 I want to introduce you now to Dr. Angus Lawler, who is an assistant professor here in UCD in the School of Computer Science and Informatics. He's a funded investigator also in the Insight Centre for Data Analytics. His primary research interests are in the area of machine learning, artificial intelligence, explainable AI and recommender systems, and he applies this to sports analytics and medical imaging. He's been actively involved in a lot of different research projects in this area over the years and is working on some very significant collaborations with industry. He's also worked on grants from Enterprise Ireland and SFI and his research has led to several granted patents as well. So over to you Angus. Thank you. Thank you Brian. So what I want to talk about today is some of the work we've done on a data definition of hitting the wall for marathon runners. And again this work really started I think when a little bit after Barry had started to work on data analytics of of running and he said we got this problem and so I kind of got into working on the data as well and also started running on the back of it to kind of better understand the data and the kind of processes behind it. So I think on my first marathon I don't think I hit the wall I kind of feel like it's a club you want to reluctantly be a part of when you talk to marathon runners they're like oh yeah I hit the wall like seven or eight times and I kind of wish that would happen to me somehow but it didn't happen the first time so maybe the next time. And this has worked done with our student Yakima now works in Kipman Labs. I'll just take you through some. So some of this has been kind of referenced in Barry's work before. When we're talking about the marathon the interesting thing about the marathon is really the the quantity and the quality of data that it generates and this is something that's changed over the last few years as well as people have better trackers that can collect better quality more high resolution data. This kind of really helps us as computer scientists. They're much better watches that have more accurate GPS they also have heart rate detectors and so on. So this is kind of the opportunity for us and in terms of participation as well in the last 10 years I found these statistics that in countries like India they've had over 200 increase in participation in marathon running. So there's been a fast increase in the number of people who are participating in marathon running and if you look at the countries that have improved their average marathon times in that same 10 year period you can see here that Ireland has improved by about seven minutes which is pretty significant on average and countries like Switzerland have improved by 14 minutes so this isn't a case of just putting on your running shoes and going for a run. Training for a marathon is a multi-month event that requires significant preparation training and so on and these people need guidance which is where the recommended systems the computer science side comes in and also these are people that are really pushing to improve their performance and again that's something where we think we can we can help them. So building these kind of smart coaching systems, recommendation systems to guide or adjust or suggest new human behaviors is kind of the target of most of our work. Now when it comes to the wall again there's a famous picture of this this poor runner in the London marathon people kind of say well you know it when you see it and that's largely true you know this kind of jelly leg syndrome or where the legs just don't obey what the brain is telling them to do is fairly obvious but in fact many runners who report hitting the wall don't experience these kind of symptoms they experience collapse fatigue a whole range of other symptoms and there isn't a precise definition for what exactly hitting the wall means. We also find that there's a vast amount of advice that people give about how to avoid hitting the wall and much of it is based on intuition or guidance from coaches and not always kind of referenced or backed up by literature. So again most of the things we can these kind of common tips here do weekly long runs run at least 118 to 20 miler don't go out too fast this again is the dangerous Barry mentioned of starting too fast and finishing slow taking breaks walking breaks during your marathon again is again a common tip and of course good nutrition and hydration is also really important and again in this meta analysis that Alison mentioned in her talk it was found that the one kind of feature that most predicted whether you would hit the wall or not was whether you'd done one at least 132 kilometer run in your training period all the other factors were found to have variable kind of impacts on your performance. So this is something that we're trying to look at can we come up with a data definition of the wall and then relate it to things like your your training history and maybe make recommendations for how your training might change or how you might change your pacing during the race in order to avoid hitting the wall. So what exactly is the wall well again these numbers vary by marathon but somewhere around 40 percent of runners who run marathons report having hit the wall it usually starts after the halfway mark so I think 33 kilometers most common onset point for the wall of the runners who do hit the wall they 70 percent of them experience a slowdown so this you might conclude that 30 percent of them don't actually hit experience a slowdown which is kind of remarkable they report hitting the wall but they don't actually experience a slowdown in their pace and of course as Barry mentioned as well males are much more much more likely to hit the wall than females again the reasons for this are not understood but everyone seems to have an opinion on it. So what can we learn from this data study right so so we try to analyze the wall using the data that we had a lot of the data came from the marathon data that Barry had scraped and we also got some data from strava.com which has activities and also training data and some of the the kind of statistics are there so if you look at the pace profile of a marathon runner this is this would be quite an accomplished marathoner here as I'm sure you can tell so we can see that kind of features here that would indicate this is a person who's run a very good marathon the finish time here is under 200 minutes you can see they have quite an even pace there are variations here but it's quite an even pace and what I'm showing here is the average pace and so you can see at the start they go a little bit behind their average pace and then they gradually speed up over the course of the race so again this would be a very nice runner who did a very good job of pacing their marathon and this poor runner did kind of everything wrong here right so you can see here they started off way too fast and then by about the 26-27 mark they suffered a pace collapse and you can see their pacing kind of went down to that walking pace here at the end you can see these big peaks here they're maybe trying to recover by stopping to take a break to walk or maybe take take on some some water some nutrition and then interestingly enough after the kind of peak here for the final couple of kilometers they actually recover start running again and almost recover their pace to their mean pace right so and again this would be fairly typical for somebody who hits the wall and so this led us to try and come up with a way of identifying runners who hit the wall from runners who don't and so we use these clustering techniques to partition marathon runners into those who hit the wall and those who don't and these are the averages of these two clusters so this blue cluster is those who hit the wall that's this would be a typical race profile for those runners and what you see and this this orange one here is the runners who don't hit the wall now you can see in both cases they tend to go out a little bit too fast and then finish a little bit slower again this is like an average but you can see for the runners who hit the wall they go out way too fast right so there's a big difference in how they start the race and this kind of oops this point here at the end there's about a 20 slowdown experienced by those who hit the wall so again this starting pace here is noticed to be about 10 to 15 percent quicker and again this work is you can find us in in this paper here so this led us to identify those features a little bit more precisely and then these features we can use to to categorize or analyze the kind of pacing of runners and put it into some of our machine learning techniques so we came up with this definition here where you know we we've kind of quantified and identified the specific phases of a race and how it might relate to the wall so you can see we have the wall starts we have the duration we have the intensity the recovery the peak slowdown and so on and so these features what we'll use when we're building our machine learning models now again Barry has presented some of these results a little bit more detailed analysis of the marathon record so I'll just go through these very quickly one of the things that we find is based on based on age so I'm showing here the the the axis here is age we've got male and female runners and we see that as expected men are much more likely to hit the wall than females now again there's a subtlety which Barry is mentioning here was that we don't know what the target is of these runners so some of them are targeting maybe a personal best in that case they might be more likely to hit the wall we don't know what the training what their kind of fitness level is or any of the other things which would impact on the wall like their nutrition or their hydration so a lot of things that are missing again this is just a data analysis of many marathon runners we also find that the intensity of the wall is greater for for men than for women again it's not a huge difference but but definitely men experience the wall more intensely and we also find that women tend to recover after having hit the wall they tend to recover a little bit better so again I don't know if this is because they know what kind of behaviors to change so in terms of walking or reducing their pace or taking on nutrition or or water is it just that they're more informed better informed or they just respond to you know the kind of feelings of fatigue or pace collapse we're not exactly sure what the reasons are but again this result has been is kind of banned for all of the marathons that we've analyzed we also we also looked at this in terms of the finish time of the runners so this is a kind of a measure of whether the runners are elite runners or recreational runners so these would be the elite runners over here and we found that the recreational runners more of them hit the wall than the elite runners now there are more recreational runners but they can they're more likely to hit the wall than than the elites and we also find that the impact here so in terms of the intensity the impact or the kind of the intensity of the wall on the the finish time of the recreational runners is larger than on the elite runners so again this kind of gives us a good motivation for improving and building these recommender systems models that can help the recreational runners because there's a lot more people in this category here that we can kind of assist than than these kind of elites which you probably have professional advice anyway and so then in terms of the proportion recovering we can see then that the recreational runners are also more likely to recover from the wall than the elites yeah okay and and again I think you know if you want to look at more detail in terms of the training in terms of people who are trying to achieve their personal best there's more details in Barry's paper now one of the things we also looked at we were interested in was was not just the wall in terms of the kind of the impact or the intensity but also the psychological effect and one of the things we noticed in the self-reporting of people hitting the wall is that it seemed to be driven by you know course related features the weather whether there was a crowd or there wasn't a crowd or certain features like the elevation in at certain points in the race and so we looked at some of the the leading marathons and trying to relate certain parts of the race with the probability of somebody hitting the wall so this is probably the audience that's most likely to recognize marathons from their elevation profile does anybody know which races these are Boston no so which one is London no this one is London does anyone know this one no it's New York this one is Berlin so I have on the next slide so this so Berlin is actually a pretty flat marathon course which is why most people go there to I think achieve if they're kind of targeting a record so Berlin tends to be fairly easy and you can see in Berlin most of the people who hit the wall they tend to hit it right at the end so this is really their fatigue but you see interesting kind of behavior in the case of the London marathon where when they get to this kind of canary wharf part it's it's there's a big spike in the numbers of people who hit the wall and it seems that the canary wharf part of the London marathon there's not that many spectators and it's a kind of a windy part of the course and the number of bridges you've got to go over it's really tough and a lot of people hit the wall at that point and again you might relate that to a psychological effect and they're clearly not I don't know why they would be more fatigued here than they would be for any other marathons and we also see similar effects in New York for example the Queensborough bridge is this this point here and in the New York marathon you've got to go over this bridge and the elevation of the bridge is quite high and there are again not many spectators so a lot of people hit the wall at this point and also there's a big spike here just at the very end there's like a little hill in Central Park here coming up which is really tough right at the very very end and again a lot of people experience the wall here at that point so again when it comes to the wall it's not just purely physiology there are effects of the course the profile weather and kind of psychological aspects as well so what we tried to do then was to take features so based on these features that we used to to analyze people's profiles or racing their their running profiles we were able to build runner representations which came from their training data we were also able to build a training set of data of runners who successfully ran the race and were very similar to these runners in terms of their training history and their ability and their kind of likely finish time and we were able to then recommend a pace to those runners that they should follow in order to avoid hitting the wall during the marathon so this is work that was published in our Pace My Race paper which was presented at REXIS a number of years ago now and again one of the nice things here is that we're we're kind of using these recommender systems models to build to build pace recommendations that are personalized to the runner they count also for their training profile and they learn from runners who successfully ran a race at similar paces but avoided the wall so they're also course specific as well because we can train them with runners who run on the same course now one of the things we've we've also tried to do we've tried to put some of this work into apps which can make these recommendations this is difficult to do it's difficult to recruit runners for these kinds of tests to kind of build the apps and get them to use it so we're just kind of showing a mock-up of what our app would look like here and one of the things we do in the Insight Center is to try and explain the recommendation so we realize it's no good to say to somebody here's what the model predicts in terms of what your pacing should be we need to provide an explanation for why the model thinks that's the right thing to do and so when it comes to training for a marathon race which is you know an activity which might take four months or six months of your time providing these kind of explanations for guided behavior is is essential so this is something that we're working on how do we provide these explanations what what form should they take and what's the most effective form they take to guide a person's behavior in the right way now there are other apps which provide recommendations for pace during the race but most of them provide you know like average pace so they just block the pace out and they tell you if you're going above or below average but but getting a recommender system to run on a watch is a kind of a trickier task but again this is kind of where we see in terms of the future incorporating much more information we'd really like to have information on nutrition and hydration and sleep quality this is all information that's missing from this kind of data analysis right we have pacing information we don't have all the other information that's necessary for making really good recommendations so just then to to kind of conclude I think as everybody knows hitting the wall is an important phenomenon to study it's something that anyone who's training for a marathon really wants to avoid these large-scale data analysis give us some novel insights into what exactly the wall is and how people experience it and also ways that we can avoid it so again basing these kinds of decisions on evidence of what people do to avoid the wall we think is kind of the right way to go so we need to understand the wall in order to learn how to avoid it and there's also a lot of scope here for you know recommender systems or these kind of smart models that can guide human behavior you know when you think about training for a race again a long period of time and it's often very difficult so the kind of suggestions that you make to runners are often not ones that they kind of readily accept so how you do this and how you change behavior is again a kind of an interesting problem for the kind of recommender systems that we're looking at and of course I think there's an argument here for much broader human sensing again there are all these sensors are out there about what we eat about the sleep and different kinds of exercise that we take and again we don't have those integrated for a large number of runners so that's something we'd the kind of a direction we'd really like to go in so that's that's all I have to say thank you thanks very much Angus I don't know if there are any questions from the room or are people scared to ask questions about the wall okay we'll have one when you say someone's going out too fast so for these runners we know their training history so we we know the activities they've done for the previous four to six months and we use these models that we've trained to predict their their their finished time in the marathon so that's how we know if they're going out too fast well you can find them in the papers so so I think Barry mentioned the regal model or the regal model which is actually pretty good the area there is about seven to ten percent we've built machine learning models which get down to I think three percent kind of error so I don't think there's a repository online but if you want to come and talk to me afterwards if anybody wanted to invest quite a lot of money in a startup that that would be another route to getting that out there with one more okay so I think Allison is gone but what so they did a meta analysis of thousands of papers research that been done on this and they found that the one feature which predicted whether you were so let me get this right I think this is the right way around to the one feature which was most predictive of whether you were going to hit the wall or not was if you had not done a 32k run no not the more one yeah this one okay okay we want one more question English thank you very much for sharing your resume of resume of great data I have one I'm kind of locking the horns which are just small things because traditionally knowledge and orders were always so that if it's a physiological effect that happened when you burn it's simply when you burn out your carbohydrates and you're going to break down your fats and the fats have broken down more slowly and that's the reason that you slow down and that happens the only thing I can agree with what we're saying there is that does happen around the 18 20 21 mark that depends on the pace but in my view it's largely a physiological effect right so I would agree with that yes so the collapse of pace is largely you've just simply run out of energy yeah but then but people do take on nutrition during the race they take on water and so on so the precise point at which it happens and the effect is again you know things that we're trying to understand but Angus what I thought I understood from what you were saying is that the psychological element may be what triggers it you know and that it is physiological but that there may be a psychological trigger for it sure and also people recover that's the other interesting thing they recover within a you know seven to ten kilometers they often get back to their average pace again okay thank you very much