 And for the third time, welcome to those in the room to this session on AI and machine learning. I'm Brenno Horos. They're working at the HIST Center at the University of Oslo. I have the pleasure of simply hosting this session with Elmarie Klassen from HIST South Africa, Bahid Rustami from Macarons, who will be presenting use cases for using the HS2 data and integrating with these new methods. So I will just give the floor directly to Bahid. Please welcome. Hey, hello from my side as well. Can you hear me well? Okay. Hey, so thanks Brenno. My name is Bahid Rustami. I'm a machine learning scientist at Macarons. So I'm a very technical person, but I will try to keep the talk as non-technical as possible. So I'll give a short overview of Macarons and who we are. And then I will try to explain a bit what is AI and machine learning and hopefully demystify this buzzword that you hear all over and hopefully keep you until the end of the talk. Because also a perfect time probably for a short nap after lunch. And then I will go to one of our use cases in Sierra Leone when we work on forecasting essential medicine and vaccine. So Macarons is an AI company and our vision is to create more from less by making pervasive system intelligence. And we are working towards a work where nothing is wasted and where each resources generates the greatest impact. And it's only possible by using AI and you will get to know more how this is possible later. And also our mission is to extend what is possible with finite resources, solving the global problem of ethically matching supply to demand. And also this bring us this kind of big challenges of adopting whatever AI and machine learning methodology exists out there to be adopted to the low resource setting. So I will talk mainly about our work in global health, but we are working in transforming supply chain in basically every type of supply chain in agriculture, environment in education and in industry. So Astriata is our core product and is a platform that I will go to in the next slide showing you what are the functionality of our product. But here is a map of the deployment that we have with different government in the across, so in Africa, Asia and also United States. And in dark blue, you can see our staff that is actually also distributed all over the world, and we are a fully remote companies. So, so Astriata as I mentioned is a is a platform and is our core product and it has three pillars. So, Astriata infrastructure Astriata forecasting and Astriata behavior. So Astriata infrastructure tries to answer simply the question of what is there. So if you think of, for example, a map at the national level. So what what Astriata provides you is that you can zoom in in this map go to the health facility level and and see what is there what is in the shelf there like both in terms of staff and also in terms of material which are there. And also it gives you a on top of that it gives you a kind of optimization strategy to how to redistribute the resources which are there to to actually use them efficiently. And Astriata forecast tries to answer question like what's needed where and how how much so basically it gives you a number that what is the need of a specific products that you're looking for in next month or next quarter or next year. And and Astriata behavior tries to answer question like who will show up for care and who won't. And this actually gives you a good understanding of that who for example is a patient who will drop out from education and you can or will be a loss to follow up. So basically combining all these three together we can have this comprehensive solution that at the end helps the decision maker to to to decide how to allocate resources. So now just talking a bit about the so the name of the session is AI and machine learning with DHS to so you all know about DHS to I'll try to explain the what is artificial intelligence versus machine learning and what is the so deep learning maybe some of you heard about. So artificial intelligence is the I think terminology existed forever even. I think pre-socratic philosopher talk about or even Greek mythology and the idea is to create intelligence which is artificial so mimicking human intelligence but making artificial kind of version of that, which can be outsourced and kind of come up with some limitation that I might have. And there are different approaches people followed in artificial intelligence and machine learning is one of them which been quite successful in last decades because of the power of computation which grow and also because of different approaches that developed both in academia and industry and and the idea there is to learn from data without explicit instruction. And then within the machine learning approaches the one of the most successful one in the recent decades is a so-called deep learning where you when you use a neural network which is inspired by human nervous system and is only inspired because still far away from how the real brain works. And but it's been quite successful in solving real world problem. Now the example I want to talk about is that so what are the potential of AI solution compared to traditional solution I think most of you are familiar with many different types of solution you developed for your problems in more classical settings. And I want to hear now compare what what AI gives you compared to that so if you think of traditional solution usually you have a problem. So you try to find solution for that so you define some rules and then you can combine the data you have with some rules. Then you get some outcome and based on that outcome you try to make decisions so you can think of like Excel sheet you have and some like operation you do between your columns and rows and get one number and then make decision with that. The characteristic of this type of solution is that is a static so as far as the rules are fixed outcome is always the same. Eventually over time the error increases because outside world is changing but the rules you put in are fixed so so the error that you make with your solution goes up. It needs perfect data so even if you have a perfect rules for your problem and the data is noisy you will get a bad outcome. And this kind of solution become absolute very fast because as I said the outside world is changing unless you're changing your rules constantly this kind of solution doesn't work. On the other hand with AI solution what we do usually is that first of all when there is a problem for example in global health we talked to the expert in the in the domain. We just gather some prior knowledge of like what is the problem limitation and we don't need the exact solution but we just need some prior knowledge. We combine this prior knowledge with the data we have at hand and then first we just let the methodology that we develop to infer the rules and then based on those rules we try to do prediction and as the new data comes we constantly improve our prediction. So characteristic of this type of approach is that so it's dynamic as the new data comes the prediction you make constantly is improving. The error eventually decreases because the algorithm you have a place actually like learns from the new data which is coming and also it doesn't need perfect data and this is also one big. Direction also at Macro is we have to really like come up with solution how to impute data how to deal with missing values and grow like intelligence solution to deal with that. And also the important part of this is adaptable and to just summarize this slide I think if you think of pandemic as a good example when the pandemic. What happens with the traditional solution is that the so the outside world, the pattern that demands everything's changes like hospitals are occupied for different reasons, but the rules you have for your solution are still fix really big error on the outcome you have but what happens with the solution is that at the onset of pandemic you make a big error of how you observe the word but as the time passes in the course of days or weeks or months. The algorithm learns from the now new rules which actually exist in the outside world, so it will correct the error it makes. Okay, so now I want to go to the. So one of the use cases of our product that we grow and also deployed in Sierra Leone working with the government of Sierra Leone and MoHS. And so I will focus on the forecasting we had there but also just want to mention that we also have a HR optimization tools where basically we. We kind of like developed solution how to basically like re-staffing the different health facility to make sure that all the medical cares are supported. So, so for forecasting what we did was that they're talking to the government we we got a we got access to DHS to data so so our engineers developed like API to to connect and pull data from DHS to. And then we we also agreed with the government to to focus on eight essential medicine that they had and so they told us to work on and three types of vaccine. And what we did first was to looking at DHS to data and come up with the actually like a measure of what is the error in place like right now the system that works. What kind of error the system makes and we use the three months rolling average for forecasting which means that we look at the last three months consumption I use that as a forecast for future. And that was the baseline error that we defined and then we developed our solution and here I just show you the result and then I will go to the details of the challenge and solution we had. And we could show that depends on the type of products we were looking at we could manage to improve the forecasting error by 34 to 59% for essential medicine and 56 to 89% for for the for vaccine. So now like a starting with the data so DHS to data that we had we did some quality check off the data and here I want to show you the I think this this plot is here shows you that the all the data we had so we had data from January. 2019 to January 2022. And on the y axis you can see the percentage of missing values in this data. So, so in 2019 we had like more than 80% missing value. It got better, like in 2020 but it always stayed around 40%. And this is I think for for whoever is doing forecasting is a quite big challenge to to manage to improve anything with these type of data. And also underneath I just want to show you two example of the data we had for for those who've seen time series before to have a better understanding of what type of time series I'm talking about so here you see on the left and the right in the bottom you see the two different types of medicine at one CHP. And on the x axis you see the time and also the y axis shows the the consumption over over months. You see that there are a lot of missing values also the other interesting thing in this data is that there's a lot of zeroes there are like 20% of these data which are not the real consumption so that's what we call false zeros, which could be a human error or sometimes there was a so the the missing values and they just replace it by zero. And we had to also come up with some strategy to detect those zeros because they all influence the whatever prediction you want to make. So, yeah so so that's solution without going to the technical part of that whoever's interested would be happy to talk about it afterwards. But what we do is that for forecasting usually if you think of like when you want to forecast for one variable you look and think of like for example for weather or stock market. You look at the historical data you look at your variable how it was behaving in the past, and you try to find some patterns like some trends or seasonality or some auto correlation in the in the in the this variable you see. And based on that you try to predict the future and this is what people call like univariate forecast. If you have multiple variable but we can do is that not just looking at one variable itself but you can look at the correlation across them and see that for example if the future of one variable can be detected or forecasted by by looking at the others. And these are like there are methodology already out there but all of them needs long historical data to be able to have a good forecast. So here in our case here I'm just showing you four variables we had 9000 variables so we had eight for essential medicine we had eight products and we had around 1200 health facility that we wanted to forecast for. So we had more than 9000 of these variables that I would call time series. And also we didn't have a he's like a long historical data so pair time series or pair product at the facility, we had on average 10 to 15 samples and this is I think for those people who work in the forecasting. They know is a kind of like quite impossible to do any type of forecasting in this. So what we did was, first of all, we developed quite different strategy to create more from the data which is in at hand so created different features from this data. And also try to not only focus on the DHS to data but looking at the whatever publicly available data out there like satellite imagery web scraping whatever that could be correlated with the consumption that we have at hand. And also looking at all these nine times series at the same time try to find the whatever like different order of correlation that we can find in this pattern across different facility or product to help our prediction and these are all done with the so called supervised learning or machine learning approaches that that we developed at my Christ. So going to to now to the to the to the result. I want to show you so what we did for validation is that is that we kept the last three months of the data from DHS to as a test set so we put it like a side. And then we trained our model and the rest of the data and then we asked. So first we look at the baseline so that baseline as I explained was these three months rolling average. And we asked like what is the error if you use these three months rolling average and this is what you see in red for each product so for example for amongst the ceiling here you see that if you use three months rolling average we get around 69% error. In green you can see our error so if you use macro is solution we get 37% so we had like around 46% improvement for this product and and of course depends on what product you have we managed to improve the forecasting error between 34% to 59%. So for vaccine we use actually very similar approach but here we had another prior knowledge let's say that's what I meant that talking to expert in the field that we knew that consumption of these three vaccine are actually very similar so when when they're at the facility they use one type of vaccine very very likely they use also others so we use this knowledge to first impute the data and then actually following very similar approach we had. And also I just have to mention I don't have a figure for that but in vaccine the missing value percentage that I showed at the beginning was way lower. And that also helped us to actually the improvement that we managed to have here was was drastically better so we here we had like between 56% to 89% improvement for the for the vaccine. So, so now I want to actually like tell you like what are these percentage means so so I'm talking about 56% error what does it mean in quantity so for example if you think of vaccine. So here for example I'm showing BCG vaccine. If the for example MOHS wants to actually like has 100,000 vaccine wants to distribute. If they make decision based on three months rolling average what would happen is that they would have plus minus 64,000 error to the kind of like the quantity that they distribute. So if they would use the macro is forecasting the deal actually goes down to 7000 so so you can see that depends on the total supply that you have, you can have actually like quite drastic improvement of the wastage that that that you might actually have. So actually that's what I wanted to share to sum up I want to just mention that. So all our effort is actually bringing AI in the and actually adopt all this methodology to be applicable in one of the actually most difficult and poor data environment. And this needs a lot of actual research and also practical work so we work with the also different university we have like research direction which tries to readapt all the machine learning techniques which are out there and normally actually are very much optimized for big data or for organization that they have a lot of data. But we really try to get this methodology try to adopt it to be applicable for low resource setting. And I just want to finish by by this sentence saying that I heard a lot from many people, especially in global health sector that we don't have data to do machine learning but but I just want to say that this is exactly what we are doing we really try to revisit all the methodology out there to be applicable even in this example I showed you we managed to to improve the the forecasting quite a lot. Thank you very much. Thank you very much. I think once you display you can switch. Swap displays. Yeah, top top left third icon from the right one across. Yeah. Right. Okay, should we get started. Right, I'll start down here. Hi everyone. So I've had the privilege of working with Omri for for many years. She tends to take my job over time and does better than me and leaves me with quite little to do. So I'm going to kick it off and then hand it over to her. So they're two sort of fundamental misconceptions around data science that we picked up over the years and by data science I mean AI and predominantly machine learning is the space that we're in at the moment and the first misconception is that it's mainly wealthy countries wealthy settings that should be doing AI and machine learning and they're going to they're going to benefit the most and they should do it. And what we're finding is that's fundamentally wrong because AI and I think you started seeing that in the previous presentation is is most useful for resource constrained settings, which is actually all of our settings in reality I don't think anyone here would feel there in a setting where there's too many resources, but particularly in some of the countries like African countries where those resource constraints are so so severe. The second misconception that we've picked up over the years is that other people are doing this really well. They're brilliant people out there that are doing data science in health so well, don't even bother getting into it. And that we've also picked up is fundamentally wrong and those are the few that are following some of the research that's come out of after COVID and there's some of the meta analysis on the machine learning work that was done in the previous presentation is found that actually this is a very messy space that's not done particularly well. And that the strong use cases that are being done well are very very few. And so it's a space that we decided actually probably is worth in getting into. So those are general misconceptions I'll tell you about some sort of more personal mistakes that I've made along my career that I think also helped to make the point here. So in early 2000s and one of my biggest frustrations working in South Africa, particularly as I got out into the more rural settings was, wow, health sector is a bit of a mess. We don't resource very well we don't plan very well we don't manage anything very well. Maybe I should get more into public health. And I started moving into public health and public health informatics and started looking at the sort of data that we had available to make decisions and I realized wow the data is not very good. Maybe what we should be doing is working more in digital health health informatics and starting this generate better data because if we've got better historical data, we can make much better decisions. And I probably spent a decade doing that kind of work and what we've learned is that our historical data is not usually that useful and that valuable because it doesn't tell us that much about what we currently are experiencing. And so coming full circle it's starting to get us really excited about the world of data science and machine learning because if we can have improved our systems and have better quality data coming through. But then we can work with it in a new way in a more predictive way. We, we may start getting to a point where I'll start feeling okay well this this is this is looking better. Anyway so that's about me. Let me push some buttons here. So we were going to spend some time talking about terminology but our colleagues dealt with a lot of the terminology so we won't worry too much about that. But for those that are new in this space initially and certainly we found that as well as there's a lot of language being thrown around. But fundamentally what we're talking about when we talk about data science is AI and machine learning and much more towards towards machine learning. So this is what we're going to cover cover today. So this is a picture that some of you probably know very well. And it talks a little bit about that cycle that that I was referring to but it's about how we, as we can start doing more things with data we can generate information as we can start being able to share that information and internalize it and then push it back and share it back to other people so we're just moving towards this knowledge space. What we're really trying to get is to get beyond that and get back down into into the inside and what some people talk about wisdom and get more predictive around their understanding what it's really saying about our environment and then actually use that to take decisions that then have an impact on our environment. And if you actually start generating those benefits that are pushing back to health system changes health output changes, low maternal mortality better infant mortality how can we make those sorts of decisions that's what we're actually actually going for. And in the DHIS world as you progress along your DHIS maturity you move from there and can start feeling the improvements and get really excited but while you're still in that top sector. We probably haven't generated any health system output improvements or outcome improvements. And that's, and that's one of the big problems is we, we, if we get stuck there we spending more and more money more and more resources are being sucked into this information space around trying to make better decisions, but we're not improving the health of our populations. We may not even be improving the health system much, but we need to get down to this space where we start seeing those those decisions becoming more productive and actually driving the, the outcome changes around impact. So a little about a bit about his and my his but the top we meet his South Africa's journey. In 2016, we started getting exciting excited about this because we realized okay. We're not so far behind they're not a lot of smart people doing this brilliantly so we shouldn't bother. And, and maybe it's worth it for for our setting. So we got past some of those early misconceptions, and we created a data science unit within his. It had no staff, it had no projects it had actually no know how at all, but we decided that let's make a start we'll create the unit. And we were faced with our first use case challenge, which was South African government wants to allocate 14,000 health professionals a year. So that's medical interns, medical doctors doing community service training nurses pharmacists a whole lot of people where government allocates them specifically to facilities. And the political mandate or reason for that is to get more equitable distribution of these junior professionals in a way that should help transform the health system. If you can get those junior professionals to start going up to more rural settings and the less attractive facilities in a more equitable way. You should start balancing some, some system system issues. So that was the challenge that we are faced with and we said well goodness. How do we do that. And we discovered this was actually a data science challenge and how we put together algorithms and how we start understanding those individuals, collecting information from them to say well okay you 14,000 people, where would you like to go. Then could we introduce some incentives to say well, if you go towards those less attractive facilities will introduce some incentives. So if you choose a less attractive facility in your top five choices will guarantee you get it. If you can only choose the really attractive facilities in your top five choices well you know you take your chances you may end up somewhere else. So those are the kinds of discussions we started happening with the Ministry of Health and how do we put those into some kind of a tool. And that's when we, that was the first challenge that was put to put to our team. Anyway just a little bit of a history we're going to get into some different use cases now. And it was as we got into that where we started realizing we need, we need to start appointing some people that really understand the space and we appointed our first data scientist in in 2019. And then some more use cases started coming along and these therefore that we're going to get into into a moment. And I'll be handing over to to Elmerie because then they get to the tricky stuff that I have to give to her. Thank you for the time here. Good. Yeah, yeah, so I'm going to take you through three very high level use cases that we have and then one into a bit more detail, just to show you the kind of things that we are working with. So we have a project project with Africa CDC looking at event based surveillance event based surveillance is sort of outside of integrated disease surveillance where event based surveillance looks at events that is not specifically reported or diseases that are not specifically reported so you're looking at events that comes from social media or from the media kind of thing and I'm going to go through four steps and my three use cases to how we identify, improve, innovate and impact on in these use cases. What we're looking at here is to use a P Twitter is a WHO product that sort of uses natural language processing to identify so you can add in tags like say covert Ebola or whatever, and then you get all the tweets that has happened about that. Now, for me to say I'm sick and tired of reading about covered. That doesn't mean it's an event of covered it's just my sort of frustrated tweet on social media. But if you were to say that you know monkeypox happened in a specific country. That might actually be something that is not yet picked up. And the challenge that the epidemiologist that Africa CDC have is that if you were to look at say, even 1000 tweets a week for them to go through all of those. It's a blood of human capital that is spent on just figuring out which is really events that they need to respond to, never mind then going to respond to those events so what we can do is to pull all these events we can get the 1000 events from later. But what we then need to do is to we are refining the output through a machine learning process to identify, which is exactly those tweets that are likely to be events. So we perhaps give them 100 to look through and say, are these real events. And then if they are, then we post them into the DSI system through the events that we have pulled. And then hopefully over time, we're going to learn from those 100 that maybe 60 of them were not real events and we learn how to better present them with a lesser amount of possible what they call signals that could be become events. So that is just a simple use case we're working on. And then the second one may be a little bit more interesting for the DHS to community. There is something like predictors in DHS to and I know it's very much used. The challenge that we have was predictors in DHS to is that you must select a certain algorithm, either exponential logarithmic and sometimes facility profiles and the data in facilities are very different. You might have a small facility that we looked at this specifically for sort of HIV patients in care and looking at how do we increase the patients in care for 95, 95, 95. Now, certain facilities are increasing the number of patients other facilities have a different process where where you might have started treating those patients in hospitals. Now you're looking at them treated more in in primary healthcare facilities and hospitals are going down. So using one algorithm is not always giving you a correct solutions there. So what we have done is that we developed a best fit algorithm for running this at every facility level and determining the algorithm that needs to run that that you need to use to predict that in for each facility. And then what we're doing is to link that to machine learning process. It is quite processing heavy so we're running that in a machine learning application. But pulling then also developing a DHS to front end application which can select which elements or indicators you want to run it on. It runs it in the machine learning application and then would post it back into DHS to prediction into data elements and then Essentially, we see a lot of use case for this in terms of predicting targets predicting in ideas are perhaps thresholds that you would want and essentially hopefully we can very soon contribute those to back to the community as a solution that you can use with the app. And we will see how do we implement the ML solution most likely would be something like a Python server that we can give you the the whole process if you can install it if not, we can write it as a service or something like that. And then health workforce modeling is the other big case that we have. So, you know, many of you might also be aware of the wise and model to where you look at how many doctors should you have based on population, 10 doctors but then 100,000 population so many nurses and so on. But what we found in South Africa is that those norms are not always practical for the Ministry of Health to implement and how do we develop a scientific model for health workforce modeling in terms of how many stuff do we need. So we have developed this looking at a specific WHO and South African norms for predicting And we use one district as a proof of concept. And now we are basically refining this ML model and build to build a planning model and then also what we're doing is to build a front end on a module that gives would give the Ministry, Ministry options to do scenario based planning. So if I tweak the the norms to from 10 to 9 doctors then what does what would that output be. If it is epidemic then how many ICU beds would I need and how many ICU trying stuff to a need versus maybe a non epidemic scenario and things like that. So, And then, you know, it's really to be able to add that value at yes we have HR data we know how many data there is, we have indicators but we don't have this type of intelligence in terms of helping them to say if this then what, you know, and what is the how can we help you make better decisions. What is that question that you really need answered on on that level. And so this is just some of our workforce models modeling results we looking at it over head count as well as population based. And I'm not going to go too much into this because I do want to, but you can see basically here for and it's also in South Africa, it's very much about equitable distribution. There are areas that historically got enough stuff and more than enough stuff and other places that have got a lot less stuff and patients are really struggling. So you can see here that we actually coming out with a surplus and shortfall or a gap in terms of the number of stuff that that is needed. I'm nearly done. Just to then say, in terms of the workforce planning context, there is a bit more of a in South Africa we have a public sector that serves 80% of the post relays. So population with around 500,000 health workers and then 20% of the population is served with around 400,000 health workers. So you can see that there is a lot of sort of misalignment there as well. And really here in workforce planning is to help the ministry to determine optimal staffing levels to mitigate attrition and to recruit what recruiting interventions that you have, what training capacity is required in the government worthy staff. So on all of these we basically look at attrition and just developed one simple use case. And these are the sort of steps that we are following. And just firstly to say what what is the question that we really need to answer. And that here we're saying it's determining the staffing needs by analysis of exits and considering who are those imminent retirees and or people who may leave the department and what what would their reasons be? Are they unhappy? Are they going to resign or retire or what is the reason for the for the exit? And then we needed to see what data do we have access to and where can we source new data, additional data that would help us to answer these questions. Do we even have enough data? Do we have access to this data? And then to get that access was another process. We then did an initial analysis on this data. And sort of these are some of the sort of looking at ex-employees and current employees. And this is a graph shows the reason for the resignations. The bar shows the mean and the upper and lower courts for each reason. We want to understand how variables are the reasons, how variable are the reasons for leaving for each age group. If you look at the second bar resignations, it shows that the mean is 38 years of age and the range is between 30 and 48 years of age with some outliers above 80, which could be a data quality issue. And this graph explores the years that people have been an employee as a service. And we wanted to know if we are with money as a factor to make people stay. The education level in this is the size of the bubbles and the lowest salaries, a lot of them are leaving earlier. So, and below 10 years of service, while the longer you stay and the higher salary you earn, the longer you are actually showing. So, this is really around knowledge discovery and understanding what data that we have. And the next steps then is to look at the finding the use case and decide whether we actually have a viable case to go further into this. Because it is expensive, you need to run, you need to get a lot of data, you need to spend time doing all these things. And honestly, not all data science projects are successful and actually give you a result that actually is worthwhile the investment that you make in it. So we decide whether in our early explorations, are we actually having something worthwhile to go further into or should we just stop here and rethink or not. And then what we're doing, the data engineers, which usually the repeated data, those are tasks such as take a position processing and governance. And we then go to the data scientist role of training the algorithm. And as described before we separate the test set, more or less 20% of the data, and we use the race of the data to train the algorithm. So what we did is, if we could predict from data that we had, which employees would actually leave in the next three months or not. And through this process, you essentially iterate this training. You know, repeat that until you're happy with your, your algorithms and your model. And then once you trust your model, we deploy it on a subset of employees of the December 2020 data, comparing the predictors predicted results against what who actually then the resigned. So we predicted who would resign in December 2020 and then compare it against who actually resigned in 20 or who left in 2020 December 2020. So you can see on the left, the actual employees who who left the department and the prediction of employees who are likely to leave the department. So if it is a one, then they are likely to leave and if not, if it is a zero, then they were not likely to leave. And then you can see for those who had actually left and that, you know, versus what we had predicted to leave. And we got quite an accurate result there. So then essentially what we need to do is to display this data for management to make decisions. And this is just some of the graphics around that. And on next steps is to scale up the data sets to refine the accuracy, identify and access additional data sets to adapt the model to predict further into the future. And essentially what is the value for this for the Department of Health, they basically can understand why do people leave, who is likely to leave and where will the gaps be in the future. So if we have only a few big iterations in the country and, and 30% of them is going to leave then how are they going to replace those, those iterations because then they can take action. Maybe they could prevent that attrition, through a person said to approach plan for training or recruitment of new staff, limit the cost of the training, and limit the loss, limit the loss of the main expertise and over their overall ability to plan better. So this is the strategy 2030 of the National Department of Health, in terms of attaining universal health coverage. And really what they need to do is to plan for effect efficient workforce to attain universal health coverage and this work is essentially touching on all of these green parts of their strategy. And, yeah, and for us as his essentially our endeavors to get into data science are growing we are learning we have got data science to data scientists that we will have on staff shortly, and it creates new opportunities for us, and adding value to our clients because we're moving away from just looking at data is data is data and information but looking for that intelligence, worse than man impact that we could reach with that data. Thank you. And just the acknowledgments and for the projects that we are working that this on is a CDC project. And this is some information about this. Thanks.