The contribution of non-malarial febrile illness co-infections to Plasmodium falciparum case counts in health facilities in sub-Saharan Africa

Background The disease burden of Plasmodium falciparum malaria illness is generally estimated using one of two distinct approaches: either by transforming P. falciparum infection prevalence estimates into incidence estimates using conversion formulae; or through adjustment of counts of recorded P. falciparum-positive fever cases from clinics. Whilst both ostensibly seek to evaluate P. falciparum disease burden, there is an implicit and problematic difference in the metric being estimated. The first enumerates only symptomatic malaria cases, while the second enumerates all febrile episodes coincident with a P. falciparum infection, regardless of the fever’s underlying cause. Methods Here, a novel approach was used to triangulate community-based data sources capturing P. falciparum infection, fever, and care-seeking to estimate the fraction of P. falciparum-positive fevers amongst children under 5 years of age presenting at health facilities that are attributable to P. falciparum infection versus other non-malarial causes. A Bayesian hierarchical model was used to assign probabilities of malaria-attributable fever (MAF) and non-malarial febrile illness (NMFI) to children under five from a dataset of 41 surveys from 21 countries in sub-Saharan Africa conducted between 2006 and 2016. Using subsequent treatment-seeking outcomes, the proportion of MAF and NMFI amongst P. falciparum-positive febrile children presenting at public clinics was estimated. Results Across all surveyed malaria-positive febrile children who sought care at public clinics across 41 country-years in sub-Saharan Africa, P. falciparum infection was estimated to be the underlying cause of only 37.7% (31.1–45.4, 95% CrI) of P. falciparum-positive fevers, with significant geographical and temporal heterogeneity between surveys. Conclusions These findings highlight the complex nature of the P. falciparum burden amongst children under 5 years of age and indicate that for many children presenting at health clinics, a positive P. falciparum diagnosis and a fever does not necessarily mean P. falciparum is the underlying cause of the child’s symptoms, and thus other causes of illness should always be investigated, in addition to prescribing an effective anti-malarial medication. In addition to providing new large-scale estimates of malaria-attributable fever prevalence, the results presented here improve comparability between different methods for calculating P. falciparum disease burden, with significant implications for national and global estimation of malaria burden. Electronic supplementary material The online version of this article (10.1186/s12936-019-2830-y) contains supplementary material, which is available to authorized users.

Estimates of the proportion of malaria infections above and below a pyrogenic parasite density threshold were derived from active case detection studies. Analysis was restricted only to active case detection studies as these fully enumerate all febrile illnesses within the study site. The parasite density threshold dataset comprised a subset of the exhaustive dataset of P. falciparum active case detection studies compiled by Battle et al, 1 updated to include all studies published until the end of 2017. The restricted set of studies included active case detection studies, where blood parasite densities of infected individuals were collated, and matched, site-specific, ageappropriate PfPR was recorded over the longitudinal study. Full inclusion criteria for this analysis were as follows: i) parasite densities of symptomatic individuals were given (either as raw values or above/below a case definition parasite density threshold; for studies that presented multiple case definition thresholds the threshold closest to 5,000 parasites/μl was used), ii) the study reported only clinical incidence amongst children between 0 and 5 years of age. For each study, the total number of microscopically-diagnosed P. falciparum positive febrile individuals above and below the malaria case definition threshold of 5,000 parasites/μl was collated, for consistency with previous models. 2 A total of 11 studies met the inclusion criteria, resulting in 31 active case detection records. Full details of the studies utilised are given in Additional File 2.
Datasets 2-4: Empirical estimation of MAF Datasets 2-4 consisted of paired observations of malaria and fever positivity, whereby the proportion of malariaattributable fevers within malaria-positive fevers was derived empirically. Under this framework, the number of malaria-attributable fevers within a study site, , is given by Where + + is the number of individuals who are both P. falciparum-positive and fever-positive, + is the total number of P. falciparum-positive individuals, − + is the total number of individuals who are both P. falciparum-negative and fever-positive, and − is the total number of P. falciparum-negative individuals. This framework can be applied to any study site where paired observations of fever and malaria are available. In this analysis, three datasets were compiled and used for empirical estimation of MAF. These datasets are labelled "datasets 2-4" and are described in turn below.

Dataset 2: Household survey data
This dataset was used for empirical estimation of MAF and was composed of 41 cross-sectional, nationallyrepresentative georeferenced surveys of malaria prevalence in children less than five years of age across 21 countries conducted between 2006-2016 in sub-Saharan Africa, obtained from the Demographic and Health (DHS) Program. 3 Full details of the surveys used can be found in Additional File 3. In each of these surveys, interviewers visited houses (selected as a geographically and demographically representative sample of the population) and a finger-or heel-prick blood sample was taken from any children under five years of age present and tested for malaria with an RDT. For each child receiving a malaria diagnosis, their caregiver was asked whether, in the past two weeks, the child had a fever. Diagnostic and fever history outcomes were collated from each child, with children aggregated at geo-located cluster-level to represent community-level groups, resulting in a total of 156,670 paired observations of malaria positivity and fever history in the same individuals. A limitation of the household survey data is that an individual who was malaria-positive at a previous point during the two weeks preceding the interview may have sought and received antimalarial treatment, and their blood antigen concentration may have reduced sufficiently to appear RDT-negative by the time of interview (referred to as the "treatment effect" in the main manuscript). The treatment effect is accounted for in both models described in subsequent sections.

Dataset 3: Literature review
This dataset was used for empirical estimation of MAF and was composed of paired observations of malaria positivity and fever collated from published articles. Two search strategies were utilised to identify these studies. First, every article published between 2006 and 2017 cited within the Malaria Atlas Project database (the largest repository of global malaria prevalence data, updated annually via exhaustive reviews of published literature), 4 was systematically searched for paired observations of fever and malaria positivity. Second, a systematic literature review of the PubMed database using the MeSh terms "malaria", "fever" and "diagnosis" was conducted. For both data sources, studies assessing a population in a malaria-endemic community were collated, using the following inclusion criteria: i) malaria diagnosis was conducted (using any diagnostic), and ii) diagnostic outcomes were paired with individual fever status, whether measured at the time of diagnosis or individually recalled recent fever history. For each study meeting the inclusion criteria, the location of study was recorded and the number of individuals in each community falling in to each of the following four categories was documented: i) P. falciparum positive and fever positive, ii), P. falciparum positive and fever negative, iii) P. falciparum negative and fever positive, and iv) P. falciparum negative and fever negative. The review of the MAP PfPR database and the PubMed review yielded 71 and 23 observations, respectively, of community-level paired malaria and fever measurements, comprising a total of 123,649 individuals.

Dataset 4: Program for Resistance, Immunology, Surveillance and Modeling of Malaria in Uganda (PRISM) study
The final database utilised for empirical estimates of MAF was sourced from the Program for Resistance, Immunology, Surveillance and Modeling of Malaria in Uganda (PRISM) study. 5 The PRISM study collected data over five years (2011-2016) on malaria-related metrics across three sites of differing malaria transmission intensity in Uganda, with the aim of improving understanding of the complex relationships between malaria parasite, host, and vector. In each of the three Ugandan sites, approximately 100 households were enrolled between 2011 and 2013 and followed up longitudinally for malaria infection. The malaria diagnostic outcome of all individuals at enrolment was collated, and paired with measured fever (axillary temperature > 38 o C) at the time of diagnosis.

Modelling approach for estimating relationship between MAF and PfPR0-5
For this analysis, a model was applied where the relationship between MAF in children under 5 years of age and PfPR0-5 was learned. The DHS Program data (described above as Dataset 2) was used to inform the shape of the relationship, and the remaining three datasets described above (Datasets 1, 3 and 4) were subsequently used to learn scaling factors, as these datasets are informative for higher values of PfPR0-5 but lack MAF observations where PfPR0-5 <0.1. Additionally these datasets were unbiased by the "treatment effect" of DHS Program data, described in the previous sub-section.
We assume this relationship takes the following form, the motivation for which was informed by localised regression fits to the response data detailed above: Where is the number of individuals in a given population with a malaria-attributable fever, is the number of individuals in a given population with a malaria-positive fever (regardless of fever causality) and is the proportion of malaria-positive fevers that are causally due to malaria. It was assumed that this proportion is negatively exponential and is shaped by three parameters, , and , where controls the minimum proportion of malaria-positive fevers in a population that can be due to malaria, and controls the rate of decline with increasing PfPR0-5 until the minimum proportion of fevers, , is reached. The relationship is scaled by the parameter .
represents the proportion of individuals who are P. falciparum-positive (whether febrile or not). The model used a numerical optimisation function to learn the parameters , , and (plus an additional standard deviation/observational noise parameter, σ) to fit the relationship between malariaattributable fevers (as a proportion of malaria-positive fevers) and PfPR0-5 in a two-step process: first, all four parameters were optimised using only the DHS Program dataset, allowing the shape of the relationship to be learnt; second, and were fixed to the optimised values identified in the first model fit using the DHS Program dataset, and the model re-fit and σ, using the remaining datasets. Using this method, the intercept and rate of decline in was learnt using the DHS Program dataset (i.e. Dataset 2), and then the relationship was re-scaled by the non-DHS Program datasets (i.e. Datasets 1, 3 and 4: the parasite-density literature review dataset, the paired observations of fever and malaria dataset, and PRISM dataset, as detailed above). For direct comparability between datasets, only observations for children under five years of age were included. All modelling described above was completed in in R using the TMB package. 6 Details of the literature review and PRISM datasets can be found in Additional File 4.

Additional factors investigated
Additional factors which may impact the relationship between MAF and PfPR0-5 were also investigated. First, recent declines in PfPR0-5 may affect the relationship due to residual exposure-dependent immunity, so DHS Program data were segregated by the magnitude of declines in PfPR0-5 during the two years preceding the survey (measured at cluster-level for each 5km x 5km pixel) using cartographic model-based geostatistical estimates for 2000-2016 7 and the model was fit to each data subset to investigate differences. Second, the parasite density threshold dataset (used for the secondary model calibration) was based on active case detection studies, meaning that the prompt and effective treatment rate for malaria cases is typically close to 100%. 2 The treatment rate in all the other datasets was far more variable. The potential effect of differing treatment rates on the relationship between MAF and PfPR0-5 was investigated by segregating the DHS Program data by the treatment-seeking rate within each cluster.