Summary of the results
This geo-epidemiological study identified a high diversity of landscapes within the METF region corresponding to a gradient from intact (i.e. dense forest) to heavily human-modified landscapes, characterized by different types of agriculture. Within this diversity of landscapes, three landscapes (DSC, DSG, SG) and a single climate were associated with malaria-affected profiles. Three eco-epidemiological zones with contrasting dynamics were identified by combining these environmental risk factors. One of these zones exhibited longer persistence of P. falciparum, up until 2018; another zone showed high P. vivax incidence after 2018; and the third displayed stable seasonal P. vivax incidence without P. falciparum.
The role of climate was consistent with the literature
One major climate (Midland climate) was linked to malaria-affected profiles irrespective of species or spatial scale. This climate presented optimal conditions for mosquito and parasite development, according to the literature: moderate rainfall between April and October, an annual mean temperature between 18 and 32 °C, and limited temperature variation [27, 28].
Forest is not a homogeneous environment against malaria
Residing in villages at the forest fringe or inside forested areas was identified as a risk factor for malaria infection in the GMS [29]. Forest covers 76% of the surface of the METF region. Of the 19 landscapes (17 at the region scale and two specifically identified in the Northern Township), 13 landscapes (70%) had ≥50% of their surface covered by forests. An association of sparse and dense forest only was identified as the closest landscape for most villages. This illustrates how in this study area, most of the villages were at the forest fringe or inside the forest. Yet we previously identified heterogeneous malaria transmission within this region [14]. Only three landscapes were associated with malaria-affected incidence profiles, confirming the heterogeneity of forested environments relatively to malaria risk.
The interface between forest and agricultural lands is at high risk of malaria in the Northern Township
In the Northern Township, DSC and DSG landscapes were associated with malaria-affected incidence profiles. They stood out due to their high proportion of forest (≈90%) fragmented by patches of cropland or grassland/shrubland. Another landscape (Sparse forest, cropland) that was not associated with incidence profile, was highly fragmented but had a higher cover of cropland (32%) and a lower proportion of forest (58%), especially dense forest (20%). The other landscapes had the same proportion of forest cover, but the forest was less fragmented by agricultural patches. In the literature, the role of forest fragmentation in malaria transmission, specifically with respect to the type and proportion of forest or agriculture, remains unclear [30,31,32,33]. These findings highlight that, in the Northern Township, a balance between three components (the relative proportions of forest cover, forest fragmentation and agricultural fields) defines at-risk landscapes for malaria, which combine sustained human presence in farms and Anopheles-favourable environments in nearby forest [34,35,36].
In addition, we identified specific dynamics in relationships with specific agricultural types. Long P. falciparum persistence and stable P. vivax incidence (zone 1) corresponded to cropland located in broad valley bottoms, indicative of inundated rice paddies. Shorter P. falciparum persistence and increasing P. vivax incidence (zone 3) corresponded to patches of grassland/shrubland located on slopes, a pattern matching traditional Karen taung yar agriculture, a rotational farming, 10-year cycle which is practised on slopes [37]. This suggests that the type of agriculture in forested areas could be used as a proxy for topographical and land-use conditions that support heterogeneous dynamics and malaria persistence. It is also possible that inundated rice fields, when located at the forest fringe, directly support higher transmission.
In the Northern Township, the persistent malaria burden in children under the age of 5 years suggests that transmission may still take place within villages, while the lower incidence in adults may be explained by partial immunity.
The specific context of the southeast border
In the southeast border area of the study region, eco-epidemiological zone 4 stood out due to its proximity to the SG landscape associated with highly seasonal P. vivax cases in adult males and few P. falciparum cases. Two factors may have contributed to this unique dynamic. First, the predominance of P. vivax over P. falciparum likely resulted from longer access to early diagnosis and treatment for malaria, which depletes the P. falciparum reservoir faster than that of P. vivax [2, 12]. Second, a combination of environmental changes and occupational exposure may also explain this specific epidemiology. In contrast with the Northern Township, the landscape associated with this dynamic has been mostly covered by sparse forest and agricultural land since intense deforestation between 2004 and 2010 [26]. The exact occupation responsible for higher malaria exposure is unclear, but we hypothesize that it may be related to seasonal farming activities carried out outside villages. Such activities usually involve adult males (e.g. seasonal migrant workers), who live on the farm over prolonged periods from the onset of the rainy season until harvest. The vectors to which they are potentially exposed may originate from small patches of sparse vegetation (e.g. along streams) rather than rarer, dense forest or forest fringes.
Study strengths
Using the unique dataset of weekly malaria case reports collected by the METF MP network, we were able to study the association between malaria incidence and forested environments in unprecedented detail. This dataset allowed the study of malaria incidence at village scale without requiring further spatial aggregation. Compared to the annual parasite index, incidence profiles allowed the temporal characterization of intra- and inter-annual variation in amplitude, seasonality, and tendency over a 4-year period. They enabled us to identify villages that were long-term hotspots or showed malaria persistence. Plasmodium falciparum and P. vivax dynamics could also be studied in parallel or combined. This was especially relevant as METF interventions were targeted at P. falciparum and had less impact on P. vivax, leading to different persistence patterns.
In addition, malaria transmission through mosquito bites can occur within villages or in remote locations where there is human activity and the precise location of infection is usually unknown. It is therefore difficult to estimate the distance between households and transmission sites, as this depends on the particular activity carried out (e.g. agriculture, logging, etc.), geographical accessibility (topography, road access) and means of travel. To overcome this, we described the climate and landscape over the entire region regardless of the location of the villages. We also relied on high-resolution land use and land cover data, and used fragmentation indices, which are key determinants of landscapes at risk for malaria in the Northern Township. Fragmentation indices are important for the characterization of interfaces between types of natural environment, such as forest, and human-modified environment (e.g. cropland), and where the probability of contact between vectors and humans is higher.
To study the association between malaria incidence profiles and environment, we chose to use a tree-based model and random forests to overcome intrinsic issues associated with regression models, i.e. small sample size for some incidence profiles, unknown distribution, and potential correlation or interaction between covariates (i.e. climate and landscape).
This analysis ultimately identified eco-epidemiological zones, which included villages within a similar risk area but with heterogeneous malaria incidence profiles. Beyond following World Health Organization recommendations concerning malaria risk stratification (i.e. based on transmission and receptivity), this approach is of interest in an elimination context where outbreaks are more stochastic, as illustrated by P. falciparum incidence profiles [9]. These results could help in the targetting and planning of surveillance by identifying villages where outbreaks are more likely. They could also be used for the planning of routine activities to target additional effort where it is needed.
Study limitations
The first challenge in this study was the overall low malaria incidence in the METF region over the study period. Indeed, 81% of the villages were classified as having a very low incidence profile for P. falciparum, and 69% a very low incidence profile for P. vivax. The other 10 profiles only corresponded to a small number of villages (from one to 46). The imbalance in these distributions limited the subsequent analysis of factors associated with each incidence profile and the use of regression models. The CIT analysis mostly distinguished malaria-affected from malaria-free incidence profiles, except for P. vivax incidence in the Northern Township. However, by combining environmental factors associated with both Plasmodium species, we were able to describe sub-regional eco-epidemiological zones, which corresponded to specific local environments and were characterized by specific trends and at-risk populations. These results suggested a relationship between malaria dynamics and environment at a spatial scale higher than that of the village. These findings agree with previous reports of P. falciparum prevalence hotspots clustering within a radius of 10 km in this region [12].
We relied on the incidence of clinical cases of malaria diagnosed by MP as a proxy of malaria transmission. This assumption likely led to the underestimation of transmission, since individuals with sufficient immunity will not necessarily develop clinical malaria episode upon infection. On the other hand, using the incidence of clinical cases of malaria to determine transmission, we could not distinguish P. vivax infections resulting from an infective bite from those due to a relapse. This could have resulted in an overestimation of P. vivax transmission, or inability to identify some seasonal patterns (e.g. zone 1). However, in the analysis of 1441 recurrent episodes pooled from two trials conducted in the same setting, 95% of P. vivax recurrences (88% relapses and 12% reinfections) were detected upon active follow-up, while only 5% were associated with sufficient symptoms to trigger an intercurrent consultation [38]. In these trials, active participant follow-up may have led to early detection and preventive treatment of some relapses before they became clinical. Despite this potential underestimation, clinical relapses are thought to only moderately modify the incidence dynamics.
Finally, even if the LULC classification allowed a 10-m resolution, validated with field observations, there were still some classes that were difficult to distinguish, i.e. roads from rivers, and agricultural patches from low vegetation. These elements of the environment may modulate vector presence and thus the association of environmental factors with incidence profiles. In addition, this study focused on land cover patterns, such as fragmentation and vegetation density. Identifying differences in floristic composition across forested areas was not feasible here, but could allow for further characterization of forest heterogeneity.
Generalizations and perspectives
The aims of this study of malaria dynamics at the local level were to improve our understanding of the epidemiology of this disease in an increasingly heterogeneous context and to contribute to intervention planning. The analysis identified eco-epidemiological zones with higher malaria incidence where specific interventions could be targeted (for more details, see [14]). These zones aggregated villages sharing similar environments, and thus receptivity, beyond local incidence heterogeneities. The use of this type of zoning is valuable for the identification of areas prone to malaria resurgence, where surveillance is strategic. Indeed, at low transmission levels, favourable environmental conditions are necessary, but not sufficient on their own, as the presence of a parasite reservoir is also required.
Health systems often rely on the collection of aggregated surveillance data (geographically, temporally, and by age). This study demonstrates how rich surveillance data could inform an understanding of malaria epidemiology and intervention planning, especially in a highly heterogeneous context. Our results support the current trend of investing in state-of-the-art epidemiological information systems. However, the size of the eco-epidemiological zones reported here suggest that analyses performed at the scale of a health facility catchment area (e.g. a dozen villages) could yield relevant results for larger regions or country-wide studies.
Additionally, eco-epidemiological zones were characterized by differences in climate, landscape, and agricultural systems, which are biologically relevant factors that may be applicable to other forested areas. Studying structural environmental factors at a broader scale in the GMS could help identify areas favourable to Anopheles presence and determine proxies of human exposure. Combining existing lower resolution LULC with other remotely sensed data sources (forest loss data, or the detection of logging or fires leading to deforestation) could provide community-level assessments of receptivity and human activities in or at the fringe of forested areas [39].