Forest malaria in Cambodia: the occupational and spatial clustering of Plasmodium vivax and Plasmodium falciparum infection risk in a cross-sectional survey in Mondulkiri province, Cambodia

After a marked reduction in malaria burden in Cambodia over the last decades, case numbers increased again in 2017–2018. In light of the national goal of malaria elimination by 2025, remaining pockets of high risk need to be well defined and strategies well-tailored to identify and target the persisting burden cost-effectively. This study presents species-specific prevalence estimates and risk stratification for a remote area in Cambodia. A cross-sectional survey was conducted in 17 villages in the high-incidence province Mondulkiri in the dry season (December 2017 to April 2018). 4200 randomly selected participants (2–80 years old) were tested for Plasmodium infection by PCR. Risk of infection was associated with questionnaire-derived covariates and spatially stratified based on household GPS coordinates. The prevalence of PCR-detectable Plasmodium infection was 8.3% (349/4200) and was more than twice as high for Plasmodium vivax (6.4%, 268) than for Plasmodium falciparum (3.0%, 125, p < 0.001). 97.8% (262/268) of P. vivax and 92.8% (116/125, p < 0.05) of P. falciparum infections were neither accompanied by symptoms at the time of the interview nor detected by microscopy or RDT. Recent travels to forest sites (aOR 2.17, p < 0.01) and forest work (aOR 2.88, p < 0.001) were particularly strong risk factors and risk profiles for both species were similar. Large village-level differences in prevalence of Plasmodium infection were observed, ranging from 0.6% outside the forest to 40.4% inside. Residing in villages at the forest fringe or inside the forest compared to outside was associated with risk of infection (aOR 2.14 and 12.47, p < 0.001). Villages inside the forest formed spatial hotspots of infection despite adjustment for the other risk factors. Persisting pockets of high malaria risk were detected in forested areas and in sub-populations engaging in forest-related activities. High levels of asymptomatic infections suggest the need of better case detection plans and the predominance of P. vivax the implementation of radical cure. In villages inside the forest, within-village exposure was indicated in addition to risk due to forest activities. Village-level stratification of targeted interventions based on forest proximity could render the elimination efforts more cost-effective and successful.


Background
In the last two decades, Cambodia has seen a marked decrease in numbers of reported malaria cases from approximately 140,000 cases in 1999 to 60,000 in 2018 [1]. Likewise, mortality by malaria steadily reduced, with zero deaths reported for the first time in 2018. Similarly remarkable reductions have been observed in the other countries of the Greater Mekong Subregion [2]. Since 2014, these countries have even pursued the goal of elimination of all malaria by 2030 [3].
This ambitious aim however faces several obstacles. After an all-time low of approximately 20,000 reported cases in 2016, numbers have increased again in 2017-2018 [1,2]. While the annual case numbers increased for both Plasmodium falciparum and Plasmodium vivax in 2017, they declined for P. falciparum in 2018 but increased further for P. vivax [1,3]. Consequently, P. vivax now accounts for almost three-quarters of all malaria cases in Cambodia [2,4].
The national malaria control programme (NMCP) is based on mass distribution campaigns of long-lasting insecticidal nets (LLIN), diagnosis by light microscopy (LM) and rapid diagnostic tests (RDT) in the public health sector, RDT-based testing by village/mobile malaria workers (VMW/MMW), and treatment of blood stage infections with artemisinin-based combination therapy (ACT) [5]. The spread of multi-drug resistant P. falciparum malaria in Western Cambodia [6] might explain the resurgence of P. falciparum cases in 2017 and the change of first-line treatment from dihydroartemisinin-piperaquine to artesunate-mefloquine their subsequent decline in 2018. For P. vivax, such resistance is not reported. However, this species has dormant liverstages called hypnozoites that escape acute blood stage treatment and cause relapsing infections. Therefore, although LLINs do reduce P. vivax transmission and ACT is effective against blood stage P. vivax infection, neither intervention targets the hypnozoite reservoir. Hence, there is a less marked effect of these interventions on P. vivax than P. falciparum.
The only licensed drug with hypnozoiticidal effect is primaquine, typically administered at 0.25 mg/kg daily over 14 days in Cambodia. The roll-out of P. vivax radical cure is part of the national treatment guidelines but has not yet been implemented because of concerns over the drug's potential haemolytic effect in patients with G6PD deficiency [7,8]. In summary, the control and elimination efforts of the NMCP are successfully targeting clinical P. falciparum malaria (understandably in the context of drug-resistant P. falciparum [6] in the country) but to a lesser extent to P. vivax.
Another obstacle to elimination is the high prevalence of asymptomatic infections [9][10][11][12]. Neither health facilities nor VMWs reach these asymptomatic carriers as they do not seek diagnosis or treatment. Although of lower infectivity [13,14], they remain however a potential source of onward transmission.
Alongside the sustained reduction in countrywide incidence, malaria has also become increasingly fragmented, affecting fewer, usually remote provinces. The highest incidence of malaria in Cambodia and across the Greater Mekong Subregion is now particularly found in the North-Eastern provinces such as Ratanakiri and Mondulkiri [1,3]. These settings are characterized by a high proportion of ethnic minorities and agricultural or wood-logging activities as the main income sources. Exposure to Anopheline mosquitoes and Plasmodium in Cambodia is understood to primarily occur in the forest rather than peri-domestically [15][16][17].
Together with the Western province Pursat, the highest incidence per province in 2018 was reported for Mondulkiri [1]. However, previous studies on population-level burden and risk of malaria in Cambodia focused on the country's West and North [12,[16][17][18][19][20][21] or the Eastern province of Ratanakiri [10,11,15,20,22,23]. Here, using a cross-sectional survey conducted in a rural area of Mondulkiri in the dry season in 2018, estimates of the prevalence of Plasmodium infection are shown and key risk factors identified.

Study area and census
Seventeen villages were selected in the Kaev Seima district, Mondulkiri province, in the Kingdom of Cambodia. The rainy season in Cambodia normally runs from June to October with the high transmission period from June through December [1,4,24]. Approx. two-thirds of the population in Mondulkiri is comprised of national ethnic minorities, with the Phnong ethnic group comprising the largest proportion [25,26]. In November-December 2017, a census was conducted by visiting each household in the 17 villages and collecting basic demographics of household participants such as age and gender. A person's household was defined as location of main residence in exposure was indicated in addition to risk due to forest activities. Village-level stratification of targeted interventions based on forest proximity could render the elimination efforts more cost-effective and successful.
Keywords: Forest, Occupational risk, Spatial, Vivax, Hotspots, Cambodia, Greater Mekong Subregion the village according to adult members of the household or the village head. GPS coordinates were collected using Garmin ® GPSMAP ® 64s devices. When no adult household member could be found, demographic information was obtained from the village head's registry book.

Cross-sectional survey
Based on the census, a random selection of households was drawn oversampling small villages to ensure sufficient coverage (Additional file 1: Table ST1). Selected households were visited from mid-December 2017 until mid-April 2018 and all household members aged 2-80 years who had resided in the study area for at least 3 months were invited to participate in the survey. Upon informed consent, a questionnaire on household variables was administered to the head of household or another adult household member by trained interviewers. A questionnaire on individual-level variables was administered to each consenting household member. Children were interviewed assisted by a parent (or rarely another caregiver in the absence of the parents). Data were collected on tablets and run through automated data quality checks within days after the interview. In case of missing data or discrepancies, the field team was informed for immediate resolution if possible. Fingerprick blood samples were collected as thick and thin film slides and in K + EDTA-microtainers. Participants were screened for symptoms, i.e. feeling sick or feverish on the day of interview, having felt feverish over the preceding two days, or having an axillary body temperature of at least 37.5 °C. Upon any indication, the participant was administered a standard malaria RDT (Malaria Ag P.f/P.v, Standard Diagnostics Inc., South Korea) and referred to a local health care provider for treatment if positive.

Detection of infection
The microtainer blood samples collected at the interview site were stored in 4 °C ice boxes. At a field laboratory, they were separated into plasma and cell pellet and frozen at −20 °C. Following transport to the main laboratory at Institut Pasteur of Cambodia in Phnom Penh, cell pellets were stored at −20 °C and plasma at −80 °C. Infections with any of the four human Plasmodium parasites were determined by real-time PCR [27]. In case of a positive genus-specific result, qPCR specific for P. falciparum, P. vivax, Plasmodium malariae, and Plasmodium ovale followed. All positive and a random selection of 10% negative samples were assessed by independent double LM readings of asexual and sexual stages and parasite densities calculated. No discrepancy in parasite densities of above 30% occurred. Parasites were counted per approximately 500 leukocytes and densities were inferred assuming 8000 leukocytes per microlitre blood.

Descriptive analyses
Villages were classified based on the forest cover in a 750 m radius around the households, computed from the land cover analysis in [28] (Additional file 1: Figure S1). Villages with ≥ 50% of households with ≥ 10% forest cover in their vicinity were considered "inside the forest", with ≥ 30% of households with ≥ 5% forest cover as "at the forest fringe", or "outside the forest" otherwise. Because of very low sample sizes, the two small, neighbouring villages Beng (11 individuals) and Gaty (95 individuals) were analysed as one. Population prevalence was estimated based on post-sampling weights assigned to each participant according to their representation by village, gender, and 10-year age bins compared to the census population (raw numbers of positive survey samples accompany the estimates in brackets). Categorical covariates were compared using the chi-squared test or the Fisher's exact test when low strata sizes required it.

Risk factor analyses
The association of covariates with infection by P. vivax, P. falciparum, or all four species was assessed by mixedeffects logistic regression with random intercepts per household and village. Those covariates that were statistically significantly associated with Plasmodium infection at two-tailed α = 5% in univariate regressions were included in the multivariate model. The villages' proximity to the forest, work-unrelated overnight travels (incl. to forest sites), and work in the deep forest were considered part of the multivariate model a priori in order to assess the association of forest exposure with risk of infection. Having slept outdoors or under a bed net last night were kept as part of the model based on causal reasoning or prior knowledge due to their association with infection risk in previous publications, e.g. [11]. Gender and age were retained in the model as proxies for riskrelated behaviour and thus potential confounders. The other covariates comprised potential proxies for socioeconomic status or further exposure variables and their subset significantly associated with infection risk was assessed by backwards variable selection. Akaike Information Criterion (AIC) was used to assess model fit. If collinearities and interactions were identified, covariates were retained comparing fit of the respective models by AIC. Statistical significance in the multivariate regressions was calculated per covariate by likelihood ratio tests. Spatial hotspots were identified by a purely spatial scan statistic via a discrete Poisson model with maximally a third of the population in a scanning window, adjusted for all covariates of the final multivariate model except for the villages' forest proximity [29].

Software
All questionnaires were applied on tablets using the REDCap (Research Electronic Data Capture) software hosted at Institut Pasteur in Paris [30,31]. Data quality control as well as all descriptive and analytical statistics were performed in R 3.6.3 [32]. The SaTScan ™ software version 9.6 was used for the spatial scan statistic [33].

Census, survey representativeness and survey population
From among the 10,053 individuals in 2351 households identified in the census, the survey recruited 4200 participants from 1147 households and oversampled smaller villages on average (Additional file 1: Table ST1). A map of the survey households and the categorization of villages by proximity to the forest is shown in Fig. 1a. Mean age was 26 years in both the census and the survey and with 51% (5135/10,053) and 53% (2231/4200) women, respectively, women were slightly oversampled in the survey. In men, ages of 16 to 40 years were mildly underrepresented (Fig. 1b). The main income source for the survey households was by large majority farming (89.8%, 1030/1147). In terms of mobility, three-quarters (75.9%, 3188/4200) of the survey participants reported work-unrelated trips to forest or field sites in the last month (8.5%, 358/4200) or any work trip in the last two months (75.8%, 3184/4200). By vast majority, these trips were short and frequent: In only 4.4% (140/3188) of the instances, stays for longer than a week were reported and among those who reported any work trip, 89.4% (2847/3184) went for work at least once a week.

Prevalence was highest in men of working age
Estimated prevalence was more than twice as high in men as in women, i.e. 10 Figure S3). While there was no difference in genus-wide prevalence across age in women, men showed an elevated risk at working age (p < 0.001), regardless of the proximity of the village to the forest (Fig. 3). A fitted interaction term of gender and age was significant in the strata of villages outside the forest (p < 0.001), but not across those villages at the forest fringe or inside the forest.

Risk profiles for P. vivax and P. falciparum infections were similar
Risk of infection was associated with individual covariates for both P. vivax and P. falciparum in a highly similar  (Table 4 for behavioural variables, full list of variables with odds ratios in Additional file 1: Table ST2). For both species, prevalence was associated with workunrelated overnight travels and highest if those occurred to forest sites (22.4%, 60/268, and 14.9%, 40/268, for P. vivax and P. falciparum, respectively) and lowest for urban destinations (1.7%, 1/59, and 0%, 0/59, for both p < 0.001). Infections of both species were also more prevalent among those who reported work trips to sites in the nearby and deep forest (assessed separately). In particular, a significantly higher prevalence was observed in participants who reported work trips into the deep forest (25.5%, 37/145, and 18.6%, 27/145) compared to those who did not (5.7%, 231/4055, and 2.4%, 98/4055,   for P. vivax and P. falciparum, respectively, p < 0.001).
Once clustering at household and village-level was taken into account in univariate mixed-effects logistic regression, work in the nearby forest was no longer significantly associated with risk of infection though (Additional file 1: Table ST2). There were fewer infections among those who reported the use of standard protection measures such as sleeping under a bed net (6.1%, 236/3880, and 2.7%, 104/3880, vs. 10.0%, 32/320, and 6.6%, 21/320, if no net was used, p < 0.01 and p < 0.001 for P. vivax and P. falciparum, respectively). A higher risk in men, at working age, in villages at the forest fringe or inside the forest, and with indicators of lower socio-economic status was also found similarly for both species (Additional file 1: Table  ST2).

Working in and travelling to the forest were strong risk factors of infection
Behavioural covariates related to activities in the forest were statistically significant in the multivariate model of Plasmodium infection (  Table ST3). Other covariates linked higher socio-economic status with lower odds of infection, such as living in a house with a roof built of relatively high-quality material (aOR 0.50, p < 0.01). The covariate on recent information on malaria via TV (aOR 0.37, p < 0.01) most likely also acts as a proxy for higher socio-economic status, i.e. being able to afford a TV in the first place.

Descriptive characterization of the risk factors forest work and forest travels
Work-unrelated overnight travels to forest sites was a predominantly male domain (reported by 10

Residing in a village inside the forest was an independent spatial risk factor
Living in a village inside the forest remained associated with risk of infection when adjusting for the significant demographic, socio-economic, and behavioural covariates (aOR 12.47, p < 0.001, Table 5). Households from all four villages inside the forest also formed hotspots of Plasmodium infection, i.e. purely spatial clusters of elevated risk of infection that cannot be explained by  the other covariates (Fig. 4, Additional file 1: Tables  ST4-ST5).

Discussion
This is the first detailed study on the prevalence of Plasmodium infection and associated risk factors on population-level in Mondulkiri province in Cambodia. The overall prevalence of Plasmodium infection is consistent with that observed in the neighbouring province of Ratanakiri [10,11,22] or other endemic provinces in the West of the country [16].
Plasmodium falciparum has long been the predominant species countrywide [9,34]. However, following a scale-up in the VMW programmes, a steady decrease in reported P. falciparum cases was observed in 2009-2011, while numbers of reported P. vivax cases increased [24]. In recent years, cases of both species were reported at an almost equal share [4] and in 2018 P. vivax accounted for approximately three quarters of reported cases [2]. That P. vivax started to predominate over P. falciparum in Cambodia even earlier is suggested by other studies in high-incidence provinces where molecular and serological diagnosis identified a higher prevalence of P. vivax than P. falciparum infections [11,12,[15][16][17][18][19], consistent with the presented study in Mondulkiri.
7.2% of the P. falciparum infections could be identified by symptoms and a positive result by LM or RDTs, i.e. the diagnostics used by the health system, compared to only 2.2% for P. vivax. Low sensitivity of RDTs for P. vivax is a commonly reported problem [12,[35][36][37][38]. In addition, higher proportions of asymptomatic infections for P. vivax are regularly observed as the relapses can lead to higher levels of immunity and thus overall lower parasite densities [39,40]. Data from studies using membrane feeding assays suggest that the densities of asexual stages and gametocytes in the asymptomatic, LM-positive samples in this survey are infectious to mosquitoes [14,[41][42][43][44][45]. It can thus not be ruled out that these subclinical infections contribute to the ongoing transmission in this area.
More asymptomatic infections, lower diagnostic sensitivity for P. vivax, high transmissibility, and the lack of radical cure by primaquine all might add up as possible explanations why the control efforts have been less successful against P. vivax than against P. falciparum. In order to address this high burden of P. vivax, Cambodia is now piloting the use of primaquine against P. vivax infections in four provinces. While an essential step, the effect of radical cure will be limited without point-of-care diagnostics that are more sensitive for P. vivax infections and ideally a test allowing the detection of hypnozoite carriage (irrespective of blood stage parasitaemia). New advances in the use of multiple antigens to detect antibodies as serological markers for recent exposure and thus for potential hypnozoite carriers are promising in this respect [46].
With as many as 40% of the inhabitants infected in the highest-prevalence village in this survey, it may be necessary to complement the current malaria control efforts by active or reactive case detection schemes [47,48] to reduce the infected (and potentially infectious) reservoir and eventually reach the goal of malaria elimination. Targeting such resource-intensive interventions to high-risk groups will render them significantly more cost-efficient. The similarity of the risk profiles for both P. falciparum and P. vivax presented here is encouraging as it suggests that such targeted interventions would appropriately address both of the two species. This study emphasizes the predominant role of forest malaria transmission in explaining elevated risk of infection in remote rural areas of Southeast Asia. Previous cross-sectional surveys have also identified an association of infection risk with either work in the forest [15,17,49] or time spent therein [16,50]. However, this study is unique in retaining them both as statistically significant risk factors in one multivariate model while also adjusting for gender, age, socio-economic proxies, and other behavioural covariates. While work in and travels into the forest are most frequent in adolescent and adult men who are usually considered the forest-related risk group, this behaviour also occurred in women and from children to participants in their sixties and thus actually extends to a much broader range of the population.
On the relatively small spatial scale of 17 villages in an area of 22 × 26 km 2 , the prevalence ranged from less than 1% to above 40% in the dry season. The trend towards higher prevalence in villages at the forest fringe and inside the forest compared to those outside the forest is in line with the widely accepted notion for forest-related mosquito abundance and exposure in Cambodia. Other studies have also identified a higher prevalence in villages at close proximity to the forest [21,49,50]. It is apparent that the association of risk with living near or inside the forest was retained in this study even if adjusting for demographic, socio-economic, occupational, and other behavioural differences.
Not more than 60% of infections could be attributed to the strong behavioural, forest-related risk factors. In villages outside the forest or at the forest fringe, most infections indeed occurred in working age men. This predominance of this occupational risk group was much lower in forest fringe villages and entirely lost in villages situated inside the forest. Taking gender and age as proxies for risk-related behaviour and occupational activities, this could indicate that in villages outside the forest, risk can be explained almost completely by such occupational exposure. By contrast, for villages within the forest, the more homogeneous presence of infection across all age groups in both males and females indicates additional exposure inside the village. This is in line with other studies in which household-level risk factors in forested areas were important to explain infection risk [10] or infections clustered in households over 2 years [51] in Ratanakiri, the other Eastern province of Cambodia.
The current WHO report on malaria eradication calls for subnational stratification of intervention programmes [52,53]. This study suggests that such a stratification of interventions could happen along a gradient of villages inside or outside the forest. Population-oriented interventions such as targeted mass drug administration or test and treat programmes with sensitive molecular or serological diagnostics [46,54] (together with their costs and risk of overtreatment) could be justified in forest villages. By contrast, approaches targeted at risk groups only may be effective at lower costs for the NMCP and the local population in residential settings outside of the forest.

Conclusion
Despite a substantial reduction in malaria burden in Cambodia over the last decades, this study demonstrates that pockets of high malaria prevalence persist in the country. Given that the vast majority of infections were asymptomatic, the study strengthens the argument to enhance malaria elimination efforts by measures of (re-)active detection of also asymptomatic infections. Plasmodium vivax infections were detected at a higher prevalence than P. falciparum infections, more often asymptomatic, and less detectable by RDT and LM. Consequently, novel tools for the identification of hypnozoite carriers could play a key role in the presented setting as well as the roll-out of routine radical cure treatment. The study corroborates the notion of forest malaria in the Greater Mekong Subregion by demonstrating the independent association of travels to the forest, of forest work, and of living in close proximity to forest with malaria infection risk. However, the study also demonstrates that infection risk is less confined to forest-goers in sub-populations that already live in forested areas and suggests within-village transmission therein. A focus of interventions solely on forest-goers could thus be insufficient to reach the goal of nation-wide malaria elimination in due time. In the light of the presented results, interventions could be targeted at whole villages inside the forest, and targeted at stratified risk groups for villages outside the forest.