^{1}

^{2}

^{3}

^{4}

^{5}

^{6}

^{7}

^{8}

^{1}

^{*}

The authors have declared that no competing interests exist.

Conceived and designed the experiments: MT BC SV. Analyzed the data: MT RH BC RD CB SV. Wrote the paper: MT RH BC RD CB SV. Provided the data: RH. Designed the toolbox to perform wavelet analysis: BC.

Determining the factors underlying the long-range spatial spread of infectious diseases is a key issue regarding their control. Dengue is the most important arboviral disease worldwide and a major public health problem in tropical areas. However the determinants shaping its dynamics at a national scale remain poorly understood. Here we describe the spatial-temporal pattern of propagation of annual epidemics in Cambodia and discuss the role that human movements play in the observed pattern.

We used wavelet phase analysis to analyse time-series data of 105,598 hospitalized cases reported between 2002 and 2008 in the 135 (/180) most populous districts in Cambodia. We reveal spatial heterogeneity in the propagation of the annual epidemic. Each year, epidemics are highly synchronous over a large geographic area along the busiest national road of the country whereas travelling waves emanate from a few rural areas and move slowly along the Mekong River at a speed of ∼11 km per week (95% confidence interval 3–18 km per week) towards the capital, Phnom Penh.

We suggest human movements – using roads as a surrogate – play a major role in the spread of dengue fever at a national scale. These findings constitute a new starting point in the understanding of the processes driving dengue spread.

Dengue fever is a mosquito borne viral infection. It has become a major public health problem during the past decades: only 9 countries were affected in the 1970s; dengue is now endemic in more than 100 countries. In the absence of any vaccine or specific treatment, control of dengue fever is currently limited to vector control measures, which are difficult to implement and hardly sustainable, especially in low income countries. To implement efficient control measures, it is crucial to understand the dynamics of propagation of the disease and the key factors underlying these dynamics. In this study, data from 8-year national surveillance in Cambodia were analysed. Dengue fever follows a recurrent pattern of propagation at the national scale. The annual epidemics originate from a few rural areas identified in this work. This study also suggests additional evidence for the role of human movement in the spatial dynamics of the disease, which should be accounted for in control measures. These results differ from the current knowledge about dengue dynamics and are therefore of interest for future research.

Cambodia is a low-income tropical country, hyper-endemic for all four serotypes of dengue infection. As such, dengue epidemics occur every year during the rainy season and result in considerable morbidity and economic burden. The basis for these recurrent epidemics is an increased vector activity during the rainy season and complex interactions between hosts and viruses with short lived cross-protective immunity

The objective of this study was to determine whether there was a spatial-temporal pattern of dengue propagation in Cambodia repeating year after year. The characterisation of such patterns is important to understand the forces driving dengue spatial spread and aid better control and logical allocation of public health resources.

Cambodia is a low-income country located in South-East Asia, divided into 24 provinces and 180 districts covering 181,035 km^{2}. Out of the 13.4 million people in 2008, more than 80% live in rural areas; 1.3 million people (9.9%) live in Phnom Penh, the capital city. Cambodia is lagging behind other countries in South-East Asia in terms of economic or demographic development. For example, unlike Thailand, the demographic transition has not occurred yet. The South-West and the North of the country (see grey districts in ^{2}). The rest of the country is composed of flat plains with few cities (mainly Phnom Penh, Kampong Cham, Siem Reap and Battambang, see

Mean annual incidence rates (in number of cases declared per 100,000 people per year) are calculated over 2002 to 2008, for districts with more than 20 people per km^{2}. Cambodia is surrounded by the Indian Ocean (bottom left), Thailand (West), Lao (North) and Vietnam (East and South-East). Phnom Penh, the capital, is represented by a circle, Siem Reap by a triangle, Kampong Cham by a square and Battambang by a lozenge. Blue lines represent the Mekong River, going north to south, and the Tonle Sap River linking the Tonle Sap central Lake to the Mekong River. Green lines represent national roads. Grey districts have less than 20 people per km^{2}.

Cambodian National surveillance recorded 109,332 dengue cases during 2002–2008, a period over which the reporting process was stable. Cases were reported passively from public hospitals and actively from 5 major sentinel hospitals located in the cities of Siem Reap (1 hospital), Kampong Cham (1 hospital) and Phnom Penh (3 hospitals). Cases were clinically diagnosed using the 1997 World Health Organization (WHO) case definitions, allowing clinical and paraclinical (haematocrit and platelets count) distinction between classic dengue fever, dengue haemorrhagic fever and dengue shock syndrome ^{2} from the analysis, dismissing 3298 declared cases (3.03% of the total declared cases). Since patients' districts of residence were recorded, we calculated weekly incidence rates for each of the 135 remaining districts. Denominators for incidence rates were interpolated linearly using the 1998 and 2008 national censuses. Based on age distribution similarities between provinces, and on similarities between 1998 and 2008 age structures, we assumed that age structure was homogeneous over the country and have not standardised incidence rates according to age. As national surveillance data were made available to be utilised for such temporal and spatial analyses and have been routinely published nationally and once internationally

Time series of dengue incidence were square-root transformed to stabilise the variance, subsequently centred to zero-mean and normalised to unit variance. We then performed a wavelet transform and used wavelet phase analysis to describe dynamic patterns of dengue. This spectral method is well-suited for the analysis of non-stationary time-series such as epidemic curves. It is particularly useful to filter the data in any given frequency band and extract the phase of any given periodic component

The four districts are: district #306 (black), a rural district located around Kampong Cham; Phnom Penh (red); District #307 (green), a rural district located mid-way between #306 and Phnom Penh; District #104 (blue). (A) Weekly raw incidence rates (in number of cases declared per 100,000 people per week). (B) Annual component of incidence, obtained by filtering raw weekly incidence in the 0.8–1.2 year periodic band using wavelet analysis. (C) Phase of the annual component of incidence, computed in the 0.8–1.2 periodic band using wavelet analysis (see

The wavelet transform was done using R software

The wavelet coefficients corresponding to a period ranging from 0.8 to 1.2 year were used to reconstruct filtered time series corresponding to the annual epidemic in each district. These filtered time series, called “annual component of incidence”, are illustrated in

The phases of the epidemic have been computed in the periodic band 0.8–1.2 year (see equation (5) in

By calculating phase differences, one can then determine in which order districts are affected by the annual epidemic. This ranking allowed us to identify districts hit early by the seasonal epidemic. Time series of the temporal lag between seasonal epidemics at different locations were thus estimated from phase differences, according to equations (6)–(8) in

In some districts identified as early districts, there were not enough cases reported to be confident that there was significant dengue transmission ongoing, given that cases are declared on a clinical basis. Therefore, for each district, we identified epidemic years using the national epidemic threshold

For a given year, to evaluate the speed at which dengue epidemic propagates along a given geographical axis comprised of I districts, we performed a regression analysis: Y_{i} = β0+β1 X1_{i}+ε_{i} , with i in [1, I], Y_{i} the annual mean of the temporal lag between district i and a district located along the axis and selected as a time reference to compute temporal lags, and X1_{i} the corresponding distance as the crow flies, in km, between the centres of district i and the reference district. The speed of dengue propagation along the axis was estimated as the inverse of the regression slope β1.

To compare the dynamic pattern along J different axes, we performed, each year, an analysis of covariance _{ij} = β0_{j}+β1_{j} X1_{ij}+ε_{ij}, with i in [1, I_{j}], j in [1, J], Y_{ij} the annual mean of the temporal lag between district i of axis j and a district located along the axis j and chosen as a time reference, X1_{ij} the corresponding distance separating the centres of districts i and the reference district on axis j, and β0_{j} and β1_{j} the intercept and slope of the regression line of axis j. We will call “p-value of interaction” the p-value of the hypothesis that the model slopes β1_{j} are all equal, or in other words, that the speed of propagation is the same along different axes. If the model gave a p-value of interaction below 0.05, the propagation was considered heterogeneous along the different axes, and separate regressions were performed for each axis.

As we wanted to know whether the results were consistent from year to year, we performed this analysis each year.

All analyses and figures were performed using R software

All confidence intervals (C.I.) provided were calculated using classical methods for calculating a confidence interval around a mean, unless stated otherwise.

Weekly incidence rates (cases per 100,000 people per week) were computed in each of the 135 districts where population density is higher than 20 people per km^{2} in Cambodia. Districts are ranked by increasing distance to Phnom Penh from bottom to top.

To characterise and compare the spatio-temporal pattern of dengue incidence along those two axes and determine the focal starting areas of the annual epidemic, we performed a phase analysis using a wavelet approach (see

The examination of wavelet power spectra (

Phases are computed in the 0.8–1.2 year periodic band. (A) Map of the two geographic areas chosen: the national road in blue, and the Mekong River in orange. (B) Phase of districts along the Mekong River (orange in

First, the pattern of spatial synchronicity observed is consistent from year to year. To test this result, each year, we ranked the 135 districts according to their phase during the epidemic period (weeks 13 to 39), from the district where the epidemic has the most advanced phase to the one with the highest phase delay. We then performed a Spearman correlation test on these ranks, comparing ranks from year n with ranks of years n+1. Except in 2007, correlation coefficients were all higher than 0.38 and significant (all p-values<0.03), inferring that districts are affected in a similar chronological order year after year. In 2007, the order in which districts were affected by the epidemic was not significantly correlated to the order of the previous year.

This ranking also allowed us, each year, to identify districts where the annual epidemic appears early. We have identified three starting points, all located in rural areas: district #306 and a few rural districts around (district #306 being the most early of them), district #104 and 2 other districts around, and, some years, district #805, located along the Mekong River, South to Phnom Penh, along the Vietnamese border (see ^{2}. They are all located away from the three urban centres where sentinel hospitals involved in active surveillance are. They are consistently the same, year after year.

Secondly,

To evaluate the speed of propagation along each axis, we calculated the temporal lag of the seasonal pattern in each district relative to the district #306

Temporal lags between epidemics and distances are computed relative to district #306. The lines show the linear regressions between the mean annual temporal lag of the annual epidemic in each district and the distance for 2002 (A), 2003 (B), 2004 (C), 2005 (D), 2006 (E) and 2007 (F). Colours represent the geographic localisation of each district, according to

Covariance analysis | Regression (“Mekong River”) | Regression (“national road”) | |||||

year | n |
p-value of interaction | n |
p-value |
1/slope estimate |
n |
p-value |

2002 | 40 | 0.005 | 22 | 0.02 | 11 | 18 | 0.70 |

2003 | 44 | <0.001 | 22 | 0.001 | 9 | 22 | 0.22 |

2004 | 31 | <0.001 | 22 | <0.001 | 8 | 9 | 0.07 |

2005 | 38 | <0.001 | 19 | <0.001 | 8 | 19 | 0.09 |

2006 | 40 | <0.001 | 22 | 0.004 | 13 | 18 | 0.22 |

2007 | 52 | 0.001 | 26 | 0.006 | 17 | 26 | 0.378 |

Number of districts where an annual epidemic occurred according to the national threshold, along the Mekong River, the national road (

P-value of the regression slope estimate.

Inverse of the estimate of the regression slope β1, in km per week (see

To test the role Phnom Penh plays in this heterogeneity, we ran the same analysis, but excluding Phnom Penh districts. The propagation remained significantly heterogeneous at an alpha-level of 5%, except in 2002 and 2007, and remained significantly heterogeneous at an alpha-level of 10% in 2002 and 2007 (results not shown).

The removal of districts with low population density has no effect at all on results shown. The exclusion of low incidence years did not modify the two conclusions arrived at in the paper (onset of the seasonal epidemic in highly localised rural areas, and heterogeneous propagation in the country), but removed the potential bias linked with the very low number of cases during the non epidemic years (see

To sum up, our results show that the seasonal epidemic consistently starts in the same three rural districts in Cambodia. Then the propagation is not homogeneous in the country. In districts located along the busiest road, dengue epidemics appear simultaneously and early (with all districts being hit in less than a month), whereas districts located along the Mekong River get hit by the seasonal epidemic later, with the delay being proportional to the distance to district #306.

In 2007 an exceptional epidemic occurred in Cambodia with DENV-3 as the dominant etiological agent, and a four fold increase in weekly incidence rate compared to the four previous years (

Regarding the onset of the national epidemic, there is a tendency for the dengue season to start in few rural districts more often than in any other district moving from rural areas towards urban centres, with, for example, annual epidemics in districts #306 and #104 (

To our knowledge, it is the first time that a recurrent heterogeneous pattern of propagation of dengue is revealed in surveillance data at a national scale. This pattern of propagation could come from actual transmission of dengue viruses between districts, via infected mosquitoes or humans, or correspond merely to spatial differences in the emergence of the epidemic, differences due either to spatial differences in local (within-district) dynamics, or to forcing by extrinsic factors such as climate (Moran effect). Given the limited dispersal range of

Climate is quite homogeneous in Cambodia, with a narrow temperature range across the country, and the monsoon striking the country from South to North within a month only, around April-May. Climate, by influencing the vector's life cycle, is clearly driving the dengue seasonality observed in Cambodia. It could easily explain the high synchronicity of epidemics observed along the national road if this synchronisation was observed country-wide. However, two of our observations preclude climatic forcing to be the underlying factor for the spatial-temporal pattern observed. First, climate is more homogeneous in the country compared to the observed spatial-temporal pattern of dengue fever. The existence of a very slow oriented dengue propagation along the Mekong River, going North to South, with the epidemic in some districts lagging more than 3 months behind the one in districts that are only less than 200 kilometres in distance cannot be accounted for by climatic differences between districts. Secondly, the epidemic starts as early as in March in the three districts identified as starting points, whereas the monsoon only begins end of April/beginning of May.

We believe the heterogeneous propagation observed is related to the heterogeneity of traffic on the roads of the country, where traffic allows the movement of human carriers or transported mosquitoes infected with dengue. The national road –the busiest country road connecting the two main economic cities within 4–5 hours – probably explains synchronicity. By contrast, the existence of dirt roads along the entire Mekong River, on which traffic and population movement between smaller villages are slower and more difficult than along the national road supports our hypothesis. Despite its epidemiological relevance, our understanding of the relationship between human movement and pathogen transmission remains limited

The higher synchronisation and higher incidence levels across the country in 2007 argues for an association with higher net reproduction ratios of infection due to lower herd immunity when a new serotype is introduced

Major limitations of our analysis include underreporting inherent to passive surveillance data and potential selection bias leaning towards underreporting from urban centres compared with rural areas. However, underreporting would only affect the amplitude of the epidemics in each district and therefore have little effects on the study of synchronism when using the wavelet phase analysis approach

One could think that not standardising incidence according to age could impair the results obtained. But age standardisation only affects the results by modifying incidence levels, and as our results do not rely on differences in incidence levels, this is not a real concern.

Our findings and speculations require further research for additional evidence. Firstly, reasons as to why the three areas identified as being hit early by the epidemic repeatedly year after year (districts #306, #104 and #805) are unclear. This warrants further filed investigations to identify specific factors that trigger the epidemic in these settings.

Secondly, the data we analysed here can only reveal the spatio-temporal dynamics and help make hypotheses on underlying factors. Given the results of this study, and particularly the heterogeneity of dengue spatial dynamics, surveillance data on the spatial distribution of the serotypes (and genotypes if possible) of the co-circulating dengue viruses would help validate (or not) our hypotheses on dengue dynamics in Cambodia, using independent data.

Lastly, given the potential benefit in term of disease control from demonstrating the efficient role of humans' movement in dengue spatial transmission, one might consider further investigations in this direction, collecting data and using dynamic models for instance.

(PDF)

^{2}.

(PDF)

(TIF)

(TIF)

^{2}.

(TIF)

(TIF)

(TIF)

(MOV)

(MOV)

We thank Tristan Rouyer for providing an R toolbox (translated from Cazelles' Matlab toolbox) for performing wavelet analysis, Kévin Cazelles, Christophe Menkès and Morgan Mangeas for their help with parts of the analysis, and Yann Quéau for useful comments.