Recording of Behçet’s UK Conference & AGM 2023 held ON Saturday 21 October 2023 in Telford, UK and online via Zoom Webinar.
Programme (Morning Session 2):
Introduced by Rachael Humphreys
CHANGING INTERNATIONAL LANDSCAPE AND CHALLENGES OF BEHÇET’S: Dr Deva Situnayake, Clinical Director at Birmingham Behçet’s CoE and Consultant Rheumatologist
GETTING TO KNOW THE BASICS OF MACHINE LEARNING AND ARTIFICIAL INTELLIGENCE (AI): Daniel Situnayake, Head of Machine Learning at Edge Impulse
AI – CHALLENGES THREATS AND OPPORTUNITIES IN THE NHS: Dr Anmol Arora, Academic Foundation Doctor at Cambridge University
Okay everybody we’re about to start again hopefully everything uh goes according to plan so uh having set the scene I’m going to get give you the first session which as it should be is looking at bashet from a a Global Perspective but I’ve only got 15 minutes so that’s a bit tricky isn’t
It um but hopefully what you’ll find I’m presenting you about to you about this morning will also Link in with what I’m about to say later on in the afternoon okay so first of all um I’ll nail my colors to the mask and I say basically
Is it a disease or a spect A syndrome and I’m firmly in the syndrome camp and I noticed that that appeared in the newsletter recently that Ali jaad from the London Center wrote an article on that uh so there’s a bit of philosophical discussion about this but
Basically as we’ll see later our genetic technology is enabling us to at least for some of the syndrome uh of bashet to dissect out those and we’ll hear later on from Aslam alabadi how that relates to children who get bashet um but this slide really summarizes some of that
Discussion and um you’ll see see that diseases and syndromes are disease constructs and they have to fulfill these two criteria at the bottom to enable doctors and for that matter patients to understand about their disease and how it will affect them in the future so give us an understanding
Of that and actually they’re only going to be useful if they give us some insight into the natural history and guide treatment decisions okay so this sum slide summarizes the uh principles behind the syndrome to disease Spectrum uh discussion and as you’ll see it it can’t be classified as a disease until a
Precise causal mechanism has been clearly defined and I think the jury is out for that with bashet at the moment uh we understand bashet as mostly a polygenic disorder A syndrome uh for which there is no precise causal mechanism although you’ll see again from my talk that there is this thing uh
Concept that’s emerged called the MHC class one opathy concept and I’ll explain a little bit about that uh more uh later on and you’ll hear about monog gen mimics from eslam alabadi so I gave a lot of thoughts um as to what I should say in this meeting
Since I only have 15 minutes so I choose to S to use international data to try and address these three questions that’s number one question why did I get it um how could I pass it on and how will it affect me in the future okay so let’s look at the first question
Um and I’m going to use some data from South Korea uh because the team in South Korea did a national populationbased study to look at the familial incidents and risk for getting bashet and the headline message is is that they did find a familiar risk for getting bashet
But also real revealed a prominent Ro role as always for Gene environment interactions that we were discussing about earlier on so in the context of s South Korea this slide really shows you how the country as a whole has rapidly transformed itself over a fairly short period of time and the country’s become
More wealthy so that slide shows the line going up for GDP and express similarly for per capita of the population and obviously that’s had a big impact for the society overall and also for individuals and one of the things that that yellow bar shows you during that period they measured a
Massive fall in the number of cases with bashet syndrome over that period of about 80% so what was the reason for that and they speculated that because South Korea was a fairly genetically homogenous population with not much M migration uh that this could be due to environmental factors such as changes in
Infection hygiene or socioeconomics in some way but all of these things would interact in this so-called genome environment interaction so let’s delve into the detail a little bit more now the advantage in South Korea is that uh the collection of data is mandatory there is
No opting out it just has to happen uh they have a F 50 million population and data on diagnosis are collected and they’re medically confirmed for bashet twice um and what they also do is they collect family histories for their patients as well so they what they have what we call kinship information
And what they found in the study first of all was that the prevalence of bashet syndrome in the population was 26 per 100,000 of the population and that as you’ll hear later on is roughly well a bit more than twice as much as we have
In the UK at least and the incidence was 1.5 per 100,000 of the population so from a cohort of the population of 50 million they found a cohort of 22 million people that they could follow between two 2002 and 2017 for the occurrence of bashet for the first time they found 31,000
Patients who had bashet in the country as a whole more women than men as you can see um 8,000 of them were fathers 15,800 uh were mothers 2,500 sons and 4,554 daughters and they were able to work out the risk of getting bashet syndrome and this is the data that
Summarizes what they found so first of all uh 53,000 individuals had affected first deegree relatives with bashet syndrome and they found 284 familial cases who develop bashet that’s there’s more than one uh case of bashet in that family and it’s important to note that overall that’s 4% of the familial cases
Compared to the number of cases that had bashet overall in the population so it’s a small proportion to start with okay and then what they did is they worked out the familiar risks so if you look at the bottom number there where it says no family history of
Bashet there was a large number of individuals who had no family history and. 22 per 100,000 person years uh got bashet okay and they’ve classified that as one so that we compare all the other examp going up the slide with that number so that we can work out the
Relative risk of a person having bashet and what you can see is if youve got a positive family history of bashet you’ve developed it in the family then that in you’ve got a we’ve measured an increase in risk of 15 times compared to the normal population okay if you have more
Than one F first degree relative affected you can see it goes up massively to 163 times so that shows you how powerful that genetic risk is similarly for an affected twin an effective sibling was again relative risk of 15 times increased over Baseline and if you have an affected father or
Mother the risk of it being passed on was 13 times more than it should have been if it it shouldn’t be passed on genetically so that gives you a handle of the overall genetics so what they found from that study was first of all that the genetic component of bashet
From families was overall quite a small proportion of all the cases that were occuring sporadically overall okay but what they found was that there was an increasing magnitude of familial risk with increasing degree of genetic similarity amongst the first deegree relatives is which is what we would expect having more than one affected
First-degree relative significantly increased the risks but one of the things I also found was that it was far more common in women than it was in men uh even though as you can see from the previous Slide the risk of passing it on was the same for men and women so these
Authors took all this information into account and concluded that apart from the genetic factors there were very powerful genome environmental interactions going on that accounted for that difference what they could be we don’t know so going back to our question yes there is a genetic component that
Increases the risk of passing it on by a fact of 13 whether you’re a father or a mother uh with bashet in absolute risk that risk is actually very small but something odd is also happening to explain the higher incidence and prevalence in women compared to men and also the Unexplained
Cases of besets that develop it without that family history so they again speculate that there’s a powerful gen environmental interaction going on uh and this must be operating during childhood to influence uh the cases that don’t get bashet so what are those factors so something funni is going on
So that leads me on to the next bit of the talk which is something that we see in clinic all the time and I referred to earlier on and on the right hand side of that slide you’ll see all of the elements of what we call the so-called bashet phenos type okay though
Individual elements and then on the left hand side of the the slide you see the ven diagram and you see some clusters of features of bashet that tend to group together more than we would expect by chance if it was a completely random process so for example more patients get
Mucocutaneous symptoms mouth ulcers genital ulcers and arthritis and lumps on the legs called arith aosm and that’s much more common in women than it is in men um uh then there’s a vascular and a GI cluster and a an ocular and a neuroc cluster so why do these things go
Together like that and we really don’t truly understand the mechanism for that so I’m going to share with you some data now from again another uh country Japan which also is a fairly homogeneous country from the genetic perspective with little in the way of migration the population stable over time and what we
Found from this paper here reported here was from two cohorts a Yokohama University cohort and then the Japanese Ministry of Health and labor welfare database which they have in their country to collect all the new cases with bashet syndrome and what they found was that things were changing over time
They found roughly similar results in these to two cohorts but they found that the uh proportion of the patients in the cluster who had the gastrointestinal manifestations and women that tended to be inre ining over time from 18.5% to 28% and that there was a change in the
Cluster amongst the patients who had eye disease with relatively milder ocular involvement and so again that led them to speculate that there was something weird going on and there was an En environmental process that we don’t know of that was influencing this and they speculated that it could be something to
Do with the changes in the uh country moving moving toward a more western style of diet and westernized lifestyle so that remains to be proven but that’s interesting to think about that and obviously all of these Gene environment interactions that could occur that could account for this are explainable through
So-called epigenetic mechanisms that’s the way that the environment influences our genomes and it we now know that these mechanisms can be heritable as well so they can be passed on so here’s another stud along those lines reported at last year’s international bashet meeting which more or less found the same thing in another
Japanese cohort again showing on the right that group b which is the group that had vascular or intestinal involvement was increasing over time and why is that and I’ve put on the left uh a slide that shows the frequency of gastrointestinal involvement in patients with bashet in different parts of the
Globe and you’ll see that it isn’t the same in all places so it’s more common uh in the west UK America it’s also appearing to be more common in Japan but in the Middle East it’s less common so they don’t see so much of that so what
Is the mechanism of that and what are there important environmental factors that we don’t yet understand that are influence this in influencing this expression of the syndrome so let’s let’s Now quickly look at genome environmental interactions which are thought to operate as I said through modifiable epigenetic or Gene
Regulatory mechanisms but they are heritable so they can be passed on so these could be important factors in relations to the difference between populations that I’ve shown you already and also might explain why one child in a family might get bashet with a family history of bashet but another might not
Okay so can is there anything we can learn about that from the literature so I’m putting up one final uh one slide here that is providing a tantalizing Glimpse to the answer for that question but unfortunately because we don’t have the detailed registry data that they
Have in South Korea or Japan we can’t answer this question directly in the UK uh for bashet so I’ve taken the inflammatory bowel disease literature instead okay because this is a condition that is similar in some respects to bashet so on the left hand side you’ll
See uh that there are a panel of uh lines and essentially what these uh investigators did is they compared they looked at families who had a family history of bashet within their family and that’s all the data on the left hand side and they also compared families with a history of inflammatory bowel
Disease uh with those families that don’t have a family history of inflammatory bowel disease and that’s the right hand panel and then what they did is they connected that information with all the epidemiological information that they had about those patients past history and experience during childhood and in their electronic health records
In this Danish patient register for inflammatory bow disease and there were two things that they found the first thing was that it seemed to be that patients with a of anking spondilitis in their family were more likely to get inflammatory bowel disease than those who didn’t have that history so that’s
Something completely Rel unrelated to inflammatory bowel disease that seems to be influencing the risk and as you’ll hear about later on that links in with what I was saying about this so-called MHC class one opathy concept because anking spondilitis is part of that group so that was interesting but the most
Interesting thing was they looked at their anti antibiotic exposure and infection history going back right into childhood and what they found was that there were two signals that came out from that and what it showed was that if you’d been treated with antibiotics during childhood especially if you’ve been
Given broad spectrum antibiotics during childhood then that modulated your risk of developing inflammatory bowel disease in your sibling okay so it’s one of the determining factors that seems to modulate that risk and they found that both in the families who had a family history of inflammatory B disease but
Also in comparing the families with a a family history of inflammatory v v disease with those that that that weren’t and that I won’t point to them because I’ve got to go away from the microphone but that came out as a consistent pattern so what is that
Telling us what is what’s driving that so summing up all of this information that we’ve heard genetic determinants are the predominant drivers famili for familial aggregation of bashet syndrome and we’ve been able to quantify the risk of that in South Korea it will be different in England but we could potentially and conceivably
Do that with a proper registry um there’s more cases in women than in men and that appears to be a phenomenon that’s emerging in Japan present in the west but not uh recorded in the Middle East why is that is there an environment Al Factor what is the uh effect of sex
On the propensity devel developing bashet and does that somehow interact with these environmental factors we’ve seen that there are differences in frequencies of the gastrointestinal or gut involvement depending on which part of the planet you are when you develop bashet again could that be something to do with environment socioeconomic
Factors infection history or exposure we need to find out and then when we Deep dive into the genome environment interaction we find these two signals of the so-called MH C class one opathy link the anking spondilitis link the impact of deprivation as a socioeconomic factor and then Clues from the Danish registry
For the role of antibiotics could that be influencing either the microbiome or could it be modulating uh the gut microbes in some way that then changes the risk of developing inflammatory bowel disease and obviously that’s possibly a concept that could be transferable to the concept of bashet syndrome
Overall so I’ve now finished my my section and I think it’s the time for my son to give his presentation has he logged on or so we’re going to go to the recording okay so there must be some technical uh problem you’ll set that up yeah so just to give a background to
Daniel uh uh is my son he lives in America and he’s been he’s been working in the field of artificial intelligence first for Google uh and then now for another company where he’s the head for machine learning and artificial intelligence and I’ve had many interesting conversations with him over
The years about this uh and I’m sure they’ll continue so I’ve asked him to give us an Insight in into how these things work deep learning neural networks and so forth so hopefully he’ll give you an understanding of that and maybe he’ll be available later on I’m giving a talk today titled getting
Familiar with artificial intelligence so a bit of background on me um I work as the head of machine learning at a company called Edge impulse we’re a software company that builds tools to help developers create AI um and I’ve also written a couple of technical books on AI and machine learning
As you might have noticed um from the name Dr sitan is my dad and he asked if I could come and give a talk today to all of you about Ai and machine learning and and what those are just a basic introduction so the real first question is what is
Artificial intelligence but I don’t think you can really answer that question without answering a kind of even more basic question to begin with with which is what is intelligence and it turns out this is actually quite a big question and quite hard to answer um it’s the kind of thing
That philosophers debate over uh but we can probably come up with some basic answers to help guide our discussion so I’d say intelligence involves understanding what is going on in a situation it also involves predicting what might happen next having some kind of expectation about what follows from one thing to
Another it’s also about knowing the right thing to do in a given moment and there are many many other things that are part of intelligence but I think these first three are kind of key for what we want to talk about today so if we start talking about what
Is artificial intelligence well really it’s just doing all this other stuff using a computer so can we make a computer program that can understand what’s going on in a situation predict what might happen next or know the right thing to do if we can do that then we
Can create something that is you know justifies the name artificial intelligence so here are some examples of artificial intelligence one of my favorites is a robot vacuum so these little things um they’re really cool you just set them off around your house and they’ll roll around on their little wheels and vacuum
Your floor and they’re smart enough using artificial intelligence to navigate around your house they can figure out where they are they’ve got cameras and other sensors and they can make sure they vacuum the whole floor and then they go back to their little home base at the end and charge their
Battery so that’s a really nice example of of AI I think everybody always thinks of robots and AI together but another thing that’s been in the news a lot recently around AI is the idea of chat Bots chat GPT for example and AI assistance so these are computer
Programs where you can have a conversation with them and they can try and reply to your conversation and help you out or um give you information in some way so they’re using AI to create convincing replies to your questions another example of AI that you use every day maybe without realizing is
Search engines so companies like Google they have big teams working on AI and they use AI behind the scenes to decide which websites are the best match for your query so if you’re writing a question about where you’d like to go on holiday the sites that get shown are
Going to have been chosen in some respect by AI there’ll be AI involved in figuring out which of the millions of websites out there are the ones that should appear at the top of the list there’s also a lot of use in a of AI in modern business software so if you work
In business um you may use AI in your software to help you analyze the complex statistics that come out of running a really large business for example um same with uh any kind of technical software these days there’s a lot of use of AI to help understand the huge
Amounts of numbers that we generate on a day-to-day basis so I want to to talk a little bit about how AI works so AI is basically made up of computer programs called algorithms and some algorithms are really simple so I picked a super simple example here which is a thermostat in
Your house so you set the temperature you want and the thermostat will turn on your heater until it gets to that temperature and then it will turn off your heater when it gets there so the algorithm inside a thermostat is very simple there’s some code here it doesn’t
Matter if you don’t understand that um the the concept is just that you look at what the temperature is right now and then you look at what temperature you want to switch on the heater and you only switch on the heater while the temperature is lower than that so if the
Temperature is less than 18° you will switch on the heater if the temperature gets higher than 18 de then you switch it off that’s an algorithm um even though it’s super simple uh it’s it’s still an algorithm that’s kind of the the most basic that they get but
There’s a little bit of intelligence there right there’s something about knowing the right thing to do when the temperature gets below 18 degrees we’ll switch on the heater that’s a little bit of intelligence captured in some computer code that’s an AI algorithm essentially it’s it’s the simplest you’re going to find
But there are also some very complex algorithms out there and for example our little um Robot vacuum there has a ton of different AI algorithms being used within its code that are helping it navigate around a house and clean the floor and not go over the same patch
Twice and find its way back to its charger at the end of cleaning so some of these systems could be very very complicated and in fact some of them can be really hard to make so that’s what we’re going to talk about next how do you create algorithms so algorithms are
Generally created by a computer programmer or a mathematician they’re a pretty well understood thing they’ve been around for thousands and thousands of years some are very easy to create like the example we gave earlier with the thermostat but others like the robot vacuum are very complex and require lots
Of knowledge and fine-tuning so maybe lots of mathematical knowledge maybe lots of time spent making sure the algorithm handles all of the different cases that can happen many algorithms are so complex that they’d be impossible to create by hand so even if you’re an expert programmer or a master mathematician
It’s so complicated you wouldn’t be able to sit down and come up with it it’s just too much and there’s a good example here so here’s our nice easy algorithm where if the temperature is less than 18° then we switch on the heater inside of our thermostat that’s the algorithm
That’s running another way to look at that would be imagine if we had a camera instead of our thermometer in the thermostat imagine we had a camera looking outside at the street and we’re trying to say if it looks cold outside let’s turn on the heater so that sounds
Simple to a human being like us you look out side and you can see oh it’s snowy there’s snow on the ground there’s snow falling through the air that looks pretty cold I know those things don’t happen unless it’s cold so maybe we should switch on the heater but if we
Were trying to train an algorithm to do it it’s really really complicated this picture it’s a whole bunch of pixels how do you figure out from that whether it’s hot or cold how do you describe to the computer what snow is and whether it means that something’s going to be hot
Or cold that’s something that’s actually really really hard to do and you struggle to sit down and create an algorithm that can reliably do that by hand so this is where something called machine learning comes in and machine learning is a term you often hear in conjunction with AI and it sounds kind
Of magical you know a machine how can a machine learn but it’s actually quite simple and I’m going to explain the basics to you today so it’s essentially a way to create algorithms automatically so instead of going in and doing all the programming by hand you have the
Computer do some of it for you so instead of writing an algorithm by writing computer code you actually collect data about a situation and we’ll see more of that in a second you then train the machine learning model to make sense of it so the model is just a term
We use in machine learning to mean the algorithm that we’re going to be training so as an example here I’ve got a lot of pictures that I’ve collected of cold so if you look at any of these pictures as a person you can see oh that looks a bit
Chilly um we’ve got a snowman we’ve got some snow fields all of these represent what we think of as cold weather so if you saw this out of the window you might turn the heater on on the other hand we’ve got a few pictures of some warm weather so we’ve
Got the the desert there we’ve got a tropical island some palm trees some nice flowers in a sunny um field and these all represent warm so this is my data um each one of these pictures is a piece of data that I can use to train my machine learning algorithm so let’s see
How that works so I start out with this ml model and to begin with it’s untrained it’s it’s basically dumb it doesn’t know anything um and the first thing we do is feed in a bit of data so first we’re going to feed in this picture of a a
Sunny flower on a sunny day and we’ll ask the machine learning model is this representing a warm day or a cold day and to begin with it doesn’t really have any way of knowing that it just kind of randomly gives a result so let’s say it
Gives a result that hey this looks like a warm day so we actually know that that’s correct so well done ml model even though you’re not trained you got the right answer basically by luck so in the next example we’re going to feed in a different image so we’ll
Feed in an image of this Snowman on a a really snowy day we feed that into this ml model and remember the model’s untrained it doesn’t have any clue about anything and it’s still telling us it’s a warm day clearly this is not a warm day otherwise the Snowman would have
Melted so we know that the model got it wrong so this is where the training comes in so what we do now because we know the model got this wrong we’re going to adjust the model very slightly so that next time this image gets fed in
Or one that’s quite similar to it it will give the right answer so we’re telling the model hey you got this wrong we’re going to change you very slightly so that next time you see this image maybe you’ll get it right so we then do this same process millions of times we
Feed in all of our images and we take a look at what comes out of the model and if it’s wrong we tweak the model very slightly and then we do it again and once you’ve done that millions of times with lots of data the model learns to
Give you the right result so now if we put in an image of a snowman in the snow or even an image that’s not exactly the same but is similar the model should be able to tell us that hey this looks like a cold day and that’s basically how
Machine learning works there are more complex models there are um all sorts of different types of models for different types of data there are some for images there are some for just time series values there are all sorts of different kinds of models but they all kind of
Work like that you basically feed in the data and then you adjust the model each time so that it gets a little better and after a lot of repetition it will figure out what you’re trying to tell it so an AI system is built out of lots of different algorithms usually so you
Might have some machine learning algorithms you might have other types of algorithms and they’ll all be connected together so in our little Robot vacuum for example it’s not that it has one algorithm or one machine learning model in there there’ll be a lot of different little pieces that are all working
Together to help the robot do the correct thing and as a system you feed in inputs and it will give you an output so for the little robot um it has cameras and sensors attached those are the inputs and the output is how does it steer itself around how how does it move
And working together all these different algorithms allow the little robots and navigate around your house so there are some big benefits to Ai and machine learning and the primary benefit is that computers are really fast and efficient at repetitive tasks so if we train our little robots to be
Able to vacuum our house that saves us time we don’t have to worry about doing that ourselves and there are lots of other tasks that are effect that are repetitive um that we can use Ai and machine learning to do for us machine learning can also help us find hidden
Patterns in super complex data so we showed the example of those images it would have been very hard to write an algorithm by hand that can determine which category it is cold day or a hot day but using machine learning we find those hidden patterns in the images and
We end up with a model that can represent these patterns for us and we can use it to make predictions but there are some big drawbacks of AI and machine learning so one of them is that in order to work you need lots of data and it can’t just be
Any data it has to be high quality data um so that means that you know if you have the hot and cold weather you have to make sure that all the pictures of hot weather are actually hot weather and all the pictures of cold weather are actually cold weather if there’s some
Messiness there and some of them are wrongly labeled as hot or cold um the the model won’t be able to learn very well and if you only have a few images it won’t learn very well either you need to have lots once you’ve trained your model it
Can be quite hard to understand how it actually works you can measure that it works you can say Okay 90% of the time it’s getting it right but it’s hard to know how it’s working and why it’s getting it right and that can be difficult if you need to have a good
Explanation of how the thing is working another big thing is that AI systems lack common sense because these are just algorithms that have been trained on a computer they only have a little sort of narrow view of the world and they can’t really understand what’s
Going on so I’ll show you a good example of that we trained our model there on a picture of a palm tree on a beach it looks very nice and hot and sunny so maybe our model starts to learn like oh if I see a palm tree that means it’s a
Hot day because palm trees are in hot places um so it must be hot outside however you could just as easily get a picture here of some palm trees with some snow and if we fed this into the model and the model was looking for palm trees and the model says hey
There’s a palm tree that means this is a hot day well sadly you know that’s wrong here we got some palm trees in the snow so that’s a good example of how the model doesn’t have common sense it’s not necessarily going to be able to figure
Out what’s going on um based on the data that it’s been shown during training so some final thoughts about AI really AI is just a helpful tool for understanding data and for automating repetitive tasks it allows us to get a little bit of insight about some data um
And it can do that very quickly and very cheap deeply AI already led to lots of real benefits across the board in all sorts of Industries and use cases it’s a real tool it’s been around for quite a long time and it can be very effective but it
Isn’t a magic wand you can’t just apply AI to something and magically figure it out it’s another tool that experts have to use to build systems that actually do real work one thing to think about right now um there’s so much hype about AI you
Hear it in the news all the time people are always talking about AI um and this is really common when there’s a new technology and there’s actually this thing called the Gartner hype cycle um which describes how people talk about technology over time and when uh new technology appears so recently we’ve had
A lot of this stuff called generative AI which is like the AI chat Bots things like chat GPT um when something new like that arrives it’s very impressive and it gets written about a lot and people get very excited and it goes through this big peak in expectations people think
Wow this is so cool it’s so exciting it’s going to change the world um and you end up with this peak of inflated expectations where the reality of the technology doesn’t quite ma match up to what it can really do um and then after after realizing that people start to get
A bit disappointed and we’re around there now you’re starting to see articles that are like hey AI actually isn’t that great it doesn’t work that well in lots of situations um you know why is everyone getting so excited about AI in the first place and then you go through the trough of disillusionment
Where everyone’s really down on it like oh AI That’s a you know a flash in the pan it we thought it was going to revolutionize the world and it turned out to be nothing but then after that when people start to really build useful stuff and people see the impact of the
Technology on their lives you get into this plateau of productivity where we see hey yeah this is a real technology it has uses um it can be applied to problems and help solve problems but it’s not going to cure every problem in the world um and that just happens
Naturally over time with every technology and where we are with AI right now is about where I drew that star so people got very very excited about it over the last year we’ve started to see that draw down a little bit and there’s a bit more kind of
Realism coming in soon there’ll be a bit of kind of despair of like oh wow this is really not working the way we thought but after that eventually we’ll get to a point where everyone’s just sort of bored really and we’re we’re not talking about AI all the time but it’s become
Another useful tool in our toolbox for helping make our lives better so thank you very much I hope this was a useful talk and I’m sure there’re going to be some really fascinating talks following this one so I’m looking forward to the rest of the day thank you
Very much and have a good one [Applause] byebye thanks to Daniel for that basic introduction I’m sure there’s a lot more to be said but hopefully gave you some understanding of it um now we’ve got anol who is I think coming in live from Cambridge and anal works with alist
Deniston and he’s going to I think delve into a little bit more of the detail of of the medical applications the pitfalls and the difficulties uh uh in healthcare so over to you anal and thank you very much for your uh contribution to the meeting thank you very much Dr sit for
Inviting me and it’s a real honor to follow that talk from Daniel as well so hi everyone I’m anal I’m an academic Foundation Dr Bas in King Cambridge and I’m here to give a quick talk about our research looking at AI in Opthalmology based out of University Hospitals
Birmingham with Professor Dennis as he said and based out of morial eye Hospital in London now we’ve heard a definition of artificial intelligence in the last talk and I completely agree with Daniel that it is difficult to define AI and actually I feel like everyone has their own definition really and the definition
That I prefer to use is that the concept that artificial intelligence is the ability of a computer system to perform tasks we usually associate with requiring human intelligence and that’s very similar to what Daniel said in my mind I distill that down into three key points learning reasoning and
Self-correction and the idea is simple that if we were to expose an algorithm to a large data set then it should be able to learn and reason in order to derive its own conclusions so if we were to expose it to that data set of hot and
Cold images it should be able to learn in reason to Define conclusions as to whatever a new image is hot or cold the key Point here is self-correction the idea being that AI algorithms should be able to learn from their own mistakes that learning and reasoning process should be automated
Because it wouldn’t make sense for us to constantly be interrupting and making changes whenever whenever an algorithm makes mistakes there wouldn’t be any real time saving benefit there so all three of these are equally important although the cell correction is more equally important and of course there’s
A massive range of use cases in Ai and Healthcare and it’s growing every day as well the ones we usually hear about are computed a diagnosis and predictive medicine on the left hand side here but it’s also important to recognize that there are behind the scenes uses that
Might end up being implemented even sooner and that can have a huge benefit to the NHS and there I’m talking about things like bed allocation or simulations for medical education even things like automated drug creation using AI to predict novel drug targets that’s a real emerging area of
Research okay so that’s all for my introduction I want to start with a real world example now and as I said the partnership between Birmingham and morefields it’s focused on AI and Opthalmology analyzing fundo fundus photos like the ones you see on this screen so this is a retinal fundus photo
It’s a photo of the back of someone’s eye and if you’ve ever been to spec saers I’m sure you’ll have had one of these taken now I can safely tell you that the funest photo on this screen is normal I know that because I pick the photo from Wikipedia now a trained consultant
Opthalmologist would probably be able to go even further as well as telling you that the photo is normal they’d probably be able to say that this lighter pigmentation around the vessels in this darker pigmentation here it might indicate that the person this photo belongs to is a relatively young
Person and if there were pathology in this photo we would be able to pick it up things like retinopathy or glaucoma or papal Adema and considering that this is actually a very simple cheap and easy test that only takes a few seconds to perform that’s quite a lot of
Information that a human opthalmologist might be able to get out of this photo but I don’t need to be in the room to to Guess that you’re probably not very impressed by that that’s completely fair enough but what if I were able to tell you that if we exposed an AI
Algorithm to this same photo this cheap quick and easy photo that we’d be able to get information like age sex blood pressure cardiovascular risk or even risk of dementia and now I’m hoping that that is impressive because that’s the current Frontier of AI research in healthcare and that’s the idea of
Ocomic so ocomic is the idea that we can get an insight into systemic disease through Imaging of just the eye the idea being that if we use things like fundus Photo simple Imaging that we can look Beyond eye disease and start to look at cardiovascular disease or neurological
Disease now this broad idea that eyes are a window into the soul or a window into a wider Health has been around for a long time but it’s only in the past couple of years with developments in AI that Professor Denniston and the team here in Birmingham coined that term
Ocomic and it all culminated this year really with the paper at the very bottom here a foundational model for generalizable disease detection from retinal images and I won’t go into the technical detour detail of the paper but just for a high level overview the model was
Called R found it’s a huge AI model trained on publicly available data sets as well as data from morfield Eye Hospital linked to people’s medical records so we can see their systemic diseases as well as their eye problems and we’re looking for three things of course we want to be able to diagnose
Eye disease from eye photos ocular disease diagnosis that’s the Benchmark we need to be able to do that uh for it to be used in Opthalmology The Next Step which is actually quite impressive is looking at prognosis or progression of disease trying to predict that from simple eye
Photos and then the part that got the headlines was ocomic trying to predict systemic disease from simple eye photos things like stroke risk of heart attacks or even Parkinson’s disease now this all sounds revolutionary and life-saving I hope when I describe it like that but before everyone Rings Specsavers and tries to
Get a fundus photo booked in just be aware that it’ll be a while before this is widely implemented at scale and just linking it back to Daniel’s talk we’re going from the peak of in inflated expectations all the way down to the TRU of disillusionment here on the Garten
Hype cycle because there are lots of different barriers to adoption of AI in healthcare and these are by no means specific to the NHS in fact the NHS is probably ahead of the curve when it comes to promoting AI now every one of these barriers is a presentation of themselves so I’m not
Going to spend too much time dwelling on them the only one I really want to pick up on now is stakeholder involvement because it’s very well characterized that for any Innovation to be implemented in healthcare it needs to be acceptable to all the major stakeholders involved of course it needs to be
Purchased by Healthcare organizations in the UK that would be talking about hospitals in America we might also be extending that to insurance companies and accountable care organizations naturally it needs to be regulated by the state especially in a publicly funded Healthcare System it needs to be approved and acceptable by
The state who would be funding It ultimately Academia or industry tend to be the people who are developing these Innovations and they of course need to be on board for it to be actually used in the real world in this case we’re talking about an innovation that’s being created by Academia
That morefields Birmingham partnership but it could well be another Innovation being created by a drug company for example now perhaps the most important and final stakeholder is the public if anything’s going to be implemented in the NHS the public have the final say really on what they’d be
Willing or wouldn’t be willing to use especially considering that the organization and the state themselves are directly accountable to the public so if the public aren’t on board if Charities aren’t on board if patient involvement groups aren’t on board none of this will ever be implemented and that’s why public and patient
Involvement is Paramount to all of this AI research and both morefields and Birmingham are really leading the front there in terms of involving patients of course there is a need to ensure safety Beyond those barriers this is a very new challenge to Regulators because I mentioned that the start that
AI incorporates those elements of learning reasoning and self-correction now those aren’t oneoff things when you to develop the algorithm they’re constant processes along the life cycle of an algorithm normally when we talk about medical devices they’ll be tested in a lab environment tested on people and at
Some point they’ll be marked as yes this is safe to use and they can be distributed now with AI these algorithms change progressively over time we may implement it one day but the algorithm changes within two three or four years and it’s really important to ensure that
We’re keeping up with those changes and we’re constantly keeping an eye on their safety over time and that’s a real challenge because we can’t just have static medical approval there we need something that’s continuously evolving and watching over I’m coming to the end of the talk
Now but I just wanted to draw attention to two of the big initiatives here at Birmingham standing together is a really exciting project looking to develop standards for diversity and data sets it’s an international collaboration including universities from across the world as well as patients and publics
From across the world and companies like Google are involved as well and there’s a really exciting paper being published here looking at diversity in data sets and the standards we need to try and promote that uh but that’s been embargoed for a press release in the
Next couple of weeks I’m not sure I’m allowed to say very much about that one but please do keep an eye out for it and the second thing is building out research data or in for structure and this is critical to AI implementation in the NHS Daniel alluded to the fact that we
Need a lot of data to train algorithms and in healthcare that’s particularly difficult because we need to make sure that that data is kept incredibly safe secure and private but at the same time we need the infrastructure that data might be able to be shared between researchers and between organizations so
That everybody can be represented in the data sets that AI is trained on so Birmingham and morfield have led the way in Opthalmology through developing this Insight Health Data research and there are ongoing wider efforts to try and build out National Data infrastructure as well okay that’s all I wanted to say
Apart from thank you very much for listening if anything I said caught your attention I’d really recommend you checking out one or both of these lines of work on the one hand we’ve got rep found and the novel technical work going on to try and build out Cutting Edge
Algorithms that are at the top of of healthcare Innovation right now and on the other side we’ve got the implementation science looking at how do we actually get this stuff implemented in the real world what are the barriers to adoption we need to overcome and standing together is a perfect example
Of looking at algorithmic bias there so thank you very much for your time there’s been plenty of media coverage and I’m hoping that you’ll see some more of it soon I’ll Stick Around for any questions in the chat but feel free to contact me later on by email or however
Else as well my emails on the slide thank you amazing thank you uh Dr anal um we do have some opportunity for some live questions actually I is he still on Zoom or is he gone oh he’s there there we are he’s back so we do have an
Opportunity for some live questions if anybody has got anything in the room yeah okay an thank you very much for that first that that fantastic talk of yours and um can I ask you the first question um which really relates to the r found uh paper in nature that was
Published um which was I thought a fantastic paper um one of the things that um I’d like if you could explain to us the concept of the AI and how you trained it and it trained it more or less trained itself using unlabeled data uh and how that’s a so-called Foundation
Model in other words something that can be a a sort of Force for good across the globe could you explain that yeah absolutely so foundational models are really quite difficult to make actually because they require huge amounts of data set the idea being that if you’ve got enough data you can pull
It all together into one big model as you say it can be unlabeled data as well and the AI algorithms can start to pick up patterns of their own accord within that data set so com back to Daniel’s talk before we saw hot and cold images
Where all labeled so you train some an algorithm on some hot images on some cold images and if you present a new one it can defi it can try and predict which category it might fall into here we’ve got enough data that we can just pull it
All together the algorithm picks up on the patterns itself and with some Human Assistance admittedly and some labeling we can then get to some kind of predictions from it now one of the key things here is it was trained on more than just one type of algorithm so so I
Sorry one type of data so I focused on retinal fundus photos because that’s what we’re most familiar with that’s what’s most common but actually the algorithm was able to pick up OC scans as well so if you go to your Specsavers or Vision Express they might offer enhanced Imaging of your retina and
That’s what these o scans are and they’re very common in Opthalmology and the idea here is with more than one Imaging modality we’re able to generate predictions that’s really actually quite novel uh and I’m sorry I didn’t focus on that too but a lot of the foundational
Models we see in other domains really just pick up on one Imaging modality this specific ability to pick up on more than one is is really important here of course in terms of the development and testing we can’t just brain it all on our own data and tested on our own data
And that’s why there were so many different data sets used here as well as the moral Z Hospital data which was used for the majority of training there were publicly available data sets and UK biobank was used as an external validator so I hope that’s a little bit
More of an overview uh I wasn’t sure how specific to go in terms of the detail during the presentation but I’m glad someone asked about it that was fantastic and I thought the other great thing that your the team did was to allow their uh Foundation database
Resource to be available for the public goods so that other uh AI teams can use that too uh which obviously is the great example uh for how we should share our data and learning um so that’s wonderful are there any other questions in the in the room for
Anal no okay well that’s fantastic thank you very much uh for helping us out uh at the last minute and um you did fantastically well thank you very much have a good [Applause] day okay so for those of you that are in the room we’re now going to have some
Lunch so please do help yourself take a break there are some breakout areas if you want to um sort of go to there or you can bring it back to your table and just aim to be back in the room seated by about 20 to 2 um and for those online
Uh please take a break uh remember to join using the afternoon link so that should have been sent to you in an email you might have had a reminder in the last hour as well because it’s a different link to join the afternoon sessions and we’ll see you back again
About 20 to 2 all right so go and enjoy mix