We live in an increasingly interconnected world. The rapidly increasing movement of people, pathogens, vectors, livestock, food, goods, and capital across borders creates both economic opportunities and health risks 1 . Epidemics, just like climate, do not respect national borders and can threaten human health and social stability. Since the Millennium, the appearance of the SARS coronavirus in 2003, the avian influenza (H1N1) in 2009, the Ebola virus in West Africa (2014-2016), the Zika virus in the Americas (2015) and the novel coronavirus identified in late December 2019 in Wuhan, China and still ongoing (2020), amongst others, has demonstrated the speed at which emerging infectious diseases can spread with devastating effects 2 .
Many infectious diseases are climate-sensitive; climate acting as an important driver of spatial and seasonal patterns of infections, year-to-year variations in incidence (including epidemics), and longer-term shifts in populations at risk 3 . Climate impacts both the virus and the vector. Evidence to date shows that arboviruses of global public health importance, including Zika, dengue, yellow fever, chikungunya, and Rift Valley Fever, have mosquitoes as part of their epidemiological cycles. Some Aedes-borne diseases have experienced an overall increase of 30-fold in the last 50 years, causing more than 50 million infections worldwide every year 4 . In the United States of America, the vast majority of Aedes-borne infections is imported from endemic and often neighboring regions -like the Caribbean, Central and South America-by travelers who become potential new sources of transmission. Autochthonous transmission in the continental USA has been already observed for chikugunya virus (2013) and Zika (2017), and risks are likely to increase with anthropogenic global warming.
For authocthonous transmission to occur, the population needs to be susceptible to the disease, but there must also be suitable environmental conditions (e.g., suitable temperatures) for both the mosquitoes and the virus. Since the susceptibility can be periodically monitored via targeted campaigns, the environmental suitability for presence of mosquitoes and viruses is one of the most important pieces of information for decision makers in the health sector. Moreover, the transmission rates or the number of cases are generally more difficult to forecast than environmental suitability, due to their link to a larger number of (often entangled and more complex) predictors, involving human behavior and socio-economic conditions.
A generalized approach to model Aedes-borne pathogens is needed because multiple Aedes can serve as vectors of dengue, Zika and chikungunya. Although Aedes aegypti is the most common vector, Aedes albopictus (otherwise known as the Asian tiger mosquito) has been identified as another important additional vector because of its vector competence for several arboviruses and recent rapid spread 5 . Both vectors pose a potent threat to global health security given their ability to transmit a wide variety of emerging and re-emerging arboviruses for which there are no vaccines. Aedes aegypti and Aedes albopictus are ubiquitous in large regions of the Americas and the Caribbean.
Historical, current and forecast climate information can be combined with disease models to improve climate-sensitive health planning and targeting of resources. For infectious disease models, the goal has frequently been to explore different interventions scenarios in order to help set priorities for policy makers 6 . However, in recent years there is increasing interest in using models for real-time forecasting [7][8][9] , although there remains a significant gap in the operational readiness of the numerous forecasting systems presented in the literature 10 . Stochastic models possess inherent randomness and are widely used in climate science as well as in disease modelling to build probabilistic forecasts 11 , as they provide a more reliable assessment of the range of likely outcomes. However, probabilistic models are sometimes harder for decision-makers to interpret, and tend to be rejected in favour of simpler, deterministic, but over-confident, models. An approach that takes full advantage of both deterministic and probabilistic forecasts is presented and discussed in the following pages.
Although the historical (average) seasonal behavior -and similar statistics-of these diseases is useful 12 , we consider it not enough for decision-making, as inter-annual variability (e.g., related to El Niño-Southern Oscillation) tends to play an important role in the actual observed variations of Aedes-borne diseases, enhancing or reducing the associated risk 8,[13][14][15][16] . Hence, a formal forecast system and its associated skill assessment is required and -to the best of our knowledge-is still nonexistent for the continental United States and its transboundary regions.
Here, we describe the AeDES (Aedes-borne diseases' environmental suitability) system, a new pattern-based calibrated, multi-model ensemble of climate-driven Aedes-borne disease models for North America, Central America, northern south America and the Caribbean, based on prior work undertaken in collaboration with the Pan American Health Organization (PAHO)/World Health Organization (WHO) 8,16,17 . We built AeDES using the same general approach for both the monitoring and forecasting sub-systems, which in addition to supporting surveillance operations, simplifies the forecast verification process.
We discuss the use of AeDES to inform concrete prevention and control strategies, using the recent Zika epidemic as an example.
AeDES uses multiple ento-epidemiological models to produce estimations of environmental suitability for transmission of Aedes-borne diseases, quantified via the basic reproduction number, R 0 (red box in Figure 1). We used the basic reproduction number to assess the environmental suitability of transmission of Aedes-borne diseases because (a) it is one of the operational outbreak indices used by WHO and several other decision-making institutes and health practitioners 18,19 , and (b) it has an intuitive interpretation in terms of the number of secondary human cases one case generates on average over the course of its infectious period (assuming a completely susceptible population) 20 ; hence, values smaller than one indicate that environmental conditions are not suitable for disease propagation.
Formally speaking, R 0 is an environmental suitability (or potential) for transmission, and not a transmission risk index itself; the latter depends on more complex interactions and the definition of the involved vulnerability. R 0 works both as a suitability monitoring index -when computed using observed variables, or when estimated by an authoritative organization such as PAHO or the Center for Disease Control (CDC)-, and as a forecast index -when using actual climate forecasts of the variables required for its computation. R 0 models require a set of ento-epidemiological parameters (green box, in Figure 1) and environmental information, either actual observations if we focus on the monitoring sub-system, or forecasts if we focus on the prediction sub-system (see blue box in Figure 1). Typically, R 0 models require near-surface (2 meter) temperatures, but other environmental variables are also involved, like rainfall or even humidity. Here, we use four R 0 models already described in the literature: the Caminade et al 21 , Wesolowski et al 22 , Liu-Helmersson et al 23 and Mordecai et al 24 models. For details see the Methods section. We use multiple R 0 models to be able to better assess uncertainties, and we calibrate each of the models independently before creating the multi-model ensemble to minimize systematic errors.
The AeDES monitoring sub-system offers maps showing the spatial distribution of environmental suitability over the region of study for the 1948-present period, at a monthly timescale. It also includes additional information to provide context to the user (Figure 2). These fields were included in the AeDES Maproom (https://aedes.iri.columbia.edu) after consultation with decision-makers at PAHO. To produce the environmental suitability maps (e.g., Figure 2a), each one of the four R 0 models was run from 1948 to present, forced by GHCN-CAMS temperature data 25 (∼56 km resolution) and ento-epidemiological parameters (see Methods), and then combined. The monitor sub-system is automatically updated in the AeDES Maproom around the 8th day of each month. These maps are useful to know the recent behavior of environmental suitability, or to conduct comparisons with respect to particular years. Trends and variability analysis, or the extension of the northern border of environmental suitability can be easily too computed with this new dataset.
The additional information, such as population density (Figure 2b), and social vulnerability (Figure 2c) is offered to the user to assess potential risk of transmission. Once a location is selected, the seasonality of R 0 , accumulated rainfall, minimum, average and maximum temperatures, and frequency of rainy days (Figures 2d-g) is provided. Our team is working on adding fields such as human mobility and connectivity, which local experts in the northeast of the US have suggested as also useful to analyze potential disease transmission.
As indicated earlier, the forecast sub-system employs the output of state-of-the-art climate models and the same R 0 models used by the monitoring system. Models, nonetheless, require statistical post-processing to help correct for biases with respect to the monitored R 0 values. Following Muñoz et al 8 , a pattern-based Model Output Statistics (MOS) approach approach using principal component regression (PCR) was applied to the raw R 0 models output. Since the R 0 models are the same (using the same ento-epidemiological parameters) the calibration takes care of climate-related model biases only.
A skill assessment was conducted for each calibrated R 0 model and the final multi-R 0 model ensemble (i.e., the AeDES model), focusing on discrimination as an actual measure of the value of a forecast system 26 . Although correlations between forecasts and observation are often used to assess skill, they only provide information of how in phase or not the forecasts are with respect to observations. The metric selected for skill assessment was the two-alternative forced choice, or 2AFC, which "measures the proportion of all available pairs of observations of differing category whose probability forecasts are discriminated in the correct direction" 26 . In other words, when terciles (above-normal, normal, below normal conditions) are used, the 2AFC measures how well the system distinguishes between the different categories; a system with poor discrimination is of no practical and economical value for decision-makers. Furthermore, 2AFC has an intuitive interpretation as an indication of how often the forecasts are correct 26 .
AeDES's predictive skill (as measured by 2AFC) is well above that of the reference long-term average (corresponding to 2AFC=50%), with values ∼1.4-1.8 times larger than that baseline basically everywhere in the region under study. Skillful regions extend farther north during the boreal summer (Jun-Aug, or JJA) due to more suitable areas for the vectors because of higher seasonal temperatures (see Figures 3a,c). Also, as expected, AeDES exhibits skill improvement compared to any of the models involved in its ensemble (Figure 3), which show comparable skill distributions among themselves, both in space and time. AeDES tends to outperform the individual models everyhere, but especially in the Caribbean (e.g., Cuba, Jamaica, Haiti and Dominican Republic) and in a lower degree in the United States Great Plains, southern Mexico, Colombia's Orinoquia and the northern Amazon in Brazil (Figure 3); it also outperforms its predecessor model for Latin America and the Caribbean, described by Muñoz et al. 8 , especially in summer in western Colombia, and in winter in most of Central America and the Yucatan Peninsula (cfr. Figure 4 in Muñoz et al. 8 ). Predictive skill of the AeDES system is especially high (2AFC ∼70%-90%) in most locations of Central America, the Caribbean and northern South America in boreal winter (Dec-Feb, or DJF), with "skill hotspots" in both boreal summer and winter in Guatemala, Honduras, El Salvador, Cuba, Haiti and Dominican Republic, Jamaica, Puerto Rico and some island nations in the Lesser Antilles (unfortunately the observational dataset used for calibration does not cover all of these island nations).
Regarding North America, the Yucatan Peninsula is one of the locations with highest skill, especially in DJF, a peak season for tourism, and thus increased human mobility. In summer, almost the entire Pacific coast of Mexico exhibits 2AFC values above 65%. Overall, predictive skill over the United States in summer tends to be higher in the eastern half of the country than in the western half (where orographic temperatures naturally tends to control vector proliferation in large regions), and ranges between 50% and 90% along the United States-Mexico border and the states along the Gulf of Mexico's shoreline. Forecast discrimination skill for southern Florida is also high in summer (values ∼90%, see Figure 3c). In northern South America, the Caribbean coast of Colombia, and northern regions of Venezuela, Guyana, Suriname, French Guyana and northeastern Brazil exhibit very high skill both in summer and winter.
Hence, predictive discrimination skill of AeDES is in general high, and decision makers geographically interested in the hotspots mentioned above can take special advantage of the enhanced skill of the system in these regions to improve their response times on key prevention and control strategies, at least a month ahead of the target season.
The risk of Aedes-borne disease transmission is in general very difficult to estimate, in part due to the complexities of accurately assessing the actual risk in terms of hazards and vulnerabilities impacting the target population. The general approach should successfully integrate the interactions between humans, virus, vectors and the environment, making it a very complex system to forecast, not to mention the fact that a lot of those interactions are not yet well understood. An alternative is to identify a predictand (variable to monitor and predict) that (a) enables decision-makers to take timely, "no-regrets" actions, (b) is verifiable (can be easily obtained from the information available or the health surveillance systems in place), and (c) can be skillfully predicted for the region and timescales of interest. The information provided to decision-makers does not need to be perfect, it needs to be reliable enough to make the best decisions.
Typical choices of predictands in the case of interest are number of positive cases and incidence. Although these options generally satisfy the criteria (a) and (b) mentioned above, skill tends to be a barrier to make the best decisions in a timely manner. Low predictive capacity for these predictands is related to different reasons, but often can be traced back to the fact that they depend on a variety of complex factors -e.g., socio-economic conditions, human behavior, human mobility, etc.-, some of which are (still) largely unpredictable. Previous work have argued 8 that a potential alternative is to focus on environmental suitability for transmission, since variables like temperature, relative humidity, vegetation cover and, depending on where, rainfall, are skillfully predictable at timescales decision-makers are interested in. In this sense, climate imparts predictability to the Aedes-borne diseases transmission problem if a predictand like R 0 is used as a proxy for transmission risk, even when clearly it is not representing the complete risk picture: additional information on the presence of the vector(s), the population exposed to the disease, and circulation of the virus is also needed. Recent work by Monaghan et al. 27 is using a similar approach to the one presented here to address the vector presence/absence component of the problem, and certainly both systems could be combined to provide additional information for decision makers in the health sector.
A large amount of work in the related scientific literature has been focused on developing or improving different R 0 models (see Van den Driessche 28 and references therein), but few efforts have addressed real-time R 0 seasonal forecasts, and no such operational system -to the best of our knowledge-existed until now for Aedes spp. in North America, Central America, northern South America and the Caribbean basin. Furthermore, to better assess uncertainty in AeDES, the approach followed here involves the use of not one but multiple ento-epidemiological models, forced by state-of-the-art seasonal climate models from the National Oceanic and Atmospheric Administration (NOAA) North American Multi-Model Ensemble project (Kirtman et
There is consensus on the need of including uncertainty information on any forecast that is produced 29 . One way of providing that information is to add confidence limits if the forecasts are deterministic (actual values of R 0 in our case). For an example, see Figure 4, sketching the expected value for the summer of 2016 (Figure 4a), and the expected standard deviation (or uncertainty, Figure 4b); for reference, the monitored (or "observed") values for the same summer are presented in Figure 4c).
Another way to provide information about the forecast uncertainty is via the use of probabilities to indicate how confident (or not) the system is that a certain outcome -say, above normal R 0 values-will occur during the next season. An example of a tercile-based R 0 probabilistic seasonal prediction, again for the summer of 2016, is presented in Figure 5a, where probabilities of below-normal, normal or above-normal R 0 values correspond to red, green and blue color shades, respectively. Although this is a very useful approach, and tercile-based probabilistic forecasts have been used for more than two decades now, decision makers often require information beyond the usual three categories described above. Using the entire probability density function (see Methods), AeDES also provides probabilities of exceeding particular thresholds of interest (Figure 5b).
To illustrate the use of both deterministic and probabilistic forecasts, consider the recent Zika epidemic in the Americas 30 .
Official CDC numbers 31 for Zika cases in the US indicate that both Miami, FL, and New York City (NYC), NY, reported slightly more than 1,000 cases in 2016, around 40% of the total number of cases in the US. Most of these cases were reported after the summer of 2016, a period of increased environmental suitability and human mobility (e.g., tourism to the Caribbean).
We will focus on these two cities in the following example.
By the beginning of May 2016, decision makers using AeDES would have expected enhanced suitability conditions for Zika during JJA in basically all of the southeastern US states, but also the Caribbean, most of Central America and northern South America (Figures 4a,b and5a), where several Zika cases had been already reported. Although it was highly probable that both Miami and NYC exhibited above-normal suitability conditions (Figure 5a), only Miami was expected to exceed R 0 = 3 (Figure 5b). In fact, the decision makers could have used AeDES to determine that most probably Miami would not exceed R 0 = 3.4 (Figure 5c), while NYC most probably would not exceed R 0 = 2 (Figure 4d). These probabilistic forecasts were consistent with the deterministic ones for both cities (Figures 4a,b), and by early September 2016 -once the actual summer R 0 values were available in the monitor sub-system-, the decision makers would have discovered that the forecasts were actually very skillful (Figure 4c). But coming back to May 2016, what those particular R 0 forecasts meant? Given an original number of 40 Zika cases and a generation time of 20 days (15.6-25.6 days; standard deviation of 7.4 days) 32 , an R 0 = 2 means that after four generations -each spaced ∼ 20 days-, or in about 3 months, there would be a total of 600 local transmission cases related to the original 40. Since R 0 is proportional to the duration of infectivity, an ideal action would be to reduce the infective period of cases, such that the effective reproduction number, R, is reduced. For example, in NYC, with an expected R 0 < 2 for JJA 2016, any combination of strategies to reduce the effective duration of infectivity by over 50% would mean an average R < 1, which should stop the spread of Zika over time. Beyond the obvious vector control strategies (for which knowing in advance when it is not going to rain could be useful), increasing traveler health surveillance, reducing the symptom-onset-to-isolation times, and the mosquito bite rates via specialized clothing and personal protective items can all help decrease the reproduction number. Economic costs for fighting the Zika epidemic would be most probably higher for Miami, given the higher R 0 value forecast for the summer.
AeDES is a "next generation" system because (1) it successfully tailors global climate information to be used at regional scales, (2) pattern-based calibration targeting mean, amplitude and spatial biases is performed using a monitoring system based on the same variable that is being predicted, and (3) it produces tailored deterministic and probabilistic forecasts for user-selected thresholds of interest, including the use of the entire probability density function (also known as "forecasts in flexible format" 29 ) to better assess uncertainties.
Previous research 8,11,16 has underscored the importance of analyzing climate signals at multiple timescales to improve decision-making processes in the health sector. In particular, Muñoz et al 16 to-interannual timescale tends to explain most of the total variance observed in climate variables impacting vector-borne disease transmission, like temperature and rainfall. Hence, although the long-term climate change and natural decadal variability signals also are considered, AeDES pays special attention to continuously providing actionable information at seasonal-to-interannual timescales, which along with the weather and sub-seasonal 33 scales are the most often used for health early warning systems.
Due to large uncertainties in long-term climate projections, the present approach should in general not be used in combination with climate change scenarios. Nonetheless, the same approach is adequate for shorter-term timescales, like the sub-seasonal (roughly 2-6 weeks 33 ) or weather (0-2 weeks) ones. Providing actionable information at multiple timescales (e.g., via he IRI's Ready-Set-Go approach 29 ). The team is presently exploring when and where predictive skill at these timescales is high enough to guide decision-making processes in the health sector, taking advantage of windows of opportunities in forecasts at those timescales 34 .
All analyses are conducted for the geographical domain defined by the coordinates 126 • W-40 • W and 1 • S-50 • N (Figure 1).
Rather than focusing on particular diseases, here we considered common environmental thresholds and ento-epidemiological parameters for Aedes-borne diseases as a whole. If the parameters are well known for diseases of interests, then the same approach can be used to have tailored information for those cases. For consistency with previous studies and model validations, we used the same ento-epidemiological parameters reported by Liu-Helmersson et al 23 , Wesolowski et al 22 , Caminade et al. 21