Introduction Many infectious disease show marked heterogeneity of transmission [1]. Mathematical models predict that this heterogeneity reduces the efficacy of disease control strategies [2], and intensifying control measures on foci of high transmission is predicted to be very effective in reducing overall transmission [1]. Marked spatial heterogeneity of malaria transmission on the household level is consistently detected when analysed [3]–[9], and results from both genetic and environmental factors [10],[11]. It is unclear how stable hotspots are in longitudinal data. Malaria risk is related to environmental factors [12] including altitude [13], cultivation practices [14], urbanization [15], and distance from bodies of water [16]. However, ecological analyses to guide malaria control have been limited by the following factors: the overall models are complex [17]–[19], the same ecological feature may not have a consistent effect in different settings [20],[21], and there is marked residual variation in malaria risk despite models with detailed ecological data [22]. Furthermore, the resolution of multitemporal remote sensing satellite data (i.e. data based on more than a single snap-shot) for environmental monitoring is rarely finer than 0.5–1 km [23]. Since vector dispersion occurs on average over 0.5–1 km distances [24]–[29], this is the scale at which “hotspots” of transmission need to be identified in order to impact overall transmission. Malaria episodes have been found to cluster at this scale to form hotspots in high resolution geo-spatial analyses in Mali [6], Uganda [7], Ethiopia [30], and the highlands of Kenya [31],[32]. Here, we have conducted an analysis of malaria episodes and parasitaemia over 12 y, a substantially longer time period than has been reported previously, across three different cohorts without conspicuous variations in geography, such as nearby water bodies. We examine febrile episodes, asymptomatic parasitaemia, serological markers of exposure, environmental remote sensing data, and molecular studies of parasite clones to describe the spatial and temporal limits of hotspots, and to examine whether heterogeneity can be predicted. Methods The approval for human participation in these cohorts was given by Kenya Medical Research Institute Scientific committee and National Review and Ethical Committee of Kenya Medical Research Institute, and was conducted according to the principles of the declaration of Helsinki. Surveillance for Malaria The cohorts under surveillance for malaria episodes were located in Chonyi, Ngerenya, and Junju sublocations of Kilifi District, on the coast of Kenya between January 1998 and June 2009 (Figure 1). Concurrent entomological studies and parasite prevalence rates suggest that the transmission intensity is higher in Junju and Chonyi than in Ngerenya [33],[34], but transmission has been falling throughout the period of study [35]. 10.1371/journal.pmed.1000304.g001 Figure 1 The distribution of homesteads monitored in the three cohorts is shown within Kilifi District. The field methods used to identify episodes of febrile malaria and asymptomatic parasitaemia have previously been described [36],[37]. Weekly active surveillance was undertaken, and children with fever had blood slides for malaria parasites. In Chonyi and Ngerenya, children with either subjective (i.e., reported) or objective fever (temperature ≥37.5°C) had blood smears performed for estimating the parasite density. In Junju blood smears were done only on children with an objective fever, but children with subjective fever were seen again 6–12 h later, and the temperature measurement repeated. Blood smears were made if objective fever was confirmed at this measurement. The parents of the children in Chonyi and Ngerenya were instructed to report to Kilifi District Hospital 20 km away if the child had any symptoms of disease at any time (and reimbursed for travel expenses), and in Junju trained field workers were available at all times in the villages for passive surveillance. Antimalarials were supplied for proven episodes of malaria by the study team in accordance with government of Kenya guidelines; this was sulfadoxine-pyrimethamine until 2004, and co-artemether thereafter. Study participants may have used private clinics or bought antimalarials without the study team's knowledge, but given the availability of free treatment this was probably infrequent. Surveys for asymptomatic parasitaemia were undertaken once yearly, immediately before the rainy season. All individuals recruited to the study were asked to attend for blood sampling, and microscopy results were available for 70%–88% of participants for each survey. The Geographic Positioning System coordinates from the Kilifi Demographic Surveillance Survey were linked to each homestead in the study. In Ngerenya and Chonyi, all the residents at individual homesteads were recruited, but in Junju only children under 8 y of age were recruited. The homesteads in Ngerenya and Junju were evenly spaced throughout the study location, but in Chonyi the homesteads were distributed along a central road through the study area. Children born in the study homesteads during the period of monitoring were recruited, and so the average age of the cohort did not increase over time (Table 1). 10.1371/journal.pmed.1000304.t001 Table 1 Cohort summary characteristics. Characteristics Chonyi Junju Ngerenya Asymptomatic parasitaemia prevalence rate 35% 32% 14% Incidence of febrile malaria (episodes per child year) 0.82 0.55 0.49 Average population 818 462 428 Median length of follow-up per child (y) 2.8 3.5 5.0 Median age of child (y) 5.5 4.2 5.7 Years of longitudinal monitoring in cohort 3 5 11 Homesteads 59 149 48 Participants per homestead 13.9 3.1 8.7 Average distance between adjacent sampled homesteads (km) 0.08 0.08 0.35 Age range (y) 0–80 0–12 0–90 Area including cohort dimensions N to S (km) 7 7.6 6 Area including cohort dimensions E to W (km) 9.4 7.6 8.4 The asymptomatic parasitaemia prevalence includes both adults and children. The incidence of febrile malaria is given for participants 2,500/µl [36]. Episodes of febrile malaria were censored for 21 d after the last episode. The incidence per homestead per year was calculated by episodes of febrile malaria per homestead divided by the number of children who were 0.05). The most significant parameters from each channel were then combined in a final multivariable logistic regression model, with further exclusion of nonsignificant parameters (p>0.05). There are approximately 120 MODIS picture elements (pixels) widths/heights to each degree of latitude/longitude at the equator (i.e., approximately 1 km squared per pixel), so several homesteads are included in each pixel. The logistic regression models were therefore adjusted to take account of the nonindependence of observations by using the sandwich estimator to group observations by pixel [50]. The coefficients and constant from the final model (i.e., with p 1.0 OD, but only 37% of the whole cohort's malaria episodes would occur in these homesteads. In order to include 50% of the malaria episodes, a cut-off of 0.85 would be required, and this would require 32% of the homesteads in the study area to be treated. Table 8 shows the results of monitoring episodes of febrile malaria during 1 or 2 mo of surveillance to predict febrile malaria episodes during the following year, on the basis of the assumption that any homestead with at least one episode of febrile malaria during the monitoring period would be targeted for treatment. 10.1371/journal.pmed.1000304.t008 Table 8 Predicted performance if all homesteads with one or more episode of febrile malaria during 1-mo monitoring are targeted for measures to interrupt transmission over the following year. Monitoring Percent Homesteads Targeted Percent Malaria Covered January (short rains) 28 73 March (dry) 17 48 May (start long rains) 26 68 September (dry) 20 65 November (dry) 17 56 Discussion We identified stable hotspots of asymptomatic parasitaemia, and unstable hotspots of febrile disease in each of three cohorts in Kilifi District. The hotspots of asymptomatic parasitaemia were characterized by nonsignificantly higher mean antibody titres, a lower mean age at febrile episodes, and lower overall incidences of febrile disease. There may have been an increase in the incidence of febrile malaria in the penumbra around the perimeter of the hotspots of asymptomatic parasitaemia (Figure 5). Furthermore, hotspots of asymptomatic parasitaemia were stable over the full 7 y of monitoring, whereas hotspots of febrile malaria were not stable past 3 y of monitoring. Taken together, these observations suggest the following explanation: that a rapid acquisition of immunity in stable high transmission hotspots offsets the high rates of febrile malaria that would otherwise result. Instead, a high prevalence of asymptomatic parasitaemia is seen [51]. Unstable hotspots are not associated with prior exposure, and hence relatively low levels of immunity, and therefore produce higher incidences of febrile disease. Although the unstable hotspots are directly associated with more febrile disease, hotspots of asymptomatic parasitaemia may be critical in maintaining transmission [52]. The satellite remote sensing data suggest a primarily environmental cause for stable hotspots, but it was unclear why some hotspots are unstable. It was not simply a seasonal effect, since there was no indication of a seasonal specificity of hotspots in pooled analysis by season. Potential explanations might be a changing prevailing wind direction [53] or the temporary presence of “superspreaders” as seen in other infectious diseases [54]. Insecticide-treated net (ITN) use has expanded in Kilifi generally [35] and in the Junju cohort specifically [55], but there was no evidence that there were hotspots of ITN use in our dataset (p = 0.9). The study is based on active case detection in a sample of the total population (5%–10%). Data from the missing homesteads might alter the dimensions and frequency of the hotspots identified, and clarify whether clustering is stronger at the level of individual homestead or at the level of groups of homesteads that form “hotspots.” The resolution of remote sensing data results in an average of six homesteads per pixel. The remote sensing data predicted hotspots, but heterogeneity in risk by individual homestead was also observed, and higher resolution remote sensing data are needed to further investigate this. The instability of hotspots of febrile malaria was not simply due to variations in age of the children monitored (Table 3). Furthermore, stable hotspots of asymptomatic parasitaemia were observed in Junju, where only children and not adults were recruited to the cohort. Although the three cohorts monitored were from geographically different areas, the study is limited by presenting data from a single district, and so cannot represent the great diversity of ecology that will be seen across sub-Saharan Africa. Other cohorts undergoing longitudinal surveillance should be examined to confirm our findings. Malaria transmission varies by geographical features such as altitude [13], cultivation practices [14], streams and dams [16], house construction [56], socioeconomic factors [57], and ITN use [58]. Data on these factors were not available for the analysis presented here, although there were no large water bodies in the study area. However, a previous analysis in one of the study areas demonstrated that both environmental factors at the homestead level and host genetic factors contributed substantially to variation in malaria risk [10]. Our analysis suggests that environmental factors identified by remote sensing are associated with stable hotspots of asymptomatic parasitaemia. The environment in Kenya is seasonal, with two rainy seasons per year. The strongest individual factors from remote sensing were not the means of any index, but rather the minimum, and phases for a vegetation index and indicators of temperature, consistent with previous studies that have demonstrated the importance of temporal monitoring [12]. The remote sensing data show strong cross-correlation, and significant individual factors are likely to be proxies for more complex environmental determinants of transmission. 48 different measures of remote sensing were tested, but after a Bonferroni correction p = 1.4×10−5 and p = 2×10−4 for the two most significant individual factors. The evidence for environmental causation of the unstable hotspots of febrile malaria was less strong. Nevertheless, p = 0.006 after a Bonferroni correction for the single individual factor retained in the final model. On principle component analysis the second and third components were significantly associated with hotspots of asymptomatic parasitaemia, but the first component was associated with hotspots of febrile malaria. This finding is consistent with a complex environmental causation of hotspots rather than a single factor, and different environmental causes for the two types of hotspot. Anti-AMA antibody titres were associated with hotspots, but seroprevalence and seroconversion rates did not predict hotspots. This association contrasts with previous findings on a larger geographical scale [59]. The effects of individual variation may become more noticeable on a smaller scale [60], and smaller scale hotspots are unlikely to be as stable as environmental features such as altitude that operate on a larger scale [61]. There was substantial heterogeneity of transmission inside and outside hotspots. This heterogeneity was not simply random variation, since the previous year's incidence per homestead was strongly predictive of subsequent risk (r = 0.57). Variation in malaria risk at the homestead level has been shown to be due to a mixture of genetic and environmental factors [10],[11]. Irrespective of the cause, hotspots have a substantial effect on overall transmission in the community [2]. Here, we demonstrate two levels of clustering; clustering at the homestead level, and hotspots of groups of high risk homesteads in 1.3-km radius areas. The “gradient” effect away from the perimeter of a hotspot is consistent with transmission in the majority of the cohort being maintained by transmission from within the hotspot of homesteads. Hence, targeting the 73% of febrile malaria within the 28% highest risk homesteads is likely to benefit the wider community, since the highest risk homesteads will increase transmission in the surrounding area [1]. The genotyping data suggest that individual parasite clones are associated with a hotspot of asymptomatic parasitaemia. This finding may be because particular hosts more frequently transmit their parasites within a hotspot [54], or because particular parasite clones have adapted to a geographical micro-environment, determined, for instance, by the vector species [62]. Genotypes were not available from febrile disease episodes. Clustering of episodes by individual is reported in many infectious diseases [1], and has been relevant to the control of diverse pathogens such as Escherichia coli [54], tuberculosis [63], gonorrhea [64], SARS [65], and leishmaniasis [66]. Clustering of malaria episodes by homestead [10] and larger geographical hotspots including groups of homesteads is well described [3]–[9],[67], but there are few data on the temporal stability of hotspots. Targeted strategies for malaria control need to consider two kinds of hotspot; stable hotspots of asymptomatic parasitaemia and unstable hotspots of febrile disease. One might argue against intervening in hotspots of asymptomatic parasitaemia, since rates of febrile disease are not high, and intervention might reduce host immunity. However, where transmission has fallen in areas of high transmission, substantial direct and indirect mortality and public health gains have been described in the short and long term [35],[68],[69]. Furthermore, these stable hotspots probably feed transmission in their penumbrae for a distance of several kilometres because of vector dispersion [24]–[29]. Hotspots of asymptomatic parasitaemia can be identified by parasite surveys, serological surveys, or, more conveniently, remote sensing. Hotspots of febrile disease may be targeted by monitoring presentations to the local dispensary during the dry season, and targeting the effected homesteads during the subsequent rains. The optimal protocol would be to monitor during the dry season in September and then treat for the following year, which would result in targeting 20% of the homesteads, accounting for 65% of the febrile malaria episodes during the following year. It is optimal to cover 100% of the homesteads with any control intervention (particularly ITN distribution). However, some other interventions are currently not practical on community-wide scale, such as repeated mass drug administration [70],[71], environmental modification [72]–[74], mass vaccination [75], or (in some circumstances) indoor residual spraying (IRS) [76]–[78], but become feasible if targeted on the 20% of homesteads at greatest risk. When transmission has reduced to very low levels, transmission remains geographically clustered [9],[79], and intensified control in these hotspots is key to achieving elimination. In our setting, as in much of sub-Saharan Africa, there is not an immediate prospect of elimination. Additional targeted control measures may be viewed simply as a cost-effective way of ensuring that those most in need get the intervention, but the stronger justification is that reducing transmission in hotspots will reduce transmission on the wider community [1]. Environmental interactions are complex in determining malaria risk per se [12], and so these findings should be validated in other datasets before firm recommendations for malaria control programmes are made. Supporting Information Figure S1 Receiver operator characteristics (ROCs) are shown for AMA-1 antibodies, MSP-2 antibodies, and for the model on the basis of selected remote sensing variables and principle component analysis of the remote sensing variables. The areas under the ROC curves were 0.61, 0.67, 0.68, and 0.68 for prediction of febrile malaria hotspots by AMA-1 and MSP-2 antibodies, and for the model and principal component analyses, respectively. For asymptomatic parasitaemia hotspots the areas under the curves were 0.70, 0.72, 0.82, and 0.73, respectively. (0.61 MB TIF) Click here for additional data file. Figure S2 The incidence of febrile malaria episodes cpy is shown for the three cohorts. More intense green colouring indicates higher incidence. (0.48 MB TIF) Click here for additional data file. Figure S3 The prevalence of asymptomatic malaria is shown for the three cohorts. More intense green colouring indicates higher incidence. (0.47 MB TIF) Click here for additional data file.
