Increasing malaria hospital admissions in Uganda between 1999 and 2009

Uganda is a landlocked country traversed and bordered by large inland water bodies. Plasmodium falciparum transmission is intense and largely perennial and 1 of the 15 highest transmission intensity countries worldwide [14]. At 7 sites across Uganda entomological inoculation rates (EIR) from locally dominant vectors, mostly Anopheles gambiae s.s., ranged from 4 to over 1,500 infectious bites per person per annum [15], the latter recorded at Apac represents the highest published estimate of EIR in Africa. Areas with lower transmission intensity are located along the high altitude areas bordering Democratic Republic of Congo and Rwanda, the slopes of Mount Elgon in the East, arid areas in the North and the central areas of the urban extents of Kampala [16‐18]. The national estimate of age-under-5 mortality has been modeled to be approximately 145 per 1,000 live births for 2000 and predicted to have declined to 117 per 1,000 live births by 2010 [19, 20].

In 2001 a 5-year national malaria strategy was launched, which had as its goal to 'prevent and control morbidity and mortality and to minimize social effects and economic losses attributable to malaria in the country' [21]. The strategy included one measurable health outcome of reducing by 3% malaria case fatalities by 2005 and included a process outcome of ensuring that at least 50% of children are protected by an insecticide-treated net (ITN) by 2005 [21]. In 2005 the national strategy was rewritten, retaining a similar goal but explicitly aiming to contribute to reducing age-under-5 mortality from 152 to 103 per 1,000 live births with the number of under 5 s protected by an ITN increased to 85% by 2010 [18].

Since 2004, three successful applications to the Global Fund have resulted in approximately $200 million committed and $146 million disbursed to Uganda to support a wide range of commodity needs and implementation activities by the end of 2009 [22]. Between 2005 and 2009 the President's Malaria Initiative (PMI) provided $110 million to promote better coverage of insecticide treated nets (ITN) though free distribution (since 2007), mounting indoor-residual house-spraying activities in selected districts and internally displaced camp populations, the nationwide improvement of malaria case management and drug management supply and improved training for antenatal care providers to support intermittent presumptive treatment [23]. UNICEF and bilateral donors including the US pre-PMI, the UK, Belgium and Canada have all contributed approximately $10.6 million since 2004 for malaria control activities in Uganda [24].

Despite documented widespread chloroquine (CQ) resistance across the East African region, including Uganda, by 2001 [25, 26], this drug was widely used in combination with sulfadoxine-pyrimethamine (SP) as first-line therapy in government clinics and promoted as a national strategy to provide a home-based management package (Homapak^®) between 2003 and 2005 [27‐29]. Both CQ and SP clinical failure rates reached exceptionally high levels by 2005 [30, 31] and the policy for malaria case management was finally changed in April 2005. In 2006 the national malaria treatment guidelines were revised to support the introduction of artemether-lumefanthrine (AL) as first line therapy [32‐34] and effectively rolled out to all clinics by April 2006. However, AL funding, tendering and drug management supply to the periphery has been less than perfect since 2004 [35, 36]. Between May and August 2007 it was shown that 75% of health facilities in four districts had experienced a stock out of one of the four AL packs, including 34% of facilities that had stocked out of all treatment packs at least once simultaneously in the preceding 6 months and 13% of facilities did not have any AL packs in stock on the day of survey [36]. The early implementation problems may in part explain why only 3% of fevers among children reported treatment with AL in 2006 [37]. However, in 2008, between September and October 58% of the Public Health Facility had experienced a stock out of at least one of the four AL packs in the 3 months prior to survey and on the survey day 17% were out of stock of at least one of the four AL packs [38]. Private, retail sector availability of AL or other artemisinin-based products is uncommon; a study of 1,398 retail outlets across 38 districts in 2008 found no routine retail outlet stocked AL or any artemisinin-based combination therapies (ACT) and only 4.3% of drug stores stocked AL [38]. However in 2009 two sample household surveys showed that 21% [39] and 23% [40] of childhood fevers were treated with AL at any time during the febrile illness occurring within the last 2 weeks.

In 2000 less than 1% of Ugandan children were protected by an ITN [41, 42]. Between 2001 and 2003 there was a slow increase in activities to socially market or use voucher schemes to promote ITN distribution, however, ITN coverage among children only marginally increased to 9.7% during a national household sample survey undertaken in 2006 [37]. From 2007, using PMI support, free mass distribution campaigns and provision free of charge of ITN to antenatal clinics has resulted in an increase in children protected by ITN as evidenced by the results of the most recent national household sample survey undertaken in 2009 that showed ITN use among children had risen to 32.8% [40]. Indoor residual house spraying (IRS) activities have targeted areas of unstable transmission including internally displaced persons and refugee camps. Before 2006 IRS activities were limited. The first large-scale pilot IRS campaign supported by PMI was introduced in the epidemic-prone highland district of Kabale in 2006 and was subsequently repeated in 2007 and 2008 this time including Kanungu district [43] and extended to several districts in the North. From March to April 2008, pilot IRS campaigns with DDT were carried out in the high transmission districts of Oyam and Apac and of the households identified for spraying, it was reported that the campaign covered over 90% of the targeted population [44].

There are an estimated 65 government-supported hospitals (2 national referral hospitals, 13 regional referral hospitals and 50 general hospitals) providing inpatient pediatric care services across Uganda. For this study five relatively high pediatric admission burden hospitals were identified in consultation with the Ministry of Health to reflect the diversity of malaria transmission across Uganda [15, 45] and were most likely to provide records of admissions over the last 11 years (Figure 1). The hospital sites were representative of the national hospital distribution and previous drug sensitivity district selections by the Ministry of Health [25], with the exception that no hospital was selected in the concentrated urban population living in Kampala. Mubende hospital is located in a moderate-to-high transmission area of the Central 2 region; Jinja is located in the East Central region bordering Lake Victoria and represents an area of historically moderate transmission and the hospital serves a higher urban community than the other four hospitals; Kambuga hospital in Kanungu district serves a population that borders the Democratic Republic of Congo in the South Western Region and characterized by low, highland transmission intensity; Tororo is situated in the Mid-Eastern region on the Kenyan border and supports high perennial transmission; and finally Apac hospital is located centrally in the Mid-Northern region and experiences exceptionally intense transmission.

Pediatric ward inpatient registers at all hospitals were identified for most months from January 1999 to December 2009. Each admission entry in the registers was recorded on a tally sheet indicating the month of admission and whether an admission diagnosis of malaria had been defined for the child or whether the admission diagnosis did not include an indication of malaria.

Exact dates of birth were not available for all admissions and we have assumed all pediatric admissions were aged between 0 and <15 years. Individual register entries were not reconciled with patient notes and we have assumed that the admission diagnosis remained the diagnosis upon which each admission was managed clinically. Slide confirmed malaria diagnoses at admission were not universally available within and between hospital sites for the surveillance period and this continues to pose a significant information gap [46]. Therefore our working definitions of 'malaria' were patients admitted with a diagnosis of malaria, probably managed clinically as malaria during their admission but without documented parasitological confirmation.

For each possible 132-month series at each hospital, data were missing for 6 months (5%) at Jinja, 8 months (6%) at Kambuga, 13 months (10%) at Tororo, 14 months (11%) at Mubende and 21 months (16%) at Apac because registers were missing or torn. The missing data were handled using the 'ice' command in Stata (v.11.1; Stata, College Station, TX, USA) by creating multiple imputed data sets for missing values using other existing variables and prior information (the number of malaria and non-malaria deaths or admission as appropriate) to create an imputed value for each of the missing months. The same algorithm was used across all sites with all missing data imputations performed per sites (that is, using data from that site only).

It is important to standardize changing admission rates with time between hospitals with respect to the catchment population and natural growth in populations served by each hospital. Defining precise catchment areas is non-trivial and we have previously used addresses of pediatric admissions to explore the spatial nature and form of catchment areas at 17 Kenyan hospitals [8]. Similar information and detailed, high-resolution settlement mapping were not available in Uganda. We have therefore used an average radial distance of 30 km (that captured over 90% of admissions in Kenya) for the purposes of standardizing the catchment areas of the five Ugandan hospital sites. With the exception of Jinja, which is a regional referral hospital, all the others are district hospitals and the defined catchment area are inclusive of the subcounties where health workers in each hospital say majority of their patients originate. This radial boundary was used in combination with a 100 × 100 m human population density map developed from fine resolution satellite imagery, land cover and available information on human population counts from the highest resolution census administration boundaries [47, 48]. Population counts within a 30 km radius of each hospital were extracted in ARCGIS 9.2 (ESRI, Inc., Redland, CA, USA) for the Ugandan census date of 2002. Other methods have been used elsewhere [8] to define catchment populations but depend on the availability of reliable data on residential location of admitted patients. There is remarkably little published data on distances travelled for severe illness requiring hospitalization and this is an area that demands further investigation. Meanwhile, 30 km has been shown to be a reasoned catchment distance for hospitals in Kenya and therefore elected for use here in Uganda. District-specific population age structures and natural annualized growth rates [49] were used to project forwards and backwards to provide an estimated 0-14-year-old population estimate for each of the survey years 1999 through to 2009. Malaria and non-malaria incidence was computed using the estimated children aged 0-14 years resident in the hospital catchment area. Confidence intervals around each annualized admission rate per 1,000 children aged 0-14 years were computed using the Poisson distributions [50].

In addition we have collapsed average annualized admission incidence rates across three key time periods: (a) a period that corresponds to the widespread use of CQ-SP, little progress toward scaled ITN/IRS coverage and limited external donor financial support (1999-2003); (b) a period of transition to AL implementation, moderate increases in ITN coverage and improved external donor assistance (2004-2006); and finally (c) a period that witnessed a large increase in external funding from The Global Fund to Fight AIDS, Tuberculosis and Malaria (GFATM) and PMI, a concerted effort to provide free ITN distribution including mass campaigns in 2007, the introduction of IRS in selected districts (including Apac and Kanungu) and an established single recommendation for AL and a withdrawal of CQ-SP (2007-2009). Furthermore, data on the coverage of interventions during these periods are available from geolocated cluster sample survey data from national surveys to explore the plausible correspondence between changing prevention and case management strategies and disease incidence between time intervals and between hospital settings (Table 1).

To examine the long-term trends in admission rates we used monthly malaria admission data, standardized for under 15 population densities within each catchment areas, with moving average smoothing methods to identify any long-term trend signals within each temporal 11-year series while filtering out short-term annual fluctuations and random variation. The smoothing technique achieves this by replacing each element of the time series by n neighboring elements, where n is the width of the smoothing 'window' equal to 12 months. We employed a centered moving average including six observations before and five after the time point and vice versa, and taking an average value of this to compute the centered average. The ARMAX ('AutoRegressive Moving Average model with eXogenous inputs') model was then applied [51], which is an autoregressive model of the empirical current value of the series against one or more prior values combined with a moving average of the current value modeled against the white noise (random shocks) of one or more prior values and includes explanatory variables. The established explanatory variables for malaria trend analysis included rainfall in the preceding months (at different significant timelags), assembled from meteorological stations located close to each hospital, and changes in service use (captured by non-malaria admission rates) resulting in a predicted or smoothed malaria admission rate per month for each hospital site over the period 1999 to 2009. Correlations between rainfall at different timelags and malaria cases were assessed before including significant variables in the final ARMAX model. Before specifying the ARMAX model, tests and diagnostics were performed in the estimation of the models. Monthly malaria rates were tested for stationarity using the augmented Dickey-Fuller test [52] with a lag of 12 months. Model diagnostics and selection criteria were used to determine the most parsimonious model. Two goodness-of-fit criteria were used to guide model selection: Akaike information criterion (AIC) and Schwartz's Bayesian information criterion (BIC). The models with the lowest AIC and BIC were finally selected. All analysis was undertaken using Stata V. 11.0.

Information was assembled on 146,473 (72%) malaria admissions and 58,115 (28%) admissions where a diagnosis of malaria was not made at admission among children aged less than 15 years of age at the 5 hospitals over 11 years (January 1999-December 2009). Corrected malaria admission rates standardized for population estimates within each hospital catchment over the entire period varied between hospital sites, ranging from 6.9 per 1,000 children 0-14 years per year in Kambuga to 20.3 per 1,000 children per year in Apac (Table 1). These differences between hospitals were similar between the three important time periods and none of the annualized period incidence estimates were significantly different between time periods within each hospital site with the exception of a higher incidence at Kambuga during the second period (2004-2006) and a general tendency at all sites for the malaria admission rates to be higher in the last period, 2007-2009, compared to the first period 1999-2003 (Table 2). Non-malaria admission rates across the three time periods demonstrated either no changes at Mubende and Kambuga, a slight increase at Tororo and marginal declines at Jinja and Apac. With the exception of one period in Jinja, none of these differences were significant with overlapping confidence intervals in all estimated time period rates (Table 2).

To examine the long-term temporal signals within the monthly time series several model forms were explored. The predictions from the best fitting model that adjusted for non-malaria admission rates and rainfall at different lags while controlling for autoregressive and moving average effect within the data are shown in Figure 2. We included linear trend lines fitted to estimated average monthly changes in malaria admissions. The relationship between malaria admissions by month show increases across all five sites, although this appears less clear in Kambuga where there were exceptional peaks in malaria admission rates in 2005/2006. Significant (P < 0.001) upward trends in malaria admission rates were observed in Mubende, Jinja, Tororo and Apac over the study period (Figure 2). By 2009 the relative proportional increase in malaria admission cases from 1999 was greatest in Mubende (+350%) with a significant peak in 2009, but important increases were seen in all other sites when compared to 1999 at Jinja (+200%), Tororo (+200%), Apac (+188%) and Kambuga (+47%) (Figure 1).