Potential of environmental models to predict meningitis epidemics in Africa

Authors


Corresponding Author Dr Madeleine C. Thomson, International Research Institute for Climate and Society (IRI), The Earth Institute of Columbia University, Lamont Campus, POB 1000, Palisades, New York, NY 10964, USA. Tel.: +1 845 680 4468; Fax: +1 845 680 4866; E-mail: mthomson@iri.columbia.edu

Summary

Objectives  Meningococcal meningitis is a major public health problem in Africa. This report explores the potential for climate/environmental models to predict the probability of occurrence of meningitis epidemics.

Methods  Time series of meningitis cases by month and district were obtained for Burkina Faso, Niger, Mali and Togo (536 district-years). Environmental information (1989–1999) for the continent [soil and land-cover type, aerosol index, vegetation greenness (NDVI), cold cloud duration (CCD) and rainfall] was used to develop models to predict the incidence of meningitis. Meningitis incidence, dust, rainfall, NDVI and CCD were analysed as anomalies (mean minus observed value). The models were developed using univariate and stepwise multi-variate linear regression.

Results  Anomalies in annual meningitis incidence at district level were related to monthly climate anomalies. Significant relationships were found for both estimates of rainfall and dust in the pre-, post- and epidemic season. While present in all land-cover classes these relationships were strongest in savannah areas.

Conclusions  Predicting epidemics of meningitis could be feasible. To fully develop this potential, we require (a) a better understanding of the epidemiological and environmental phenomena underpinning epidemics and how satellite derived climate proxies reflect conditions on the ground and (b) more extensive epidemiological and environmental datasets. Climate forecasting tools capable of predicting climate variables 3–6 months in advance of an epidemic would increase the lead-time available for control strategies. Our increased capacity for data processing; the recent improvements in meningitis surveillance in preparation for the distribution of the impending conjugate vaccines and the development of other early warning systems for epidemic diseases in Africa, favours the creation of these models.

Abstract

Objectifs  La méningite méningococcale est un problème majeur de santé publique en Afrique. Cet article explore l'usage potentiel des modèles basés sur le climat et l'environnement pour prédire la probabilité de l’émergence d’épidémies de méningite.

Méthodes Les séries de dates de cas de méningite enregistrés par mois et par district ont été obtenues pour le Burkina-Faso, le Niger, le Mali et le Togo (536 districts années). Les informations sur l'environnement du continent de 1989 à 1999 (sol et type de couverture du terrain, index d'aérosol, verdure de la végétation, durée des nuages froids et pluviosité) ont été utilisées pour développer des modèles prédictifs de l'incidence de la méningite. L'incidence de la méningite, la poussière, la pluviosité, la verdure de la végétation et la durée des nuages froids ont été analysées comme anomalies (moyenne moins la valeur observée). Ces modèles ont été développés en utilisant les régressions linéaires univariée et multivariée par étape.

Résultats  Les variations dans l'incidence annuelle de la méningite au niveau du district étaient liées aux variations du climat mensuel. Une forte association a été trouvée entre la saison de l’épidémie et les estimations de pluviosité ou de la poussière avant et après la saison de l’épidémie. Ces associations étaient observées dans tout type de revêtement du sol mais elle étaient surtout plus fortes dans les régions de savane.

Conclusions  La prédiction des épidémies de méningite peut être possible. Le développement complet de ce potentiel requiert: 1) une meilleure compréhension des phénomènes épidémiologiques et environnementaux qui déterminent les épidémies et comment les données climatiques par satellites reflètent les conditions sur le sol et 2) des données épidémiologiques et environnementales plus étendues. Des outils de prévision météorologique capables de prédire des variations de climat 3 à 6 mois avant l’épidémie pourraient augmenter les délais disponibles pour les stratégies de contrôle. Notre capacité renforcée de l'analyse des données, la récente amélioration de la surveillance de la méningite dans la préparation pour la distribution de vaccins conjugues imminents et le développement d'autres systèmes de divulgation très tôt des cas d’épidémies de maladies en Afrique, favorisent la créations de ces modèles.

Abstract

Objetivos  La meningitis meningocócica es un grave problema de salud pública en África. Este reporte explora el potencial de modelos climáticos / ambientales para predecir la probabilidad de ocurrencia de epidemias de meningitis.

Métodos  Se obtuvieron casos de meningitis seriadas en el tiempo, por mes y distrito, para Burkina Faso, Níger, Mali y Togo (536 distrito-año). Se utilizó la información ambiental (1989–99) para el continente (suelo y tipo de cobertura vegetal, índice de aerosol, diferencia normalizada de los índices de vegetación (NDVI), el cold cloud duration (CCD) y la precipitación) para desarrollar modelos que predijesen la incidencia de meningitis. La incidencia de meningitis, polvo, precipitación, NDVI y CCD fueron analizadas como anomalías (media menos valor observado). Lo modelos se desarrollaron utilizando un regresión linear univariada y multivariada escalonada.

Resultados  Las anomalías en la incidencia anual de meningitis a nivel de distrito estaban relacionadas a las anomalías mensuales en el clima. Se encontraron relaciones significativas tanto para la precipitación como el polvo en las épocas pre, post y epidémicas. Aunque se encontraban en todas las clases de cobertura vegetal, estas relaciones eran especialmente fuertes en las áreas de sabana.

Conclusiones  Es posible predecir epidemias de meningitis. Para desarrollar totalmente este potencial se requiere: a) un mayor entendimiento de los fenómenos epidemiológicos y ambientales que hay detrás de las epidemias y como los proxies (indicadores indirectos) climáticos, provenientes de imágenes satelitales, reflejan las condiciones en la tierra y b) bases de datos epidemiológicas y ambientales más extensivas. Las herramientas para las predicciones climatológicas, capaces de predecir variables climáticas con 3–6 meses de antelación a una epidemia, aumentarían el tiempo disponible para establecer estrategias de control. Nuestra cada vez mayor capacidad de procesar datos; las mejoras recientes en la vigilancia de la meningitis como resultado de la preparación para la distribución inminente de vacunas conjugadas y el desarrollo de otros sistemas de alarma temprana en África, favorecen la creación de estos modelos.

Introduction

Epidemics of meningococcal meningitis present a major health problem worldwide, but nowhere more so than in the ‘Meningitis Belt’ of sub-Saharan Africa where the majority of epidemics are located and which also suffers the greatest burden of endemic disease (Molesworth et al. 2002).

Epidemic occurrence may be associated with the environment. Epidemics start during the dry season and usually subside at the onset of the rains and their location has a distinct ecological pattern, indicating that certain environmental factors, such as low absolute humidity, land-cover types and dusty atmospheric conditions, may play an important role (Lapeyssonnie 1963; Cheesbrough et al. 1995; Greenwood 1999; Molesworth et al. 2003). Work by Sultan et al. (2005) indicates that the onset of the seasonal rise in national meningitis cases in Mali corresponds with large scale atmospheric phenomena associated with the Sahelian dry season although the index they used was unable to distinguish between high and low incidence years. In addition to determining epidemic location and seasonality, the environment may also play a role in determining their temporal occurrence. While there is strong evidence that the spatial and seasonal pattern of meningococcal meningitis in Africa is determined, at least in part, by climatic and environmental factors, it remains to be demonstrated that inter-annual variation in incidence is associated with anomalous climate-related environmental conditions. This paper explores this possibility and considers the potential implication of our findings to the future development of environmental models to predict meningitis epidemics in Africa.

Materials and methods

Time series of meningitis cases aggregated by month of diagnosis and district of report were obtained for four regions in West Africa comprising Burkina Faso (January 1997–December 2001, 50 districts), Niger (January 1993–December 2001, 38 districts); and parts of Mali (July 1989–July 1998, seven districts in Segou province) and Togo (August 1990–July 1997, four districts in Savannes province), through the respective national Ministries of Health. The numbers of cases were aggregated to annual totals from September to August (to reflect the meningitis transmission year) and incidence rates were calculated based on USGS total population estimates (http://grid2.cr.usgs.gov/globalpop/africa/). We used selected environmental information, available as digital grid-based surfaces for the African continent between 1989 and 1999, to develop models to predict the incidence of meningitis (Table 1). For each district these included the predominant soil and land-cover type, and for each month a spatial average of the mean daily aerosol index (a measure of dust in the upper atmosphere, missing from May 1993 to June 1996), mean dekadal estimate of vegetation greenness (NDVI), mean dekadal cold cloud duration (CCD) and total rainfall. The distribution of the districts studied is presented in Figure 1. Environmental data explored in model development is described in Table 1.

Table 1.   Environmental data used in the model
VariableTimeSpatial resolution (km)
  1. NDVI, normalised difference vegetation index; CCD, cold cloud duration.

  2. * Missing May 1993–June 1996.

  3. Derived from data provided by:

  4. † US National Aeronautics and Space Administration – Goddard Space Flight Centre, USA.

  5. ‡ The Africa Data Dissemination Service (http://edcintl.cr.usgs.gov/adds/adds.html).

  6. § The Food and Agriculture Organisation, Rome.

  7. ¶ The Climate Research Unit, University of East Anglia, UK; The US Geological Survey (http://grid2.cr.usgs.gov).

  8. †† US Geological Survey (http://grid2.cr.usgs.gov).

Satellite data
 †Average daily aerosol index (dust)*January 1980ca.100
 ‡Average dekadal NDVIJuly 1981ca.8
 §Average dekadal CCDJanuary 1989ca.8
Total rainfall
 ¶Interpolated meteorological data (monthly)January 1951 –December 1995ca.5
 ‡Satellite derived rainfall estimates (dekadal sum)January 1996ca.8
Digital maps
 §Soil type1977ca.17
 ††Land-cover type1992/1993ca.1
 ††Population density1980, 1990ca.4
Figure 1.

 Districts in Burkina Faso (vertical stripes), Mali (horizontal stripes), Togo (dots) and Niger (circles) where weekly surveillance data were obtained for analysis.

The annual meningitis incidence, and monthly dust, rainfall, NDVI and CCD were analysed as anomalies calculated as the mean value for the biological year (incidence) or month (environmental data) over which the data were collected minus the observed value. The association between the loge incidence anomaly and environmental predictands was explored at the district level using univariate correlation statistics and stepwise multi-variate linear regression, based on a total of 536 district-years of observation for which data were available. Factors that consistently appeared in the first five predictands for a range of exploratory models were selected for the final model. To verify the findings the same variables were then used to create a second model using a random selection of 65% of the data (development data set) and this second model was then used to predict results for the remaining 35% (validation data set). Separate models where created using the same predictands for each land-cover class.

Ethical approval was obtained from the Research Ethics committee of the Liverpool School of Tropical Medicine.

Results

The seasonality of dust, NDVI, CCD and rainfall were explored with respect to national boundaries and land-cover characteristics. All variables were highly seasonal and varied between years, national boundaries and land-cover type.

The most consistent predictands of anomalies in meningitis incidence were anomalies in rainfall (August and January) and anomalies in dust (October and April). For any individual year, for instance 1990 (which would include incidence data for the corresponding biological year – in this case from September 1989 to August 1990) the best predictands prior to the meningitis season where October dust and January rainfall; during the peak season; April dust and post the peak season August rainfall.

The seasonality and intensity of rainfall, aerosol index and meningitis incidence, by land-cover type are presented in Figure 2. Compared with savanna and grassland areas, barren regions had much lower incidence of meningitis and a shorter overall season of transmission. The highest dust levels occurred during the rainy season, which lasted from April to October. The rainy season started earlier and was more intense in savanna compared with grassland or barren areas. Barren areas were dustier than grassland or savannah areas.

Figure 2.

 Rainfall (a), aerosol index (b) and meningitis incidence (c) by month (September–August, Sep = 1) and land-cover type.

Univariate regression for April and October dust; August rainfall and ln meningitis incidence anomalies for savannah, grassland and barren areas are presented in Figure 3a–c. Figure 3d represents the relationship of ln meningitis incidence anomalies with rainfall estimate anomalies in January.

Figure 3.

 Meningitis incidence anomaly by dust [April (a) and October (b)] and rainfall [August (c) and January (d)] anomalies.

A greater incidence of meningitis was experienced in years with a negative compared with positive January rainfall anomaly treated as a bivariate variable (0, 1). Using the four variables weighted by district population produced a model, which described about 38% of variance in incidence between years (P < 0.001, stepwise linear regression) as shown in Table 2 and Figure 4a. This was confirmed in the validation process (adjusted R2 0.40 and 0.34 for development and verification datasets, respectively). The model appeared to work better in savanna compared with grassland areas (adjusted R2 0.42 and 0.32 respectively) and least well in barren areas R2 = 1.87) Table 2, Figure 4b–d.

Table 2.   Model results
Predictands*Standardised coefficients
AllSavannahGrasslandBarren
  1. * All variables were significant P < 0.001.

August rainfall anomaly−0.246−0.241−0.3090.438
January rainfall anomaly 0 or 1−0.258−0.143−0.4250.349
April aerosol index anomaly−0.339−0.383−0.260−0.435
October aerosol index anomaly0.2610.2980.187−0.069
Adjusted R20.3820.4330.3720.187
Figure 4.

 Observed and predicted* ln meningitis incidence anomaly (a) all land-cover types, (b) savannah, (c) grassland and (d) barren (*All cases weighted by population density).

Discussion

Our results are the first indication that climate variability may indeed play a role in the changes in incidence between years and may therefore be of value in the development of an early warning system.

Given the limitations of the data, this report can only be considered as an exploratory attempt to establish the feasibility of more comprehensive forecasting models. The model was based only on environmental variables, and other factors, such as vaccination coverage and recent epidemic history as well as changes in the characteristics of the bacteria such as the introduction of new meningococcal clones in an immunologically naïve population such as W135 (Decosas & Koama 2002), will also be necessary for the development of meningitis epidemics. In addition, the creation of temporal models is entirely dependant on long-time series and there are major limitations with the quality of reporting across the continent. Although, this situation has improved in recent years with the development of a meningitis surveillance system across Africa, and better quality data may become available in the future, no account was made for delayed or under-reporting of cases, of their misclassification, which may undermine comparisons between regions and between months. Moreover, although, the health community has become more adept at using environmental information created for other purposes (Thomson et al. 2000), these data sets often have poor spatial and temporal resolution, poor calibration with environmental proxies of direct interest to the health analysis and inadequate representation of the uncertainty associated with the data sets. For example, we used the absorbing aerosol index obtained from the Total Ozone Mapping Spectrometer (TOMS) satellite sensors as an estimate of dust in the atmosphere. This data has been used by the climate community to map the global distribution of atmospheric dust sources (Herman et al. 1997). However, the data relates to aerosols in the upper atmosphere and it is not clear to what extent this corresponds with dust near the ground. A careful analysis of the TOMS data in relation to meteorological station visibility data (a ground based measure of dustiness) and human dust exposure is required if we wish to develop models that use data with a wide area coverage, are reliable over longer periods of time and are less susceptible to social disruption.

The results, nonetheless, are compatible with other early warning signs from meningitis epidemics. For example, the early presentation of cases in the season is often considered a warning sign of an impending epidemic (Anonymous 2000). Our results suggest that environmental conditions prior to the peak meningitis season (excess dust in October and a rainfall deficit in January) may be important in increasing its incidence and that key environmental factors could have predictive value for epidemic forecasting. Delayed onset of rainfall at the end of the meningitis season would in theory have good potential but, in the absence of skilful seasonal climate forecasts many months in advance, this information would not be available in time to decision makers for incorporation in predictive models. A deficit of rainfall in August corresponded with higher than average incidence anomalies; this may have been by chance or may be because this type of anomaly corresponds to a poor onset of the monsoon or some other climatic characteristic advantageous to the spread of meningitis. Other environmental information during the epidemic period may also be important. For example, unseasonable periods of relatively high cloud cover/humidity and or light rainfall occurring in January–February in the West African dry season could be a likely candidate for future investigations and the cessation of the epidemics is clearly related to the onset of the rains and the high humidity belt that precedes them. Medium range weather forecasting (10–20 days ahead) may provide epidemic control managers, in the midst of controlling an epidemic, with information on whether or not they are likely to get relief from the epidemic conditions by the arrival of the first rains or perhaps more importantly the humidity changes that proceed it. In our analysis, negative dust anomalies in April were strongly indicative of positive meningitis incidence anomalies – the reason for which is unclear.

Forecasting tools require the creation of improved epidemiological and environmental data sets, but equally crucial, is the development of skills for environmental prediction. Key variables for meningitis should occur before or early in the dry season. The vast majority of climate analysis and seasonal climate predictions for Africa however has focussed on the rainy season and now it is time to encourage the climate community to improve their understanding of the African dry season in terms of its inter-annual variability and predictability. Seasonal climate forecasting (up to 6 months ahead) has developed rapidly in recent years and their relevance to health decision makers is currently being explored (Thomson et al. 2006). If climatic/environmental variables were shown to be associated with meningitis epidemics, then it would be valuable if they could be predicted 3–6 months ahead as this would extend the lead-time of a warning system.

The basic model presented here and our review of the work required for their development suggests that meningitis forecasting tools, which predict a change in epidemic risk, could be feasible in the future. The creation of these models requires a better understanding of the basic epidemiological and environmental phenomena underpinning these events. Our increased capacity to process large datasets (as environmental dataset often are), the improved surveillance systems for meningitis being developed in preparation for the deployment of new meningitis conjugate vaccines and the development of early warning systems for Africa (e.g. famine and malaria), favours the creation of these models. Consideration should be given to the development of these forecasting models, which could result in improved decision-making tools fit for the 21st century.

Acknowledgements

We are grateful to the Meningitis Research Foundation who supported this work through a project grant and the Ministries of Health of Niger, Togo and Burkina Faso support for the collection of meningitis data.

Ancillary