege of being a cybersecurity project manager for one of the largest health authorities in bc so it's been a really fun time so far last summer i had a really exciting engagement with a client based in geneva where our team looked at different applications of artificial intelligence and healthcare in many different countries in the

pacific i'll just minimize my screen for a second so what i really wanted to do with all that really great information was share some of the findings on ai and healthcare with a broader audience so really that's what this presentation is about is to share with you some of the findings um again this is more my own personal research and interest on ai and healthcare i think it's really where things are going certainly there's a lot of hiccups and roadblocks with adopting healthcare ai as you can imagine funding resourcing personnel there's just so much happening that is creating some barriers and challenges to adoption of machine learning and ai and health but i also think there's a lot of

opportunity here and i think i think based on my research in asia so countries like china taiwan japan those kind of countries are adopting a lot of different health care applications things like robotics clinical decision support i think that because they have been adopting that it's really showing that this is possible it's proven in some jurisdictions um and therefore you know it's something that can certainly make its way over to canada in the future so just to clarify this is based on my own personal interest i hope you're interested as well and it's really a high level presentation looking at some of the applications some of the issues and concerns and some of the security risks

so next slide here does anybody recognize who this is kind of a trick question for you all okay can anyone tell me what this slide is about

okay i only have one monitor today so forgive me so this is lee sedol who is participating in a alphago tournament for those of you who follow ai and what's happening um alphago is an ancient chinese game and in 2016 there was a tournament where this um go champion um basically competed against alphago and alphago one and it was a hugely impactful event it really shook the um the technology and just the you know the world especially in asia and really for the first time we realized that this game which is considered to be very complex very challenging for a machine to learn lots of nuances that were considered to be something only humans could

understand was something that could be achieved and understood and competed in by an ai so that's something that was really impactful and it and it really created shock waves throughout um much of asia now this was 2016. i have some suspicion that we've forgotten a little bit of this event and just how impactful it was but in many industries this this was seen as the event that proved what ai can do if machine learning and ai have been you know used in retail and finance and insurance for many many years before this but this event because it was so public and it was sort of fun it was a game really showed folks that ai can do some

incredible tasks and and learn there are moments in the game if you watch the youtube if you're a huge nerd like i am where um this poor fellow lee sedol would just be shocked his jaw would drop and he was just completely perplexed by some of the moves the ai was making so really this showed the world that ai can learn complex tasks and can do things that people can do so you might be wondering okay emma get to the health care so ai has applications in many many industries as i'm sure you're aware finance you know insurance mining there's many many applications of ai i will be honest with you that my research has indicated to me that

a lot of ai applications are not healthcare focused right now and this is a sector that's been underserved by this new technology in general however there is also a lot of excitement in healthcare um i think because we see healthcare as this area of work that can be really impactful to society as a whole and can really improve the lives of many people many other countries not just canada have physicians shortages and are really struggling with serving large aging populations for instance china and japan i interviewed many doctors there in this engagement last summer who note that if you could have just like in it and computer science it service management tier one tier two tier three um health

care service where tier one perhaps an ai or a chat bot talks to you and says you know hey how are you feeling what are your symptoms send me some pictures and you send that in and then they determine whether to you know send you off to a to a doctor um so that's one application um and as well there's there's so many applications i think the the one that i think is most palatable if i was going to leave you with one um application that i think is palatable now um it would be radiology the reason why is there's been a history of standardized quality data sets for decades in radiology and there's a lot

of training data out there in the format of images um now images as you might know you know they can contain personal information phi however there there are standardized approaches to cataloging them and to basically linking what the images of with a label or title so really if you go into a radiologist there's this all this data is already available it's relatively easy to train ais to use that data to make recommendations and predictions and again like i said with the privacy aspect there are ways of anonymizing or de-identifying images but it's less i'm going to say it's less of a risk in the sense that the data is not as um obviously a you know a person if you're

putting tons and tons of data into a training data model it's just images and therefore it's a little bit less i suppose controversial when you start to build algorithms that are using those images to make diagnoses um patient monitoring robot assisted surgery cyber security just administration booking appointments ais that can crawl through emr electronic medical record systems to find patients that haven't been in in a while who have high risk factors and who can send alerts or text messages to those patients to say hey you haven't been in in a while and the last one there is clinical decision support or cds so that again is kind of what i mentioned with that tier 1 tier 2 is you

could have um really ais that look at data on a patient and help to determine the decision of whether they should actually go into an office to see a person that's one you know aspect of this and other aspect of this with cds is a doctor is seeing a patient and based on the records of that patient in their emr file electronic medical record system file the doctor is suggested that this patient might have high blood pressure the doctor might still be then validating that but um you can kind of see how this could save time right and save energy and resources for resource-strapped physicians and doctors and and other nurses and healthcare practitioners so

many many many applications of ai in healthcare um what is happening in canada i really want to focus on that i think that's what we're interested in today so there's been a history of digitization and digital health in canada and the last 30 years have seen you know incredible growth we all know that we're all here but really based on my research what i'd like to offer is that canada has been a little bit risk-averse and a little slower in approaching ai in healthcare and there's many reasons for that we could go into but what i really kind of want to offer is that really the last three or four years five six years as you can see on this little

timeline and i've just i've just selected a few events that i thought were interesting to you there has been really an explosion of investment from mostly public sector sources into healthcare ai now why is this clearly there's there's a seen value or perceived value of ai in in benefiting canadians and benefiting our care institutions and systems so we've had really just incredible investments in the last few years and recently as well there's been more money in the most recent budget so really looking at supporting a pan-canadian ai strategy so what else is happening in canada there's a lot of research happening in canada so we've had a lot of organizations for instance in alberta in

toronto in montreal looking at ai applications not just in healthcare but in many sectors so i have one example on this slide which is in 1993 mila was founded at the university of montreal that is a research institute that looks at ai in many different applications healthcare is one of them so for instance mila helped to develop a medical chatbot as one example so it's happening it is happening and it's something that's been really in my in my view happening more in the last five or six years um but there is a but to this and i think the the the thing that is in my mind that you can probably understand if i'll go to the next slide

is we are in my in my view based on my research lacking some some key enablers in canada in bringing um ai into the healthcare space so you might be wondering you know where is the money for this coming from you know healthcare and healthcare institutions already struggle in some respects in bringing innovative technology in so really i've sort of answered that question which is government is really really pushing for this um we're really investing in it from a canadian government perspective and certainly many universities local colleges and you know provincial governments are also investing in it so really we should be seeing i'm hoping with all this investment some really healthy new pilots um

because what's happening in canada in the last you know few few years of trying to explore ai in healthcare is a lot of research has died in this um pilot phase so you get a lot of institutions organizations universities who have great idea ideas and they pilot them and they explore some test data and then there's just a huge barrier to getting this technology to clinical applications so why is that there's there's many reasons i think a really good reason is simply that healthcare organizations are already resource stressed and it's really challenging to take the time and energy necessary to work with a university to bring in new technologies this second line of date of information here on data

is also really key so in canada we do have a lot of great big quality data sets certainly in ontario for example you can get health data there from decades and decades past so we have lots of data however is it high quality data so in some of my research from last summer when i was talking to many experts in ai many startup founders in asia one of the things they mentioned was that a lot of times data exists in very separated systems and databases and it's not being connected there might be a lack of standardization there may be fears over privacy and you know the regulatory parameters around how that data can be used

um the other thing is the data is just really messy quite often i know i'm talking to the audience who will understand when you're dealing with large data sets uh sometimes the data is not high quality data in a format that will be usable um whether it's um you know an ai solution or any other kind of solution trying to use that data um there's some huge challenges there so there's actually been some startups for instance i talked to one fellow in taiwan whose whole company is it's a healthcare ai company and they're just looking at normalizing and making the data out of emrs better quality before it'll be fed into another another ai algorithm that will look for

um some more um useful healthcare applications so to clarify there's an ai just trying to qual make the data better quality before it can be fed into different applications so i think that's something we can be wary of also happening in canada and i think in canada as well there's concerns around privacy so there's a need when looking at ai for understanding how to treat the data can we de-identify the data but i say not anonymization why is that in healthcare we need to use that data and it needs to be linked to a person in some way so for instance you might have large data sets from an emr and an electronic medical record system

and you have an ai that can just just crawl that and look for high risk patterns or issues that need to be alerted so you might want to bring in that patient who has a high risk of blood pressure so the issue is you might want to de-identify the data when the ai's crawling it but there has to be some way to re-identify patients once an issue is noted so you need to call the patient or send them an email or you know otherwise contact them to come in you have to identify them to let them know they need a specialist to look at their at their health concern um the last kind of key enabler i want

to talk about just briefly here i see i'm running out of time is culture people and skills um so key enablers of ai and healthcare really in many sectors in healthcare you have whether nurses and doctors whether they're administrative professionals whether they're you know t professionals they go to school for you know five six to ten years and then you also have these folks working in ai research and ai you know piloting and test labs who are also extremely qualified skilled people who go to school for a long time so what i'm trying to say is there's these two groups that are highly skilled and bringing them together is the challenge and the reason for that

is they're very you know busy in their own domains really you need to find people whether they're it leads cios even doctors in in many of the interviews i conducted in asia last year there's doctors who have their own ai startups i think they just don't sleep those people and that's a cultural thing right that you know working you know and having those two or three different streams of activity and really working to not just provide medical services but also looking at i.t um the other thing i'll add is in some of those countries that i was working in last year you'll have folks who went to school for medicine but were also doing a computer

science degree at the same time or they did some coding in their spare time while becoming a surgeon and that's very common in some countries and perhaps less common in canada i don't have the data on that but that's just the point i'd like to make is that translational champions are something we really need here um in canada if we want to bring more ai into healthcare so i think just to summarize what i just said um regarding funding data and sort of people is we really do need to kind of understand that a bit better in canada in the healthcare space so i think we need some trusted champions it could be you listening in

on this call you don't have to be an expert in you know machine language to you know advocate for it and learn more about it awareness is the first step gaining the trust of those who are responsible for the data getting creative thinking about what else can be done with data in an already resource-stressed system we need to find a way to pilot new technologies that is really cognizant and respectful of that in my view that really requires strong partnerships between government private sector and educational institutions so what we really need in my view is companies private sector companies that are already playing in the ai space to really whether you want to call it

pro bono or otherwise to really look at some applications of their technology that can benefit society so whether that's partnering with a research hospital or partnering with a university or a small startup we really need those those partnerships to happen to move this forward data standards and interoperability i think i already discussed that but really that's something we need to think about a lot in canada and then lastly is guardrails and guidance so evolving our regulations and standards in step with technology is necessary um and we can't just keep saying no no no patient data client data can't be used for that we really need to find a way to responsibly and respectful respectfully collect and share data in a

way that provides value and can potentially even save people's lives right some of these applications of ai can provide services and connections that not every doctor has time for there's a strong possibility that you know this technology can improve the health care that's delivered to to canadians so we're really here to talk about security i only have a few minutes left so i'll just chat for a minute or so on this people are adopting new technologies all the time we all know this without considering security and other issues so certainly senior leaders know that um as regards ai generally there's a lot of new ways to attack ai solutions um so model manipulation data poisoning there's been a lot of

recent hacks that are really exciting and interesting if you're into that on um you know like feeding feeding information into ai into alexa so alexa's listening she's always on and i'm telling her you know i need i need more grocery bags you know order some grocery bags there's there's ways that you can play a noise in the background that will give alexa other orders or other commands so data poisoning is basically injecting data into algorithms that or into training models that can buy it that can bias and misclassify what something thinks something is so model compromise backdoors really there's many different ways we can compromise ai but i also would like to add and i know this is a very high level

brief overview that there's ways i ai can also be used for attacks um and so i'll just go into that a bit more so for instance um you obviously know that you know there's lots of bots on the internet you can write ai programs that will look for um the best way to hack a certain target based on based on a whole library of different types of malware there's there's just so many different ways that um ai can be used in in in cyber crime um mimicking a real person for instance very simple example when you're logging into a website and you're trying to you know i would never do this but when you're trying to crack a password right if

there's someone trying to log in they can mimic more realistic human interactions with that login page rather than just you know um just you know throwing a ton of different password ideas at it every couple seconds they can wait a few hours you know try different things and learn from past passwords that didn't work so there's just so so many applications of ai in cyber crime so then we really need to think about how to secure ai and i know this is a really big topic and it's hard to overview in such a short time so i hope we can discuss a little farther and really my objective is to plant some seeds of ideas in your mind

so to secure ai really having some test labs where we can test different types of attacks and different types of compromise on ais standardization is key so really standardization is having you know data standards quality standards not buying products until they've been tested and vetted so that's also feeding into that next point which is attestation so really maturing canada's approach to checking to make sure that technology is enterprise class or just appropriate for the task at hand transparency a big issue that i've learned about is do you trust an algorithm if you don't know how the algorithm works so for instance um in implementing some some ais in hospitals and health care institutions some medical professionals and doctors

really don't like that they don't know how the algorithms work even if we're doing that kind of tier one tier two support model i mentioned um if they don't know that if i submit a photo of my mole if if a doctor doesn't know how that algorithm works or if a professional doesn't know um they might just say no i don't want that technology it has to you know they have to come in and see me in person anyways um so understanding transparency of algorithms is key um and then lastly auditability is really having a way to figure out how things are actually making decisions what data it's inputting so those are just some very high level

concepts and i wish i could talk about this to you for hours but the last point i think in purple there relates back to our point that we discussed about cyber crime so using ai to address ai risks is a really exciting burgeoning area and i'll leave it there because we don't have too much time but really i'd like to turn it back to you guys and hear some questions or ideas on on this topic i know it's hot i know it's a hype topic but cutting through the hype there is some value here what do you think is you know a valid way to use ai in health care what do you think are some some risks that are

coming on the horizon what are your thoughts i'm very excited to have a discussion with you

okay let's just back here

awesome so it seems like there aren't too many questions i know that was very high level but thank you so much for your time it was great to talk to you and i'm certainly you know happy to share more i have lots of information um that was a very high level summary and if you're if you're excited about this i'd love to connect so please reach out

okay we have one great question thanks ashley so it's only been recently pharmacies have began to communicate with one another and still it's not all of them it seems dangerous that they would not have a centralized database for each patient what would you say is the primary factor blocking the sharing of information between these types of bodies for a patient yeah that's a really good question and it varies by jurisdiction in canada we do have a huge precedence for not sharing health data between provinces territories um between pharmacies and it's something that as a consumer of health services i think we all find quite frustrating so what i will say is there's some good progress in this area

um for instance in i believe it was 2019 canada the government of canada um announced a grant of 49 million dollars to support something called the digital health and discovery platform or dhdp um which is looking to establish a canada-wide data platform i don't know exactly what that would look like but the idea is to have really safe secure privacy conscious security conscious linking of health data across canadian jurisdictions regardless of institution so if you're curious about that look it up it's called dhdp digital health and discovery platform so it's something that the government in canada is working on but i hear you it's very frustrating

any ai examples relating to kovid luke i don't know if i have time to answer your question um but yes a google search will find many many many examples china is where you'll find lots of exciting and sometimes scary things happening

yeah actually one other thing regarding ai and kovitt is the tier one tier two approach i met method i discussed um was used in china during covid and that was some so covet has been in in some ways a gr a good thing can i say that for really catalyzing uh innovation in ai and digital health so um in china the you know a lot of the um state um healthcare organizations did begin using an approach to having people answer online forms and surveys regarding their health um their coveted symptoms and then diagnosing or you know evaluating them based on that so whether that was done in a way that is the way we would do it in canada i can't

speak to but very interesting

i should also mention i didn't go into detail on um how can ai be used in discovering different diseases this is really fascinating so human minds can't go through genetics and like genetic data and find indicators of diseases but ai can so you can take a huge data set of genomic data and run a certain kind of ai through it and find basically indicators of certain genetic diseases which is fascinating and new and this is something that people couldn't really do before ai so that's a really exciting application so identifying genetic indicators of disease and the other thing i'll mention regarding kovid and that question about disease is google deep mind did actually help with genetic sequence

genetic sequencing of the covid virus using ai i don't know the exact details of that but check it out online if you're interested in in deep minds i think what they did is they looked at the protein structure or did some research into the um the structure of the virus using ai

is there a reason we cannot hold our own records good question ashley and i do not know the answer to that i think you're always you always have the right to your own data under almost every piece of canadian um health care health privacy legislation so you can at any time call up a clinic and get your records always no matter what clinic it is please correct me if i'm wrong but i'm quite certain that is the case but they're not in a unified database

and actually i want to talk about radiology again because i'm excited about radiology so in in china during covid um ais were used to distinguish between regular um chest radiology scans images of pneumonia or common cold and ones that were indicative of covid so there was a notable difference between the the chest scans of someone who just had a cold and someone who had coveted 19 so that's another example of just how detailed and nuanced some of these ais can be um now i'm not i'm not saying i fully understand that algorithm or how um you know legitimate it was but that was something that was used in in in china which is really fascinating if it

if it worked out

all right any other questions or comments no this was a bit less on security as than it was on more the kind of exciting more new things happening in this area of work so please let me know if you have any security ideas or questions and i'm really excited to discuss