 So, lle here I wear my other hat so speak, as you saw last week, and as I explained most of my research life has been spent working on clouds and convection, both in high resolution models and also in terms of the parameterisation of those processes in climate models such as ECAM and, of course in the IFS at ECMWF that you have been accessing under S2S. The nice thing about ICDP is that you have really a nice, should we say, freedom to look at the application side of things, and that was something I really wanted to get into ICDP to look at how one can translate forecast information into a usable end product in different sectors. I was involved in two European projects which are, how do you say, the emblem up here, so one was Healthy Futures, which was based in East Africa, that finished last year at the end of 2014, and Queasy, which was quantifying effects of weather and climate on health in developing countries, that finished in 2013, and again that was Africa focused. So I've got a number of co-authors here that have provided data, or Philippa was working with me closely on the data analysis. So what I want to do is talk a little bit about this system, but to remind you I'm not a biologist, so I don't have a 20 year experience of all the biology, but I would do my best if you have questions on that side as well. So I've got a few slides to start off with to give you a little bit of background about the disease of malaria, because probably a lot of you were like I was just three years ago not knowing anything about the disease, and I was involved in this project on modelling the effects of climate and health, so I thought what I'd better read up a little bit about the diseases and how they're modelled, and that's how I kind of stumbled into this modelling project by accident, almost one of my, Andy always says, crazy Christmas projects. So in the end there was no open source model in which I could try out my own ideas, and so Andy Moss, who was one of my friend and colleague, goaded me into basically writing my own model, which I did, so this is what I want to introduce to you today, and what we've actually done in this project. So malaria is actually caused by a parasite, there are actually six, I mean you often see four species, but now one of the species they actually think is actually two separate species, that's still a little bit controversial. There are actually four main ones that generally are found to infect man, the two principal ones are ffilsiparum and vivax, but you also find in restricted locations around the world cases of these two species, malaria and ovale. Now the most dangerous one is ffilsiparum, that's the one that's most commonly found as well in Africa, so that's the one that causes the majority of mortality. Vivax is less dangerous, there's a whole interesting background of why they think that is, about the ages, how long we've lived with these various species of parasites, but I don't have time to go into that today, but as a coffee time conversation I actually find very interesting. Vivax is the one that can lie dormant, so even if you treat it you feel fine, it can pop up again a year or two years later, that doesn't happen ffilsiparum, it doesn't have the dormant stage. So ffilsiparum is the dangerous one, vivax is the one that's more difficult to clear. If you look at a map we talk about malaria endemicity, so endemicity we're talking about basically the level of transmission, how intense the transmission is, what proportion of the population you would expect to have the parasite. You can see that these days the colours are not reproducing terribly well here, but it's essentially restricted to the tropics, and especially in Africa, so we have these, should we say, grades of endemicities going from hollow endemic where you basically have all year round intense transmission down to hyper-endemic, or you can refer to this as epidemic, where maybe climate restricts the disease to occurring perhaps only every four days. Or five, six years, so you won't have transmission every year. And of course this is important because it means that we have the capability to build up immunity to malaria, but it takes quite a long time in these exposure, immunity is not fully understood yet, but essentially if you survive until your third, fourth, fifth birthday and you're living in an endemic area, then you build up immunity. So in these strongly endemic areas it's children that are most at risk and pregnant women, whereas in epidemic areas then you have the whole section of population that can be at risk because basically between epidemic outbreaks the population will lose their immunity to the disease. Do you have a wider risk profile? So the transmission fringes are probably the areas you want to focus on if you want to look at how climate variability is going to be controlling the outbreaks of the disease where you might have potential information from a forecast system. I'm going to come on to the climate drivers in a moment, but essentially if you look at a map of Africa zooming into Africa you see that basically these are the fringe areas where transmission can be irregular. So you can see there's a band across the Sahel that's pretty much affected by the extent of the monsoon, so variability in how far the monsoon stretches north will control year to year variability in transmission. So it's the fringe of the rains, whereas these areas in eastern and southern Africa you can see basically coincide with the topography, high areas where temperatures are colder and you're near the fringe of the temperature that can support transmission. So if it's slightly warmer than usual you can go above the threshold in which transmission can be sustained. So these are the areas we're going to tend to focus on. A little bit more background again, this is the world malaria report from 2014 and again it's just to emphasise why for example in these projects we were focusing on the continent of Africa. But of course some of the work is relevant for other continents and I'm trying to build up a project now for example in Brazil with colleagues there. Annual mortality is estimated now for the most recent year of around half a million, that's quite a drop. Some of that as well is due to improved diagnostics, I'm going to talk about that. So some of it's not a real drop but a lot of it is due to the scale up of interventions. So for example now it's estimated that in Africa about 49% of the population has at least one bed net. And these are the figures here at the bottom I want to emphasise, I mean global spending in 2013 was now almost $3 billion. It's quite a lot of money, you want to make sure you target that money effectively, I'm going to come back to that theme. It's still below the targeted spending, the estimated amount that's needed to really get on top of the problem is almost double the actual figure that's available. So I've got a couple of figures now that will explain a little bit about the background of how climate actually affects the transmission of the disease before I go on to the modelling. Before I do that, one of the parameters that I'll be using as a model output to describe the intensity of transmission is basically the entomological inoculation rate. That's simply the number of bites that you receive per unit time that are actually from an infective mosquito. So you can imagine that if you get 10 bites in a day from an infective mosquito your chances of having, acquiring an infection are much higher than if you only receive one infective bite per month obviously. So it's just basically a measure of the force of infection, how many infective bites you get per unit time. And so just as a back of the envelope, rough estimate, an EIR value of around 10 per year, okay, so that's 10 infective bites per year, roughly marks the division between an epidemic area and an endemic area such as a meso endemic area where you have regular transmission but just for a short season, okay. In very hollow endemic areas for example where you have, you know, you're getting towards that year round transmission then you can get numbers as high as 500 or even a thousand, okay. So I think in Cameroon is the highest that they've measured the head estimates of around 3000 infective bites per year so it's an average of 10 per day, okay. Just to give you an idea of the range, very roughly, very, very roughly the log of the EIR translates into clinical cases, okay. That's a rough, I should have put a slide into that. So what are the climate drivers of malaria just to give you a background of the linkage? Well there are a number of ways in which climate affects malaria, okay, through a number of parameters. Basically the key ones are temperature and rainfall but wind and relative humidity also will affect the transmission. So wind, how would you think, I was always asking you questions last week so let's start again. Now how would wind affect transmission for example, any ideas? Exactly, so perhaps if it gets too strong they might try and shelter, they can get affected. You might, you have cases where they've had not much malaria but you have cases of disease outbreaks sometimes in North Australia where insects have been blown across from Papua New Guinea when they have strong winds. Insects do use winds though to track CO2. If the wind gets too turbulent and too strong it's very difficult for them to track you. In fact I find here in the summer because we have a lot of insects here that one of the best ways if you don't want to use a net when it's hot is just to put a fan on it just disperses your CO2. It's amazing the way they track a little bit of side information in brackets. So they basically sense our pheromones, essentially mainly from the ankles. So if you have smelly feet you're very attractive. So that's how they identify. A lot of these I'll talk about in a moment. There are some mosquitoes that are particularly dangerous from malaria because they like to bite people. So you can have 100 cows in one person but they will know where the person is because of the pheromone signal. They can sense perturbations to CO2 that are a fraction of a percent above background and follow the plume. So often people think about wind. They say it blows the insects downstream but it's actually the opposite. They follow the wind upstream to actually find you because they receive the signal. So somebody in the quantitative life sciences group, the biology group, they're actually looking at ways insects track. But the key variables are rainfall and temperature. So I should move on before I get too slow here. So rainfall of course is absolutely critical because a lot of the main species, there are over 50 types of the enophilies mosquito that can transmit malaria. But there are two or three key ones that are very important because they are the ones that really like to bite people. So they're basically anthropophilic rather than some insects are zoophilic. They prefer animals rather than people and some just don't care. They're not fussy eaters. They just take what they can find. And those, like enophilies gambier for example, likes to lay eggs in small, clear, sunlit temporary pools. So rainy season starts, you start to get ponding and puddling. We're talking about ponds on the scale of maybe one, five, ten metres. If the ponds get too big, you have ripples that drown the larvae. So they don't like big lake bodies. You also have predators if the water body is around for too long. So they're basically insects of opportunity. Small, clear puddles, sunlit in those cases except in high temperatures. So I'll show you in a moment how transmission in a lot of areas, if you don't have a permanent water body or a hotspot, because you can get ponding around a lake, it's very complicated that hydrology in fact, that mostly the transmission season follows the rainfall season but with a lag. Temperature has a multitude of impacts. I'm going to talk about some of those. So here we have a schematic of the transmission cycle. So when a mosquito emerges from the larvae stage into the adult stage, they don't have the parasite. It can't be passed through the egg stage. So a new born mosquito never has malaria. The female needs to take a blood meal in order to have the protein to develop her eggs. So after mating she will look for a blood meal. If that blood meal just happens to be taken from a person that has the parasites, she can acquire basically the disease, the parasites. There's a development stage inside the mosquito. That takes time. That time is a function of temperature. So the parasite development inside the mosquito occurs faster if the temperatures are warmer. Then at some point, when that process is complete, sporozoites will basically be in the sliver glands of that female mosquito. If she then takes another meal after she lays her eggs, she then will basically take another meal to develop the next batch of eggs. She can then pass on the parasite to a second, uninfected human. Temperature sensitivities. Well, I've already told you one, parasite development. Two, the egg development inside the mosquito is faster at warmer temperatures. If it gets too hot, the mosquitoes die more quickly. So their lifespan is a function of temperature. So if it's very hot, they don't live very long. If they don't live very long, there's not much chance of them biting a second person. The egg development rate is also a function of the pond's temperature. The larvae mortality rate is also a function of pond temperature. If it's very hot, they die more quickly. I learned quite recently that also the fecunditif, basically the number of eggs, is also a weaker, but it's a function of temperature as well. So the actual number of eggs that the mosquito lays. Without water, basically they don't have... So some of the species of Aedes that basically are important for riff valley fever transmission, they can sit kind of dormant in dry conditions for a long period, even years. But this is not the case for most of the Anophiles. So if the pond dries out, the larvae die, and that's the end of it. So if you get...we talked about sub-seasonal variability in rains. And then we gave the example of agriculture. Two seasons with the same rainfall, one with sporadic rain, intense. For malaria transmission, that's bad. Obviously for us, it's good. If you have just small amounts of rain all the time, you expect to have much more transmission because the ponds don't dry out in the dry periods. And there's also...I don't want to go into too much detail, but if you have intense rain, it tends to flush out the sites. So what happens is that the ponds you have overflow, and the small larvae, especially the stage one larvae, basically just get flushed out onto the ground, and then they just dry, desiccate, and they die. So desiccation is an issue. That's why you see sometimes even in very small ponds larvae as well, but those larvae often don't survive because those very small ponds don't survive long enough, they'll dry out. Only if they're in a larger-scale depression will they keep water during the rainy season. So basically that's just reiterating, as I said, the water stage, temperature sensitivities, but I'll swiftly move on. This is just basically showing, I'm going to go through all the detail of the actual development stage, but it's just showing the cycle. The thing I wanted you to just emphasise in this slide, the reason why I include it, is when you get bitten by an infective mosquito, it's not 100% sure that you will actually acquire the parasite. It's actually quite difficult, this transmission process. So the probability of acquiring the disease is on the order of about 20% to 30%. It seems, but again, data is quite sparse. It seems that if you have high immunity, there's also a kind of blocking immunity that your body is able, when you get the first inoculation, to actually combat it straight away so it doesn't actually lead to a particular case, you don't have the manifestation. So it seems when you have high immunity that percentage drops even lower. That's quite important. I'm going to come back to that in a moment. That makes the transmission statistics, should we say, quite non-linear. It's the same for the mosquito when it bites you and you basically have the disease. Again, it's on the order of 10% to 20% the chance of transmitting it to the mosquito. So there are quite a lot of loopholes that the disease has to basically pass through. And I've put the thresholds here. So the threshold for the parasite is on the order of 16 to 18 degrees and it's species dependent. So farcipurum, for example, the threshold is on the order of 18 degrees. Vivex, it's lower on the order of 16 degrees. That's why when there used to be malaria at higher latitudes, the malaria transmission in Europe was normally Vivex. Farcipurum only got as far north as Sicily, for example. So in the north of Italy, it was mainly Vivex. But in the south of Italy, you had transmission of both Vivex and Farcipurum because the temperatures were high enough to sustain. So those temperature sensitivities are basically species dependent. So to give you an example, I'm just going to take two of those temperature effects. Just two of them and show you how they combine to give you a transmission curve as a function of temperature. So on the left-hand side here, this is basically showing on the X axis we have temperature ranging from 16 to 40 and this is basically how long the parasite development cycle takes as a function of temperature. So you can see as up near the threshold 200 days where there are not many females that last that long. Most of them die much, much, much more quickly than that, on the order of one or two months. But as you get above 28, you can see we're down to on the order of the range of around five, six, seven days. Again, this is kind of a generic curve. I think this is for Farcipurum. I should have checked which one. I think this is Farcipurum. But if you basically... No, sorry, this is Vivex because it's going below 18 here. So this will be slightly different for Vivex and Farcipurum. On the right-hand side, we see basically temperature again and this is the survival rate. So you can see that the curve is roughly flat at 90% daily survival rate until you reach the mid-30s where it drops off rapidly. So if you combine those two, this basically shows you the population of mosquitoes that live long enough in order for the parasite to complete its development cycle so they can pass it on to a second human. You can see at the left-hand side, it's basically zero because it's too cold, the parasites are just not developing fast enough. On the right-hand side, it drops to zero because even though the parasite is developing quickly, the mosquitoes are dying even faster. So you get basically this sweet spot, so to speak, where you have this maximum on the order of 27 to 34. But this is just two effects. Once you add in the water temperature sensitivity, it starts to shift this curve to the left. So this is just, as I said, highlighting two of these effects. This is showing you the relationship. This is for a village in south-west Newgea. It's from a paper by a Bomberleyser Tower. The blue line is showing the seasonal cycle of the precipitation in this particular location and the dashed lines are showing malaria cases. Each line is a separate year, if I remember rightly. 2001, 2002, 2003. I think this was the mean rainfall for those three years. You can see how transmission ramps up following the rainfall season, but with a delay. So why don't we have that delay? Because you have these cycles that have to be completed. It's like a process that has to spin up. When the rains start, first of all, you start to have the mosquitoes breeding, building up their numbers. Then you've got the development cycle of the parasites. They've got to pass it on to a second person. So the whole system has to spin up. What is this delay a function of? Temperature. So the warmer it is, the faster the system can spin up and the smaller that delay is. That delay is also a function of temperature. So fighting malaria, where there's a whole host, and this list is not complete, I'm sure, of methodologies in which you can combat malaria. So you can distribute nets. People sleep under nets to protect them from being bitten. Those nets can be treated. So mosquitoes will basically die if they come into contact with the net. You can basically residual spray people's housings with insecticide. There's improved diagnostics. I mean you have a whole thing with improved drug access. The ones on the right tend to be more long-term interventions. Housing improvements, healthcare, land management. And it's quite interesting. Again, you can give a whole lecture on how our approach to malaria has changed over the last 100 years. It's gone through phases. So if you go back to the 1930s, the emphasis was on land management, drainage schemes. A lot more of this kind of infrastructure approaches. Then suddenly we invented DDT. So in the 60s, there was this big move towards eradicating malaria. It became much more attack in the vector, spraying breeding sites and so on. That found that basically in 1970 the whole programme was wound back. There was a big rebound. Now basically the key, these are why these two are in red, the key way the money is spent is through bed net distribution and residual spraying. And it's still, I mean I'm not an expert in policy, but you really see that there's a disconnect between the long term and the early. There's not always a long term. I remember when I was talking in Uganda about, is there any emphasis on housing improvements? And basically they said, no, why would we do that? And there's definitely the policies. There's a top-down approach from the WHO where they advise how to tackle the disease. And yet there have been studies that show a bed net cost $7, but for $20 you can put screening on a house. It lasts longer on the order of like 8 to 10 years and it can be cost-effective. So we're not using all of the weapons. It's a little bit like having a football team and you only send three players on the field in some ways. I'm not saying bed nets are not effective, but I personally, personal opinion, this is not the opinion of ICDP. It seems that, and again I'm not an expert in this at all, but it does seem that we're not using our whole arsenal of weapons against the disease. This gives you an idea of how things have turned around since the turn of the millennium, where one of the big goals was to tackle this disease and you can see how bed net coverage has ramped up across this as a massive success. The penetration of bed net distribution. Most countries in Africa now are massively increasing the coverage of bed nets through distribution, working closely together with the Global Fund, the President's Malaria Initiative and so on through the Ministries of Health. I'm going to skip over that one. I'm running a little bit late. So modelling malaria, how would we want to actually model malaria? The key thing you see with application models is they, unlike, we talk about diversity between climate models and so on, but underlying the climate models is essentially those same set of equations. You have the equations of momentum, you have your thermodynamic equation and so on. The differences come down into how you solve the numerics of the advection, how you basically parameterise all those sub-grid scale processes, as we emphasised last week when the model uncertainty. The situation is different when it comes to application models. You have a much wider array of modelling approaches. So as we saw, we have the statistical kind of approaches. You can use CPT to relate predictor variables to malaria cases. So you have models which are basically statistical models, which can be quite complicated, incorporate for example data as well. So the map analysis at Oxford is a basically a basing approach. You can think of it as a kind of statistical simple model, like a data assimilation that incorporates surveys to give you an estimate of the intensity of transmission. You have box model approaches, which I'm going to talk about next, which basically focus on the disease transmission in humans, but are not incorporating the climate. Then you have a number of dynamical models that try and represent the processes of transmission through basically dynamical equations that you integrate forward into time. There are not that many. There are more than I have listed here. I'm just listing some of the key ones. One of the ones is the Liverpool malaria model. That's a spatially distributed grid point model. That tries to account for the rainfall and temperature drivers of the disease. It doesn't account for population. So each grid point has 100 generic people in it. So it doesn't account for changes in transmission as a function of population. Open malaria is basically the model that's developed at the University of Basel. That was with Gates funding actually. It's very successful in terms of the way it models interventions. They're really focused on your model of point location, but they're really trying to represent how bed nets, spraying, drug distribution and so on affect that transmission. Again, it focuses on the human. So there is no climate in this model. This is the only other open source model I'm aware of. All of the other dynamical modeling systems, basically you can't access the source code. So I was motivated in these projects to, I wanted to actually try out, should we say, some ideas I had about the population density interactions. And I found there was no model that basically accounted for climate, which was open source. So I wrote mine. Before I do that actually, sorry, I did want to, I put this slide back in just to show that a lot of the models and modelling approaches have what's called an SEI or an SEIR approach where essentially you have four compartments where people, if they're uninfected, they're susceptible, S. And you basically have a rate of change between these boxes where if you have a certain biting rate, you'll have a certain rate of moving people from the S box to the E box. Then they have a development timescale around 18 to 20 days. And so you have a rate of change from basically exposed into infecting and if you have a representation of immunity, then some people after having a clinical manifestation will be moved into the recovery box, whereas some won't develop immunity and will be just basically moved back into susceptibles. So there's a very nice review paper, if anyone's interested in this approach, that really summarises all of the SEI models that have basically been put together over the last 10 years or so. I've forgotten to list it here, sorry. So I wanted to basically develop a model that accounted for temperature and rainfall impacts on the disease. I was interested in inter-annual variability, the predictability of the disease. But I wanted to improve also the representation of the surface hydrology because the Liverpool model basically just has a linear relationship or that has been modified, but they just basically directly relate rainfall to eggs vector density. So I wanted to improve the surface hydrology, account for population density and be able to go down to resolutions on the order of 5 to 10 kilometres in terms of my grid box size. If you go finer than that, once you get down to a kilometre, then you have to start worrying about vectors moving from one box to another. And I didn't really want to worry in the first instance about vector motions. OK, so this is vector in a nutshell. I'm going to show a couple of other slides, but this just gives you an overview. So it's a dynamical model. I'm going to explain this graph in a second, but it accounts for the temperature effects on those key aspects of the disease transmission. It's got a simple pond parameterisation, so I'll explain that in a second. But essentially, in a nutshell, I was working for a long time on, for example, cloud cover parameterisations. We talked about parameterisations last week where if something is occurring on a scale that's smaller than the grid box, you can't explicitly model it. So if you have a model with 100 kilometre boxes, you can't model each individual cloud. So you try and model the low-order statistics of those clouds. So one, for example, aspect would be the cloud cover, how much of the box is covered by the clouds. And that can change in time. So you try and model it. You perhaps have a prognostic equation that influences that cloud cover. On the upside down, I have a pond parameterisation. I can't model explicitly these five-metre-scale ponds, but what I want to model is how, in a statistical sense, the coverage of these breeding sites changes in response to rainfall in a simple way. And on the left, the model has population density. The population in this version is static. The people are not moving around, but we do have population density, which allows the model to represent to zero order how the population density is increasing. This is the biting rate. And basically the blue is a collection from reviews of many papers of how the bite rate drops as the population density increases. And the red lines are different ensemble model simulations with the model in East Africa and West Africa. You get this zero-order effect of the population increases, decreasing transmission intensity. Now, of course, there are a whole host of other reasons why population density increases also impact on an area. You have better access to medical care inside cities, better drainage, less breeding sites, possibly. So the background of the model is it simply tries to resolve these effects because I'm interested in predictability on seasonal timescales. So I want to get that delay as best as I can. So you have, for example, an array of boxes. So you have an adult. She lays eggs. So I just simply model the progression of the eggs forward in time. When they reach the end box, they basically hatch into an adult. She looks for a blood mill. And then the eggs, I model the way the eggs are developing. It's gone a traffic cycle. So this is that egg adult, should we say, progression. So how does rain actually affect this? Well, rains drives my pond coverage parameterisation scheme that basically gives me a breeding site availability. That relates directly to how many larvae I can support through a biomass limitation. And so that restricts basically the density of larvae that I will have. So immediately you can see some of the problems you're going to have here. How do you fix these parameters, the biomass limitation? We've been at most one or two surveys that have tried to measure this. That's in one location. Of course, you have a huge variability in food resources. It depends. Do you have pollen falling on your pond? Is it away from vegetation? So there's a huge variability. It's very difficult to set these parameters. So you can see immediately that with these kind of dynamical models, you're going to have quite a large uncertainty with some of these aspects. That's right. So they're integrating forward in time. So if you imagine if this were my clouds and I had been resolving my physics scheme, these would be my small cloud droplets and then you would have an integration equation that showed basically models how they grow by diffusion. So they would move. So this is like a bin resolving my physical scheme in clouds. It's just these are larvae and the equations are, in fact, a little bit simpler because I'll show you in a moment. So basically I'm stepping the time steps forward one day at a time. Now that's actually a very good question because some of these processes, when they're getting down to five or six days, one day time step is getting on the edge of actually resolving these. You have a little bit of time truncation problems and that's another issue of these models in the new racks. But it's difficult to go down to sub daily time steps because you don't always have the data. You know if you have daily rainfall, you want to drive the model of that. How do you subdivide that? I have been thinking of trying to perhaps subdivide with a temperature dyno cycle and there are a couple of people that say that the dyno cycle, because of these non-linear models, it does have an impact. For the moment for simplicity, it just integrates forward at one step at a time, one day at a time. And as I said, all of these relationships are both the mortality and also the actual progression rates are temperature sensitive. But the mosquito actually has two characteristics. So you can imagine for the mosquito, for the larvae, we just have the development stage. For the mosquito we have two characteristics. So what I mean by that, well for me we could model everybody in the room, we could have two numbers, our height and our weight. For the mosquito what do I care about? I want to know what the development stages of the egg. And I also want to know what the development stages of the parasites inside the mosquito if she's acquired the disease. So when a blood mill is taken, there's a certain probability of transmission of the parasite if it leads to infection. Then basically I also model in this direction the parasite development rate. So you can imagine starting off here, you take a blood mill, so you've got the protein to develop the egg. So over time she moves to the left. Now these are actually static boxes. I'm saying in each box it's how many vectors per square metre have this characteristic. We're moving from left. So it's not a Lagrangian sense. But if she's acquired a parasite, as well as moving to the left, she'll also move down through the grid at a diagonal because the parasites are developing. When she reaches the bottom row, the colours change to brown because now she has the sporozoites there ready to infect another human, which is why this is red for danger because when the blood mill is taken here, this is when transmission to basically a susceptible human can occur. Now it's actually a distributed model. So it's like you can run it for one point. In fact, when I run the lab classes used in this model, I just run it for one point. When I run the lab classes used in this model, I just demonstrate it for a single location because it's very fast to run, but you run it distributed. So each box is driven by the rainfall and the temperature. At the moment, there's no communication between those boxes though, so you can imagine it. It's like lots of single cells. So I'll come back to that at the end because the biggest key connectivity is not the vector movement. It's our movement and we move around more and more and more. Exactly, very good. So the whole reason I want to model is that these are the two characteristics that are temperature sensitive. So she moves to the left. If it's warmer, she'll move faster, further in one day. Same with the parasites. If it's very cold, we'll just very slowly move down. Each day we'll only move down, maybe just a fracture will move down one box. So these are temperature sensitive. And so the mosquitoes will keep cycling. That's a said because they lay the eggs and then they basically go off the left, these blue arrows, and back onto the right and take another blood mill. Eggs develop, lay the eggs, another blood mill. And this keeps going until of course eventually a mortality event occurs. And that mortality is a function of temperature. The bottom row here is basically if there's transmission, this is my human array. So these are my susceptibles. These are people without the disease. If they acquire the disease, then they start to move to the left. This is basically the development of the disease inside the host until they reach here when the person is basically dangerous because they can transmit to a vector and they may have manifestations, clinical manifestations. So essentially I've got an array of boxes there. That's the only part of the model that's not climate sensitive because of course we're warm-blooded. But I have an array of boxes, nevertheless because I want to get that 18, 20 day delay in order to get that spin upright. OK. So you can imagine this is basically that one of these SEI models. I don't have R at the moment because I was trying to think of the best way to put in the immunity. Immunity, as I said, is not that well understood. I think I'm going to put a simple immunity. There'll be an R box at the end. I'm just going to do it very simplistically. So the next version actually, which will be released early in the next year, will have the immunity as well. At the moment there's no immunity in the model. I'll go very quickly just through these. I'm looking at these. The main message is these relationships are very simple to a great extent. For example, the development rates are often just a degree-day effect. So you have a critical temperature and you basically have, if you're two degrees above this temperature then the process takes half as long. If you're four degrees or quarter as long compared to being one degree above that temperature. So the relationships are quite simple but the parameters like the K can vary greatly and there are very few field studies that allow you to set these parameters. Okay. I've tried to set the model up as simply as possible to be able to change all these. Everything's in a kind of a nameless. And the idea about that is I want to make it as simple as possible to be able to explore parameter space with the model and set it up for larger ensembles. So this illustrates some of this uncertainty. This is temperature on the X-axis and this is the survivability and all of these different lines and dots are all different parameterisations and measurements of the mortality rate as a function of temperature. Now, even if you look in the middle of the graph and you say, well, this is not a great difference but if you actually integrate that up this is the daily mortality rate. So if you look at the expected lifespan the difference between 95 and 90% is huge when you integrate it forward in time. Okay. I'm going to skip over the hydrology fairly quickly just to show you. It's just a very simple. You have a catchment fraction that you have to set which we're now trying to set according to the slope and the topography. We don't even account for the soil texture at the moment. Okay. So it's an extremely simple scheme that's integrated forward in time. We have tried to validate it both with in-situ measurements in Ghana. There's a paper in press now and another paper that actually compares it to explicit simulations. When I say explicit this is an extremely complicated model at 10 metre horizontal resolution for a whole village in Niger and it's basically all of the overland flow, sub-flow and everything. So the lines here are just showing our simple model where we've calibrated one parameter to get the overall magnitude. It doesn't affect the sub-seasonal. So the sub-seasonal variability is the same but the absolute magnitude to get the fraction right we had to calibrate one parameter within like a 30% range. But it's just showing one is a massively complicated 10 metre explicit assimilation and the other is just my palm fraction parameterisation with that one simple equation is actually slightly different to this now. This is the 1.2.6 version. We made some small improvements. Just showing how this is two seasons, two rainy seasons, how you can capture a lot of the variability in a very simple model of just one parameter actually. I was actually amazed at how well the simple parameterisation did. I was really surprised. And this just was swapping out the station data for a satellite. The biting, very quickly, remember I told you that the probability of actually really getting bites is not 100%, it's only 20%. Why is that important? Well the thing is you have distributions of bites. Even if we were all equally attractive to mosquitoes which we are not because some of us in this room are more attractive for various variety of reasons which are not all fully understood. One of them is just body size so I'm at a slight disadvantage because I'm taller, but I give off perhaps a little bit more CO2. But even if we were all equally attractive if there were one mosquito in the room per person we wouldn't all receive one bite. Just through random fluctuations maybe Andy would get three bites and I would escape just through random fluctuations. The model just starts off with that very first simple assumption. We just simply assume that the biting process is random so it's under-dispersive compared to reality. So the distributions of bites per person is basically puzzle and distributed. That's important of course because it's the mean transmission. Andy and I get one bite each and the probability of transmission is half so there's a probability of half that I get the disease and he's at half so in average one of us would get the disease. If I get both bites and he escapes then my probability of getting the disease is basically 0.75 because it's half, instead of 1 so you see that basically affects the mean transmission because it's one of the non-linearities in the transmission. So this shows you the climate sensitivity of the model, this temperature on this axis and you see this is an older version of the model so now actually it's reduced here due to flushing and we have increased transmission here but it's showing again that transmission peaks in this sweet spot of around 25 to 32 degrees while the transmission again is a function of population density I'm going to skip over that I'll skip over this as well because what I want to do now this is just a comparison to map but what I want to show you quickly is some of the uses that we've applied the model to and I just want to show you one example because that's relevant here of the S2S timescale so I've actually predicted so we've applied the model so far to try and look at seasonal forecasting and we've also done a few other things that I just want to list here in case you're interested so we've looked at trying to simulate transmission in Uganda in the early 1920s and 30s so in the pre-intervention period to try and see if climate could explain the variation there we've looked at climate change problems land use change impact on malaria over longer timescales and we're also working now with quite closely with IRI on aspects of basically in these models with Madeline Thompson so again just to go into that finance again I gave you a figure before this shows you how as a function of time domestic and international funding increased but it's essentially flattened out again to justify why do we want to use climate funding is basically plateaued and there was even a funding round with a global fund that was actually cancelled due to pledges of money not actually arriving in time to disperse those funds the first ramp up in some ways is easier if you want to say okay nobody's got a net you can just do a mass distribution and you know those nets are going to be distributed everybody needs to have a net okay there's a lot of questions and discourse at the moment how we actually then maintain that coverage you know do you try and target how do you know who needs to have their net replaced where your most vulnerable populations are it's much more difficult you don't want to just do a blanket dispersal because maybe a lot of those nets won't be required you're wasting a lot of money climate information therefore you may be able to have cost effective prioritisation of where you target some of those intervention processes that we were talking about advising people to retreat their nets because you expect this year to be an epidemic season for example sending out spray teams earlier if your rain onset is going to occur earlier so what we try to do is we try to look at the potential to set up a seasonal forecasting system so this has been tried already with the Liverpool model to a certain extent one of the things they didn't do which we try to do here is their initialisation of the malaria model was done in a way they basically started a little bit like climate models are started from an arbitrary state and then they basically spin up the model with the early stage forecast situation the problem with that is you don't have any information about what's gone on before if I want to start a forecast today if it's been raining more than usual my past month I'll have more breeding sites a higher vector density perhaps a higher parasite ratio because it's quite a slow evolving system the information that's occurred before my start date is really important so what we try to do is the point is none of these surveys are timed so we had to basically try to set up an analysis which was climate driven so we used climate observations or the reanalysis to drive the malaria model up to the start date to ensure that we have reasonable initial conditions I say reasonable because it's very difficult to actually validate if these are correct or not there's no real time information about vector densities and then we basically use climate weather forecast and climate forecast to drive the vector model to give us our forecast of hazard so that's the definition of system so what you've been using this week when you access S2S is this monthly EPS so we use that as well we use that in the first month this was first of all using out of 32 days we can now extend that to 48 days with the new system and then we basically slot in the seasonal forecast from month 2 onwards out to month 4 so we're using the model system that you've been analysing these two weeks in the workshop we've got one member of the real time so each one of those drives a vector model and so on so we mask we want to focus on areas with high variability I'm not going to go into details here but we're looking at how the variability like PDFs are trying to separate endemic and epidemic areas so this is like a mask of a mask has been applied here okay and then what we've done is we've looked to the statistical skills and then we've shown areas where we have high inter-annual variability not the endemic areas so all the grey areas it's not because the forecast is not skillful or it's not raining then it's simply that there's not strong climate driven variability according to our assessment so what I've shown here it's like an RGB plot the red means we have skill at this lead time in temperature blue is rainfall for example on the map can you see white dots it means that we have skill in temperature rainfall and also malaria predictions and this is validating against the reanalysis driven run so it's assuming the malaria model is perfect at this stage so it's just a test of how the predictability should we say in rainfall and temperature would translate into temperature whether the model is acting as a nonlinear operator so you can see that we have lots of yellow points but even at lead one a lot of the points are not just white but they're yellow so yellow means predictability in temperature and in malaria but not in rainfall so you can see even at lead one some of you have probably seen this already with the S2S systems that the temperature predictability extends out to a month or more but the rainfall drops off much more quickly this is one well this is actually to be strictly accurate this is you call this a two week lead because it's reading the first month for the integration is that those first 32 days 32 days so it's going from lead zero to lead 32 in days so the average lead is two weeks so that is sub seasonal so this is your sub seasonal predictability the nice thing is you don't have any black spots black spots would be no rainfall no temperature I should put a thing in the key there now what this is showing is month one, month two, month three, month four, lead I'm just showing January to June so you can see the progression towards the colours towards a lot of the areas having the black spots which means we have no more predictability four months ahead now what do you notice about this in the first month we have lots of yellows in month two, month three we have lots of points which are green now green means we have predictability in malaria it's skillful in malaria but not in the climate parameters how can that happen how can we have a two month lead how can we have predictability in malaria but not in climate if the temperature and the rainfall are incorrect how can we have the malaria correct one might be if the rainfall has occurred or the climate has occurred in the previous month and they're lagged it's a big malaria in month two it doesn't depend on whether in month two whether in month one it might just be an initial condition thing you've got observations depth of malaria population and that provides something without any weather or climate exactly so it's basically as I said the system is slow to evolve so you've got a lot of memory in it and you have the lag so it's all down to those two reasons so essentially even if you didn't use the forecast system at all and you only observed basically temperature and precipitation you can still skillfully forecast malaria transmission one to two months ahead but by adding in the EPS and the seasonal forecast we're getting out to about month two to month three so this also indicates that our EPS your S2S system is adding skill but we're not really getting a lot out of the system far otherwise we'd expect to get some there are a few points especially in eastern Africa where teleconnections with ENSO for example tend to be a bit stronger here and so on so basically all of this pretty much is coming from the initial conditions so this shows you how important the analysis system is that we've put in so basically all of this pretty much is coming from the initial conditions which hasn't been done before this is partly from the initial conditions but partly from the S2S in fact that's a test that we should do we could actually take out what we could do is actually what I've set up because you don't want to just average so what I've often been doing is my experiments where I shuffle the years in the wrong order so you have the right sub-seasonal characteristics of the rainfall but they're in the wrong year test, thank you very much ok, very quickly because I'm running out of time in my last five minutes for the hours up I wanted to show you very quickly some initial and this has been slowly evolving I was hoping this would be written up and going into the pilot stage by now but hopefully next year fingers crossed for some different funding for next year but essentially we've been trying to analyse the system in various countries we've looked in Malawi and Uganda and I'm going to show you the Uganda because that's the place where we got the ffervist mainly through the fact that we have the best working relationship with the Ministry of Health there it's very important to have these strong interactions the other thing that's nice about Uganda is as well as having the district level data they're also high quality called sentinel sites where you have high quality data where every single suspected case is tested either through a microscope test or the rapid diagnostic text kit 100% confirmed and they span from basically the highland areas in the south west to lowland endemic areas what you'll notice is and this is one of the key problems is that the time series of the health data tends to be very short so most countries in Africa move towards digital databases in the early 2000s a lot of the paper records before either haven't been digitalised or in the worst cases have actually been lost or destroyed it varies from country to country but there are very few countries what's wrong for example is one exception where they have longer term records of malaria cases and even then to the external research community they're only available on a country level and the sentinel sites is even worse they started in 2006 or 2009 but just to give you an idea this is like a comparison to the sentinel site and we're just showing normalised anomalies of the log of the EIR which as I said relates to cases the red is the confirmed cases and you have the spread is indicated by the grey so what you'll notice is that the model usually and again caviar of the short time series and I'm going to show a couple of examples it tends to get the interannual variability but if you look closely in the details you'll find that if you're looking at individual months you can be quite a long way off but you'll find that for example that it was an end of the year in 2010 and this is lead one, lead two, lead three and lead four you get the variability so what I want to show is just how we're doing so if you go to Canunga Canunga is interesting in that we actually have this double peak and you can see the forecast system picks up that double peak at the sentinel site as you go to higher lead times you can see that time intends to be off but it picks up this double peak so this is not ENSO driven so this is Mubendi I'm running out of time so I'm going to skip over some of these and you can see that it has a single peak it's picked up but there's some problems with the timing if you look at the skill for five of the six sentinel sites you basically have should we say that the skill out to four months in advance the exception is Kabali which is too cold the model doesn't stimulate so there are 15 degrees there on average there's no malaria simulated there but you have cases so that's quite interesting if you think about it normally I would ask you why that would be but I'm just about out of time so I'm going to have to just tell you I mean if you think about it if you have a single station you have undulating topography so that station might measure 17 but you only have to move somewhere where you're a little bit lower down within the catchment area increasingly a busy trade route between Uganda and Rwanda so there's a lot of traffic passing through people passing from low high transmission areas through the area often staying overnight in the area they can then basically carry the parasite with them you can have secondary transmission that way but nevertheless in my opinion to my best knowledge this is the first ever demonstration of a skillful malaria system has been tested on a subnational scale in this way it's all smooth those were smooth this is how the raw monthly figures look like so you can still see that double peak for example showing up but if you look at the month by month figures so we really need to think about how we're going to portray these forecasts that was the central data when you start to look at the district data things can get quite different this is just two districts I've picked out where you're looking at this at a longer time series it can look very good the forecast in some districts there's no resemblance at all to what's actually going on so of course some of that is model error but some of that I think is basically data inaccuracy as you really see you can have two small districts side by side very very disparate changes some of that can be the actual situation layer of the land interventions and so on but it can also be data anomalies so if you look at a map this is a kind of a typical should we say the coloured districts are the ones where we get a significant skill which are white and we're not what I was surprised about here actually was we actually got a significant skill even in endemic areas where we didn't expect to have so much climate driven variability so where are we I'm going to wrap up because I've overrun a little bit but basically we have this pilot system we've tested it out on a sub-national scale in three countries Uganda, Rwanda and Malawi and I think the forecast there anyway if you look at the weather forecast they tend to have a lower skill in terms of the temperature rainfall that could be part of it in Rwanda the system was doing actually very well especially in the eastern end of the country and Uganda we can see that against the central data it's looking pretty good although the timing's off we've still got to do a lot more work so I'm hoping next year especially if we manage to get a little bit of time for this to actually take this a step further and start to look at policy in the country and how you would integrate it so there are lots of open questions I don't have time to discuss this we've been looking at ways of trying to turn this into like up a terse and up a 10 percentile we talked about this last week with the other applications early on in the week with the flooding how these products made Frederick was saying they kind of made a product but there's a kind of interactive process where you can try it out on the eastern WF site for a while there was a pilot system looking at terse maps a little bit like they use for the monthly system where you can also mask out the endemic areas how do you bring in vulnerability assessments I don't have time to talk about that but there's a whole array of like how do you help people to actually target their interventions I'm of the opinion that you can't do too much because the experts and the people the district health officers are on the ground here are breeding sites lower lying area they tend to have higher transmission when there's transmission compared to another area they probably have a very good idea already within their districts how to apply a broader scale information a lot of people say wow you can't use a forecast system like this because I know that malaria transmission varies from one village to the next on scales of a kilometer so I'm into mind about how much you want to go down the route of this trying to do this very high resolution project so I'm going to stop there I wanted to show you in a nutshell the applications model we've developed trying to get across to you the level of uncertainty associated with some of these parameterizations in these schemes which is not just climate models that have uncertain parameters and then when you're parameterizing your convention or your clouds and so on you're often setting these parameters by one lab study that was conducted in 1962 so there's a massive need for data to improve our understanding of these disease transmissions we need to improve the way these things are calibrated and remember that a model is a tool but I think we've demonstrated so far that it's potentially a very useful tool we just need to learn how to actually use that tool for policy I'm going to stop there but I'm going to stop there