**7. ENSO and Health Forecasting**

Due to an improvement in the climate science community's understanding of the large scale mechanisms that influence climate, plus rapid advances in computing technology, seasonal to inter-annual to decadal climate forecasts have become a real prospect [2,97]. This, coupled with an increasing knowledge of the nature of climate-health associations, has spawned a number of attempts to construct disease early warning systems based on seasonal predictions of health-sensitive climate fields, so that potential health threats may be anticipated several months in advance.

A key source of the potential seasonal predictability of health-sensitive climate variables is ENSO. Given this, the hope is that with time, accompanied by an improvement in the understanding of ENSO health associations, effective seasonal forecasting of climate (ENSO)-sensitive health outcomes will become operationally possible [98]. Generally, two broad approaches have been adopted in constructing climate-sensitive disease early warning systems based on known climate and health links, namely numerical and statistical. Numerical schemes take the output from seasonal climate forecast models, usually in the form of a rainfall and/or temperature time series, and ingest this into numerical process-based disease models for diseases such as malaria and dengue (e.g., Liverpool Malaria Model, [99]). Typically the output from disease models includes disease parameters such as disease transmission, size of mosquito population, and disease incidence [100]. Statistical or empirically based forecasting schemes generally draw on a variety of statistical methods and use empirical observations of climate and disease incidence to construct transfer functions that statistically link climate disease associations. Although a simple distinction has been drawn here, between numerical and statistical/empirical models this does not mean to imply that numerical approaches do not draw on statistical methods and vice versa. In most cases, the output from both numerical and statistical models are probabilistic statements about the likelihood of a given climate sensitive disease exceeding a critical threshold and often statistical schemes, when run in forecast mode, will use the numerical output from climate models to force the climate-disease transfer functions so as to gain estimates of disease incidence. Furthermore, many dynamic disease models use statistical functions to model the relationship between disease sensitive climate variables such as temperature, and for

example, in the case of mosquito borne diseases, the rate of development of the parasite within the mosquito (the sporogonic cycle) and the mosquito biting/feeding rate (the gonotrophic cycle).

Thomson et al. [101] describe one of the first efforts aimed at seasonal forecasting of malaria in Africa based on ensemble predictions of rainfall and temperature from global coupled ocean-atmosphere climate models, and firmly established statistical climate-malaria links. The models achieved probabilistic predictions of anomalously high and low malaria incidence based on rainfall thresholds for Botswana up to four months in advance. Building on this work, Connor et al. [102] presented a framework for the integration of climate model based seasonal climate forecasts into early warning systems for climate sensitive diseases such as malaria and dengue. This work and that of Thomson [101] has been influential in guiding further endeavours related to the development of operational seasonal forecasts of malaria in southern Africa using outputs from numerical climate models run by forecasting centres such as the European Centre for Medium Range Weather Forecasting (ECMWF) [103,104]. For India, Lauderdale et al. [105] explored the feasibility of malaria forecasting by using a ECMWF seasonal forecast model to drive a numerical process-based dynamic malaria disease model. Using hindcasts from the ECMWF model, simulated forecasts of malaria were produced. These demonstrated probabilistic skill in predicting the spatial distribution of *Plasmodium falciparum* incidence particularly in regions where high seasonal and inter-annual variability of disease incidence is a characteristic. As well as showing some ability to predict the spatial distribution of malaria the seasonal forecast model was able to distinguish between years of "high", "above average" and "low" malaria incidence in the peak malaria transmission seasons with a three month lead time [105].

A number of statistical/empirical seasonal health forecasting models have been developed. For example, Lowe et al. [106] incorporated precipitation, minimum temperature, and Niño 3.4 index forecasts in a Bayesian hierarchical mixed model to make monthly predictions of dengue incidence in Ecuador for 2016. The ENSO element of this forecast system was in the form of Niño 3.4 SSTs derived from a structural time-series SST prediction model. It was found that the dengue forecast model was able to correctly predict an early peak in dengue incidence in March, 2016, with a 90% chance of exceeding the mean dengue incidence for the previous five years. Interestingly, when Lowe et al. [106] controlled for confounding due to chikungunya cases incorrectly recorded as dengue, this improved the prediction of the magnitude of dengue incidence. A similar approach was adopted by Lowe et al. [107] in the development of dengue forecasts for southeast Brazil. Poveda et al. [108] describe how satellite imagery of vegetation activity along with ENSO sensitive climate variables can be used as environmental indicators for malaria occurrence in Columbia. Armed with this knowledge, they demonstrate how statistical models and geographical information systems are applied by the Colombian health authorities to develop early warning systems for malaria. For the Solomon Islands in the western Pacific, where ENSO has clear impacts on rainfall as a disease sensitive climate variable, Smith et al. [109] applied stepwise regression to analyse climate variables and climate-associated malaria transmission at different lag intervals in order to identify rainfall thresholds associated with malaria categorised into three incidence categories. Study results not only revealed clear rainfall thresholds, but significant lag associations between rainfall and increases in malaria incidence such that drier October–December periods are followed by higher malaria transmission periods in January–June. Based on these statistical relationships an experimental early warning system has been proposed for the Guadalcanal region of the Solomon Islands [109]. Chuang et al. [110] used cross-wavelet coherence to evaluate the regional El Nino Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) effects on dengue incidence and local climate variables for Taiwan. Their work revealed the importance of non-linear and lag effects of minimum temperature and precipitation on dengue. These associations were applied in the successful prediction of dengue transmission between 2013 and 2015 [101].

While the potential for ENSO-based health forecasting is clear, despite improvements in observations and models, ENSO predictability and long-lead seasonal forecast skill, generally taken to mean the extent or lead-time for which boreal winter SST or any other ENSO index can be predicted with measurable skill, remains an issue [111]. A case in point is the 2014 ENSO forecast. Rather than a strong 2014 El Niño event occurring, as forecast, only weak warming in the key El Niño oceanic regions (Figure 5) was observed. This forecast "bust" [112,113] caused the ENSO prediction community to critically examine the efficacy of many of the significant ocean and atmosphere system components drawn on as a source of ENSO predictability [114–116]. A particular challenge for ENSO based health forecasting is the so-called "spring predictability barrier" [117–119]. This is basically a hiatus in ENSO forecasting accuracy for Northern Hemisphere spring when both dynamic and statistical ENSO prediction models display a sharp fall in their ability to predict sea surface temperature fields; following spring, the model ability to predict ENSO improves markedly. It is likely that the spring predictability problem exists because during this season El Niño/La Niña events are often in the stage of decay, following a winter peak, sliding into a neutral phase, which may persist or eventuate in a El Niño/La Niña later in the year. Consequently, the ENSO signal to noise ratio is low. Further during spring, the ocean does not exert a strong influence on the atmosphere because climatological (average) SST gradients in the tropical Pacific Ocean are much reduced and thus strong ocean-atmosphere coupling is compromised [117].

Given the expectations of the broad ENSO forecast user community related to ENSO forecasts as a panacea for climate risk management problems, much effort has been invested in improving predictability [120–122] with seasonal health forecasting scheme developers conscious that validation of predictions is a requisite part of the forecasting development process [104,123]. Further to the issues of predictability, other constraints related to seasonal health forecasting may well bear implications for the operationalisation of ENSO (climate)-sensitive disease early warning systems. Increasingly, seasonal health forecasts are couched in probabilistic terms that have been found to pose communication and uptake problems, making it imperative for forecast developers to think carefully how forecasts are provided to end users [124]. In the context of climate services based on climate forecasts, Ballester et al. [125] provide a sobering review of some of the challenges related to the construction of seasonal health forecasts. These include the capital and human resources and the associated governance arrangements required for development and implementation; the need for forecasting tools to master the complexity of the interactions between climate, disease transmission, socioeconomic disparities, and vulnerability; the imperative for integrated climate and health data sets; and acknowledgement that early warning systems and the climate forecasts on which they are based may only be effective when certain windows of opportunity present themselves, such as during ENSO events when there is a clear climate signature in a range of health responses.
