Speaker 1: Hello, everybody. Indeed, my name is Harmon Boes. I'm working at the company Atos, where I am the team lead for AI and machine learning. I'm based in the Netherlands. So my background is actually originally in bioinformatics. So I used to be working in research at a genetics department in Groningen, where I was mainly concerned in doing research on DNA, and really looking at genome-wide association studies, seeing if I can find disease markers and correlate them to research. But at one time, a researcher or a recruiter actually approached me and he was like, hey, do you want to work in the corporate world? And then he showed me a number and I was like, well, that sounds very interesting. So I guess I'll, I'll take you up on that offer, which meant I ended up at Atos. Atos is actually a really big company. So it's one of the largest IT companies in the world. It has over 110,000 employees, and we're really distributed everywhere. So we have big partnerships with Google, we have big partnerships with Microsoft, we organize all of the IT in the Olympic Games, for example. And yeah, I mean, this is a bit of shameless promoting of the company I work at, obviously, but it has it has a purpose, I promise you. Because if you look at the site here, we do a lot of outsourcing. And we do a lot of consulting and system integration, which means that we're working at a lot of different venues, we're working at a lot of different companies. And one of our biggest companies of biggest things that we work for biggest sectors is health. So we work together with a lot of hospitals. And that actually is in the very low level end of hospitals. So just in the system integrations, the security on their firewalls, you know, the access gates into the hospitals or into restricted areas, which is, for us, a very valuable business to have, right. We're also based here in Switzerland. So we have around 500 people here in Switzerland that have a lot of this expertise as well. But what it really enables us to do is actually work together with these types of companies. And now from my field, it gives me a very interesting network to approach them and say, hey, can we actually help you with industry proven standards to apply machine learning things from other industries that we've used before? And is there anything that we can use you with with the systems that you're already using that we already implemented for you? Now, one of the hospitals thankfully told us, well, that sounds very interesting. So come along and let's see what you can do. So that's a really big hospital in the Netherlands in the Amsterdam area. And it was mainly regarding the emergency departments. So what we really wanted to look at, are there any processes in this emergency department that we can optimize using machine learning or any type of analytics to actually help them and see if we can improve the situation there a little bit? Because the emergency departments and hospitals are obviously dedicated to people that need immediate medical care, and they can be visited at any time on any day. So it's a very crucial part of our society. Now, in the Netherlands, over a million people visited every emergency department over yearly since 2013. And this number is rising by roughly 50,000 to 100,000 people a year. One of these reasons is that there's a large increase in population, but also that there's an increase in population aging, which means that the likelihood of more people visiting emergency departments increases yearly. Especially ambulance usage is increasing for critical care patients. However, due to the uncertain political climate in most countries, the budget cuts are always looming and all the hospitals are very wary of actually being able to sustain the level of care that they are, and they want to look towards the future and see if how they can improve the situation that they have right now. Now, we propose that data analytics and machine learning could actually help them remedy the situation. And to do that, I made a very basic sketch to understand the situation myself a little bit. And I hope also to explain to you what's happening. So in a single hospital, you have roughly three types of patients coming into an emergency department. One of them is walk-ins. So that's just people coming off the streets as it were with a cut or a bruise or a broken arm. Those are the biggest uncertainty. So you don't really know how many walk-ins you're going to get. There are obviously some gut feelings that they have. They know it's really bad weather. There's maybe the likelihood of more people coming in or if there's a news event, there's a big crash or a big pile on collision at the highway. They obviously know, but there's also a lot of uncertainties there. Now, then you have doctor referrals. They're fairly certain about doctor referrals. Doctors really put a good diagnosis in. They call ahead and they know that they're coming in. So this is the least uncertain factor. Then you have the critical care patients. So there are people that are coming in by the ambulance. You already know a little bit about these people because the ambulance has specialized staff there. So they know how to describe the symptoms. They know how to call it in and they know how to inform the doctors properly. But this is still an uncertain factor because you don't know how many people each day are going to be coming in by ambulance. Now, the moment they come into the emergency department, there's an emergency department team that's ready to help them. Now, these nurses and doctors, they are evenly planned over the week in three different shifts. So every day you have a day shift, you have a morning shift, you have an afternoon shift and you have a night shift. They're able to handle roughly between 10 to 40 people at a consistent manner during each shift before diverting these patients to a different emergency department. Now, within this process, we can already see several things that maybe we can help with. So it's a logistics problem, right? So we can see how can we predict how many people are coming in based on historical data. There is a triage problem because they might not be using their resources optimally because we can actually look at the symptoms that are coming in and see if the relevant doctors are actually on staff or on call right now to help those patients if they are not maybe redirect them somewhere else. And there is also a simple staffing problem, right? Because they have a limited amount of staff, they don't want to overburden the staff because if it's really busy, they just need more people to handle all the inflow. But if it's really, really quiet, they might want to give people some time off and save some costs. Now, the first case that we looked at was predicting the total inflow and planning of patients. So imagine if there's 100 patients coming in, and we can predict that 50% are going to be coming in the morning, 35% are going to be coming during the day, and 15 are going to come at night. Now, using the numbers that I showed you, we're pretty certain that at night, it's going to be fine, right? They have enough people to cover what we're predicting. Now, in the morning, they might have more people coming in than they maybe can handle. So they might have to call in an additional staff member. So they add somebody here. Now, in the middle, we have the most interesting case, because we were discussing this and we were very optimistic. And we said, well, now you can actually take one staff member out, right? And they were like, well, I don't know if we should, because is your model accurate enough to warrant that? Right? Because we're not talking about some sort of delivery that Amazon is doing somewhere, right? And they might be late a bit. No, it's actually people not being helped at a critical moment. So that one is actually very telling. We can only look at, can we add staff members? We are very wary of actually predicting whether they need less staff members. So that's one of the things, sadly, that they're not getting any budget cuts out of yet, because we don't have 100% accuracy. Now, another use case that we looked at is how can we direct ambulances more effectively? So if you have an ambulance coming into a hospital, you have very extreme cases. So this is a very basic and extreme representation, but it drives to the nearest hospital, right? It takes 10 minutes to get there, and the emergency department is full. Okay, so we drive to the next nearest hospital, which is also full. And then finally, you drive to a third hospital. And yes, there is a place there, but that took 60 minutes to get there. Whereas if you already knew that those two hospitals were full, or at least that there was not staff that was able to help that patient there, you could actually have gone to the not intuitive hospital 30 minutes away, and you would have been saving yourself 30 minutes. Now, the extra thing that we want to add here is to also streamline the communication between the actual ambulance and the emergency department, because it's very specialized personnel, and they are very good at giving clear and concise instructions regarding the symptoms that each patient has. So if they're able to use a system that they can quickly and effectively put that in, then we can already see if we can match that to the relevant specialist on call within the hospital. So it won't be a matter of somebody coming into the hospital. And as you see in all of those series, somebody shouting all of the symptoms, and then somebody else going, Oh, no, we have to call this and this guy. Now it's already ready. He's standing there and he knows the symptoms of the patient and he knows how to help him. So saving time is of critical importance there. Now we have some measures of success of this. So things that we determined at the start in this pilot that we wanted to do is reduce emergency stops from an ambulance, reduce staff workload, and optimize staff planning and resource allocation. And then obviously showing the live and future patient inflow. To do this, we try to create a prototype dashboard that has to function in a form of an air control tower, where you can see the interactive view to see all the predictions based on the input data. You can make it available in a more batch manner to the roster maker within a company suggestion of who to plan in and maybe which specific nurses as well, because they might have different skill sets. And we are starting now by predicting the next shift and we're scaling up to upcoming weeks. And up to now, we can actually predict up to a month in advance. Obviously, the uncertainty gets higher, the further you go in time. The data that we use for this is three years of historical patient data. So since we're at a big company, and this is where it comes in that is very handy working at a big company, we have big partnerships. So we're working together with one of the suppliers for this information that works for hundreds of hospitals across the world. So we approached them and asked them if they could actually make this type of data available to us. And thankfully, they did. So three years is still potentially not enough. But we have a lot of nice information markets here. So we know the idea of the patient, we know the gender, we know the age range, we know the urgency that it was classified as we know the illness type, and we know the time spent. Now we can use this information to add some seasonal components. So we can add the day of the week, we can add the month. So we know if it's in the winter, we know if it's in the summer. And we can add what time they came in. So what shift was it in? Now, there are some additional events that are very important here as well. So for example, weather, we can back correlate that to actual days and see if there are spikes in there that can be attributed to bad weather events. But also we can look at events. And this was actually one of the most telling markers that we found, because Amsterdam has a very big football club called Ajax. And every time Ajax played, there was a spike in people coming to the emergency departments. So that was actually fairly telling. And we were looking at different events like that. So there's also the Amsterdam dance event, for example. And those all those types of events really showed a correlation with how many people came into the hospital. So if we know that in advance, obviously, the hospital itself knows that in advance as well. But it's also very nice to see at least the model prove us right on those instances. Now, for the baseline model, we ran into some problems. So this is the baseline model from a few weeks ago. And I'm happy to say we at least improved upon this a little bit since then. We used multiple simple regression models. And we looked at how it would perform. Now, the accuracy was fairly good at 80%. However, if we look at the max error, it was up to 22 people that we were wrong. Obviously, any model that's 80% correct has a one in five chance to be wrong in the order of magnitude of 22 is not very usable, right? Because either two more nurses are on call, and nobody shows up. So then they'll be pretty mad at us for making them incur more cost. Or if they do take our first optimistic suggestion to not plan into extra nurses, there'll be massive chaos and panic. Now, some of the things that we're still working on is improving upon this model, obviously. And we're now determining with them what is actually the minimum required accuracy to actually take a valuable action for you. But there's actually another problem that we did not anticipate because I'm a technician, and I was very happy to be able to predict something for them. But I always forget about the people, right? So you have nurses and nurses do like to know in advance when they're going to be working, right? So there's a large change management track that is needed to dynamically fill these rosters, because they want to know, when am I going to be working and they don't want to be called one shift in advance, you need to show up this evening, right or this morning. So there, it's a really big problem for us to find a mode in which we can both accommodate people's working schedules, but also find some valuable insights or actions that we can do with this type of work. Now, we also want to look at something that we do know will have an immediate effect. That is implementing proactive communications with ambulances. So obviously, there's already things like 911 or 112 in the Netherlands that already coordinates a lot of these actions. But we really want to see if we can help them be more proactive about contacting ambulances so that there are not these rather really life threatening situations where you show up to an emergency department and there's nobody to help those people. We're also running a live simulation right now, because we don't want to implement it immediately, we want to see what would have been the impact of our model. And we want to define concrete actions that have to be taken on these model predictions. Because that's one of the things that always is lacking, because we're now trying to really apply machine learning to a practical process. The question should always be okay, then what action do you take if the model predicts x, y, z. Now, some of the takeaways that I want to give you is that AI can be very powerful if applied correctly. And it can really benefit practical processes in our really day to day business as well. So we really think that we should look at different things like hospitals, we should look at a lot of different public information systems to see if we can actually apply AI there and help the general public. And for me, applied machine learning is really about identifying your relevant business process and providing a relevant insights to take action. And solutions are applicable across industry. I've worked at Atos and I started working with NLP. I started working with logistics companies. And I see all of that back in this project, which would not on first sight be very logical to think as an NLP expert and logistics expert, I'm going to implement a big project at a hospital. So I want to thank you for your time. And if you want to contact me, this is my information or I'm at the Atos booth upstairs. Thank you very much.