Spatial-temporal patterns of malaria incidence in Uganda using HMIS data from 2015 to 2019

The global burden of malaria has declined since 2000 primarily due to the scale-up of control interventions, including long-lasting insecticidal nets (LLINs), indoor residual spraying with insecticide (IRS), and use of artemisinin-based combination therapy (ACT) [1‐3]. Nevertheless, incidence rates in sub-Saharan Africa remained high at an estimated 219 cases per 1000 in 2017–2018 [3]. The incidence estimates used to monitor trends across sub-Saharan Africa are typically generated using parasite prevalence in children 2–10 years fitted in prevalence-to-incidence models [3]. Though informative, the surveys included happen infrequently [4] and may be limited in scale. Derived burden estimates, therefore, cannot adequately support day-to-day monitoring for decision making at national or sub-national levels [5].

National malaria control programmes typically depend on routine health management information systems (HMIS) data to guide programme decisions in control and elimination efforts. With the advent and extended access to web-based health information systems, such as the District Health Information System - version 2 (DHIS-2), timely access to nation-wide HMIS data and quality of these data have been shown to have greatly improved in sub-Saharan Africa [6, 7]. As such, the WHO has reiterated that timely and high-quality HMIS-based burden estimates are achievable, and can be used to inform on-going decision making [8]. Despite this, HMIS remains underutilized, especially for risk mapping, due to concerns over incompleteness and delayed reporting [3, 9, 10]. Whilst HMIS has had, and still needs, improvement, substantial discrepancies between estimates of malaria burden from the current prevalence-to-incidence model approach and HMIS-based reports persist among at least 30 high burden countries [3]. Thus, questions remain as to the reliability of HMIS-based estimates and their corresponding representation of fine-scale spatial distribution of risk to support evidence-based decision making by country-level programme managers.

Small area space-time disease models fitted to routinely reported data have been widely implemented to accurately identify contextually important risk factors and unpack spatial-temporal patterns of infectious diseases, including tuberculosis and malaria [11‐15]. These models have the capacity to explain the spatial autocorrelation in disease data, and can provide robust means of understanding ecological connectivity and relationships [16] that are critical for control processes in high malaria or other disease burden countries. Moreover, foci of high malaria risk or burden are pertinent to the principle of strategic information to drive impact under the global high burden to high impact initiative, for effective targeting of interventions [17]. This study therefore, aims to investigate a pragmatic novel small-area space-time approach using a nationwide network of health facilities in estimating malaria incidence from HMIS data, in order to identify areas of high malaria burden and risk across Uganda and assess malaria seasonality.

Results from this innovative, large-scale, longitudinal observational study suggest that with improved HMIS reporting, credible high-risk areas at both high and low spatial scales were identifiable. The study revealed a distinct monthly spatial distribution of malaria incidence across the 15 regions of Uganda, in a concurrent multi-resolution assessment, including coarse (regional) down to fine (health facility catchment) spatial resolutions. Moreover, whilst Uganda is considered a perennial transmission setting, this study revealed a nation-wide seasonal pattern in incidence rates with two peaks (major and minor), the highest during June–July and the minor peak during October. This approach may facilitate efficient implementation and optimization of targeted control activities that can leverage existing health facility systems [37]. It may also improve managers’ understanding of the heterogeneity and/or clustering of malaria burden within districts that currently form the lowest level of malaria burden assessments, though acknowledged as difficult to use or unusable for control planning [5].

This study showed that the risk of malaria by regional rank, among the highest and lowest risk regions, had minimal temporal variability. These regions maintained their status both during low and high burden seasons. These findings were consistent with extant UDHS regional stratification of Uganda where Acholi, West Nile, and Karamoja were among the highest transmission regions, and Ankole and Kigezi among the lowest. This stratification supports tailored approaches for long-term malaria control efforts aiming at elimination, as advocated in the global ‘high burden to high impact’ initiative [17] that was recently adopted as central to onward national malaria control strategies for Uganda [38]. Whilst targeted interventions, including IRS [39] and larval source management [40] have been used, further emphasis is necessary [17, 41] with implementation taking greater account of local context. Importantly, however, temporal variability of risk among many regions highlights the continued vital role of routine surveillance for planning and timely action towards control. Moreover, higher risk among high burden locations during the lowest than highest burden seasons suggests persistent high-risk in these locations, the identification of which could facilitate high precision targeted actions for effective control.

This study also identified several distinct clusters of high-risk health facility catchments, which were consistent over time though largest during the highest burden seasons and smallest at the lowest. The largest high-risk clusters were concentrated in the West Nile and Acholi regions in Northern Uganda, although smaller clusters were noted in the recognised high transmission regions of Karamoja and East-Central Busoga [19]. Conversely, the most notable low-risk health facility catchment clusters could be grouped into three categories: highland regions (e.g. Kigezi, Ankole and Bugisu) [42, 43]; regions with recent intense targeted multi-year IRS activity associated with high impacts on transmission (e.g. Bukedi, Teso, and Lango) [4, 19, 44, 45]; and, large urban municipalities (e.g. Southern Buganda) with urbanization associated with reduced transmission [46, 47]. These findings provide further evidence of identifiable candidate locations for targeted control interventions among the high-risk clusters and an approach for assessment of possible impacts of previous interventions.

Trends in annual confirmed malaria cases in Uganda declined between 2016 and 2018, despite increased reporting and proportions of confirmed cases over time, consistent with MIS findings between 2014 and 2018 [4, 19], before a sharp increase in 2019. Moreover, the relationship between regional relative risk and prevalence of malaria (among children under 5 years of age from the 2018 MIS) showed that small changes in parasite prevalence were associated with sharp increases in relative risk among regions at lower than national average risk. However, large changes in parasite prevalence were associated with small changes in relative risk among regions at higher than national average risk. This further confirms the variability of risk among many regions while pointing to strong effects of age on malaria [48]. Besides the estimated confirmed cases being lower than estimates reported by WHO and MAP per year (possibly due to study design of excluding some facilities), trends were also dissimilar with WHO and MAP cases increasing between 2016 and 2017 [18], unlike in this present study. Nevertheless, such dissimilarities have been documented [3] and are likely explained by the use in global assessment for sub-Saharan Africa of prevalence surveys that to date, are predominantly conducted among children [49]. With estimates for the whole population generated from these surveys, despite shifts in malaria burden from children to the older population following effective control interventions [48], the dynamic effects on burden may not be adequately accounted for in the prevalence-to-incidence models used.

The observed seasonality with June–July peaks and February–March troughs was consistent with reports from south western Uganda, where epidemics followed a regular July pattern except during El-nino in 1998 [5, 50] and in Gulu district (Northern Uganda) where between 2006 and 2015 biannual peaks of malaria were reported during June–July and October–November [51]. One study however, reported two peaks of malaria during April–May and September–November in Northern Uganda following the rain seasons, though unsubstantiated [52]. Findings from this present study may inform optimal timing for control activities, including IRS, mass drug administration (MDA), or community mobilization campaigns towards increased malaria risk awareness for control vigilance.

PFP facilities, a small majority of which do not report to the HMIS and were therefore excluded from this study, limit the utility of focal analyses such as presented here. This highlights an important missed surveillance opportunity. The limited capacity to detect outbreaks in settings largely served by PFP may exacerbate the severity of malaria outcomes among their most vulnerable residents, coupled with increased case management costs [53]. There are several possible initiatives to increase reporting in these facilities where a small majority seek care for febrile illnesses [4, 19, 54]. First, provision of guarantees on exclusive use of data for public health not revenue monitoring, may improve confidence and alleviate any fears of punitive intensions in their reporting. Second, ensured availability of standardized reporting tools, may offset running costs of stationery in the private facilities while it enables improved documentation of health records. Third, training of PFP managers and owners on the benefits of surveillance and/or reporting may increase their involvement. Lastly, implementation of regular feedback mechanisms may provide a means of continued evaluation that fosters risk and other assessments that are mutually beneficial.

Given that policymakers’ remediating responses as well as policy formulation processes are informed by pooled information from diverse sources, including but not limited to research, political, and funding provisions, it is unrealistic to expect these technocrats to be expert generators of the evidence from these multi-disciplinary sources. Whilst there are no simple solutions to the implementation of analyses such as in this present study, interpretation of contemporary outputs is nowhere nearly as demanding, highlighting the criticality of partnerships between policy and research dimensions for malaria and other disease control efforts.

This study had limitations. First, the disproportionately low proportion of geolocated reporting private facilities impacted on the estimates of malaria burden, especially among highly urban locations, including Kampala and Wakiso districts and others across the country. Results for the Kampala region (and Wakiso district) in this study, represent only a small proportion of the burden and were excluded from results discussions. Moreover, exclusion of non-geolocated reporting public health facilities (such as in Kitgum district), impacted on the estimates of incidence due to unidentified catchments in those places. Nevertheless, there was wide coverage of health facilities across the country with a small proportion of districts under-represented, minimizing effects of this constraint. Second, the study did not account for level of health facility and other population level factors that impact on differential health seeking behaviour, which may have inflated incidence rates and risk, where a given level or type of facility is preferred. However, in this analysis it was assumed that for uncomplicated malaria, people attend the closest health facility and some important factors such as urbanicity and primary care giver education were accounted for, though further research may be required to better understand impacts of level of health facility on care seeking for uncomplicated malaria. Third, the study did not account for stock levels of antimalarials or test kits, variations of which may impact on the number of cases recorded between seasons of full stock versus stockouts. A better understanding of the linkage between logistics management and HMIS may be required, given known associations between stockouts and increased under-five mortality or compromised treatment practices like dosage rationing and use of less effective remedies [55]. Fourth, given that health facility recruitment into the study was not dynamic, any increase in number of facilities reporting could have had impacts on study findings. Moreover, the systematic exclusion of non-geolocated facilities, may have biased study results towards more long-term established than newer health facilities, but duration of facility existence was beyond the scope of this study.

Assessment of malaria burden and/or risk in high burden countries using routine surveillance data is highly achievable. Using national routine data, this study provided needed evidence of vital concurrent assessment of malaria risk and burden among regions, districts, and health facility catchments with identifiable significant spatial clustering of risk. Targeting hotspots as an intervention approach has been shown to yield modest and transient impacts on malaria prevalence [56]. However, locations with persistently high-risk of malaria that are potential candidates for health facility-based interventions such as community outreaches, provision of LLINs, mass drug administration and enhanced case management were identified, an approach that may be beneficial beyond isolated health facility catchments. Furthermore, whilst extensive geo-spatial analytical output with scales either too large (region or district) or too fine (pixel or neighbourhood) may be challenging for control programmes to use [57], this study provides evidence of HMIS-based assessments at practical scales for districts to implement and assess intervention impacts. Moreover, in perennial settings, the identifiable strong seasonal patterns as seen with June–July highest peaks and February–March lowest troughs in Uganda, provide vital information for intervention timing. Taken together, these results show the potential in routine HMIS surveillance data for pragmatic timely identification of high-risk areas and with further research, the assessments for optimal implementation of targeted control activities and their impacts.