Malaria paediatric hospitalization between 1999 and 2008 across Kenya

In recent years several African countries have managed to rapidly scale up the delivery of key malaria control measures, notably insecticide treated nets (ITN) [1] and changing policies to support the provision of efficacious Artemisinin-based combination therapies (ACT) for malaria case-management [2]. Documenting expanding coverage is important to reconcile against international donor and government financial investment, however increasing the numbers of individuals protected and adequately treated must be linked to changes in malaria disease burdens, if future malaria control investment is to be sustained. At several sites across Africa there have been recent attempts to attribute changing patterns of malaria morbidity to scaled intervention coverage [3–14], malaria hospitalization [3–6, 9, 11, 13–18] or malaria-attributable mortality [3–5, 7, 11, 13, 14, 16, 19–21].

The popular perception is that the epidemiology of clinical malaria is in transition across Africa [22–25]. A comprehensive assessment is hampered by the selective reporting of only good news stories or the occasional editorial biases toward positive effects in peer-reviewed journals [22]. To date reports of changing disease risks have been limited to observations from selective sites limiting generalisability [5, 11, 16–18]; or reported at country-levels ignoring the heterogeneity within countries [2, 9, 14, 26]; or inadequately adjusted for missing data, health service use and populations-at-risk [7, 9, 12, 15].

Here we present a country-wide assessment of changing paediatric admissions to 17 district general hospitals across Kenya using novel approaches to adjust for the populations served by these hospitals, general service use and the periodicity effects of rainfall over a 10 year period between 1999 and 2008.

Between January 1999 and December 2008 information was assembled on 245,810 malaria admissions and 204,128 admissions where a diagnosis of malaria was not made at admission or discharge among children aged less than 15 years of age at the 17 hospitals. Imputing missing values for 45 months using multiple imputation techniques and correcting for the estimated catchment population, rates of malaria admission between 1999 and 2008 varied between hospital sites, with the lowest in the arid areas of Makueni and Wajir and the highest at Kisii in the highlands and Hola located in a predominantly arid area but along the Tana River (Additional file 2: Table SI 1). Overall admission rates were within the range of 1.49 cases per 1,000 children aged 0 to 14 years in Wajir to 21.01 cases in Kisii for malaria admission while non-malaria admission rates were between 2.66 and 19.24 cases per 1,000 children aged 0 to 14 years. Preliminary descriptive data of averaged annualized admission rates for malaria and non-malaria that were congruent with three important control policy time-lines are shown in Figures 2, 3, 4, 5, 6 and 7 (individual time-hospital specific rates provided in Additional file 2: Tables SI 2 and SI 3).

At hospital sites located in the Western/Lakeside region of Kenya malaria admission rates varied little between the three time periods (notably Bungoma) with some evidence of rising rates, remaining elevated in the second (2003 to 2005) and third periods (2006 to 2008) at Busia, Bondo and Siaya (Figures 2 and 3). The exceptions were at Kisumu and Homa Bay where malaria admissions rose in the second time-period and declined in the third. Among hospitals located in the Highlands, both Kericho and Kisii showed a modest decline in malaria admission rates in the third period with much smaller declines at Kitale during the same period (Figure 4). The three hospitals along the coastal region of Kenya show earlier declines in malaria admission rates with lower rates documented in the second period (2003 to 2005) compared to the first time period (1999 to 2002) and further declines in the third period (2006 to 2008) compared to the second period (Figure 5). Hospitals located in areas that are regarded as arid or semi-arid showed variable changes between the three time periods (Figures 6 and 7). At Makeuni and Wajir hospital admission rates with malaria were consistently low across the three time periods. At Narok the rate of malaria admissions doubled between the first and second time periods and remained unchanged during the third time period. Both Hola and Voi however showed similar categorical time-period changes to those in the highlands with no change until the final time period 2006 to 2008 where malaria admission rates more than halved.

Differences were observed in non-malaria admission rates across the three time periods (Figures 2, 3, 4, 5, 6, 7; Additional file 2: Table SI 3). Overall these changes demonstrated a converse pattern to those of malaria admissions with either no changes or marginal declines in rates of non-malaria admission across the three time periods at Busia, Kisumu, Siaya, Kericho, Kitale, Kilifi, Hola, and Voi. At Bungoma, Homa Bay and Wajir rates of non-malaria admission showed an overall decline by 2008. At Bondo, Kisii and Msambweni rates of non-malaria admission declined by 2003 to 2005 compared to the first period (1999 to 2002) but rose again in the last period 2006 to 2008. Rates of non-malaria admission rose across the time periods at Makueni, Narok and marginally at Malindi. The categorical time-defined patterns of paediatric admission for diagnoses that were not malaria were overall not the same as those described for rates of admissions that included malaria as a diagnosis (Figures 2, 3, 4, 5, 6, 7). These variations were used to accommodate general temporal changes in hospital service use during the period in subsequent time-series models.

In general, all-cause admission trends closely matched the temporal patterns of admission observed for malaria diagnoses across almost all study sites (Figures 2, 3, 4, 5, 6, 7). All-cause admission rates varied little between the three time periods at nearly all hospital sites along Lake Victoria and in Western Kenya with the exception of Kisumu and Homa Bay where admissions rose in the second time period and declined in the third (Figures 2 and 3). In the highlands, Kisii showed a significant decline in all-cause admission rates in the third period with much smaller declines at Kericho and Kitale during the same period (Figure 4). In the three hospitals along the Kenyan coast, Kilifi and Kwale showed early declines in all-cause admission rates with lower rates documented in the second period (2003 to 2005) compared to the first time period (1999 to 2002) and further declines in the third period (2006 to 2008) compared to the second period in Kilifi; but remaining unchanged in Kwale (Figure 5). At Malindi declines in all-cause admissions occurred only in the last period. Hospitals located in the arid/semi-arid areas also showed initial rise in the second period followed by a decline in the last period in Hola and Voi; rises in both period two and three in Narok while remaining relatively unchanged in Wajir and Makueni (Figures 6 and 7).

We have examined the temporal patterns of paediatric malaria admissions at 17 hospitals across the diverse malaria transmission ecology of Kenya between January 1999 and December 2008. At two hospital sites located in arid areas (Makueni and Wajir) hospitalization from malaria (and non-malaria) was extremely low and showed no obvious changes during the observation period. At 10 of the remaining 15 sites reductions in malaria admission rates were observed. These were significant after adjusting for rainfall and general hospital service use. The overall reduction in the paediatric malaria admission rate was on average 0.0063 per month since January 1999 by the end of the 120-month time-series. These sites included the three areas located along the Kenyan coast, three areas in the highlands, two semi-arid areas: Voi and Hola (along the Tana river), and two sites located in the west of Kenya: Homa Bay along the shores of lake Victoria and Bungoma bordering Uganda. Homa Bay and Bungoma, however, showed small proportional reductions in monthly malaria admissions over the 120-month time-series and not as marked as the remaining eight sites showing overall declines in malaria admissions (Table 2; Figures 2, 3, 4, 5, 6, 7). Despite statistical significance of the time-series adjusted analysis of malaria admissions, largest reductions with the clearest trends were seen at Kisii, Kilifi, Malindi and Msambweni. Reductions at the other six hospital sites were less marked and subject to large within-series fluctuations in admission rates. At Kericho an exceptional peak in malaria admissions in 2002 was an anomaly (Figure 4); the large initial peak in malaria admissions at Bungoma contributed to the perceived modeled decline in risk with time (Figure 2); and at Kitale, Homa Bay, Voi and Hola, malaria admissions rose from 1999 to 2003 through 2005, but were then off-set against declines in admission rates between 2006 and 2008 (Figures 3, 4 and 6). This was also true of Kisumu that showed no overall significant rise or decline across the combined 120-month time series (Figure 4).

A more consistent pattern of increasing adjusted malaria admission rates over the observation period was observed at Busia, Siaya and Bondo district hospitals where the rise in malaria admission rates was between 0.51% and 0.66% per month by the end of the time-series in December 2008 (Table 2). Following a rise in admissions from January 1999, Siaya exhibited a drop in malaria admission rates after January 2005 but rose again through 2008 (Figure 3).

The analysis assumed that the trend component was linear; meaning that the level of the predicted series increases by a constant amount each time and this is valid across the majority of the data series. However we acknowledge that in certain instances, the series do not exhibit a constant time trend and are extremely variable moving up or down over the course of a few years as appears to be the case in Hola and Kisumu. An additional caveat is that we have assumed a constant annualized growth in population size using inter-censal estimates from 1989 to 1999; it was not possible to derive a more precise and temporally consistent growth rate in the absence of the 2009 national census and therefore we cannot be certain of changing population size and structures.

It is equally important to recognize the 17 hospitals selected here were those that could provide enough information over a long enough time period to evaluate trends in paediatic admission and only one hospital had consistently monitored the presentation of malaria with microscopy and detailed clinical examination of all admissions (Kilifi District Hospital). Thus the majority of the data analyzed here are not based upon parasitological confirmed cases of malaria. These data were often not available and of variable quality and accuracy [32, 49]. Changes in clinical diagnosis of malaria with time may have contributed a systematic, non-random error in the data series reflecting periods of perceived epidemics in the highlands or perceived absence of malaria following wide-spread reporting of malaria declines in Kenya by the media. However, in the past malaria has been the predominant clinical and parasitological confirmed diagnosis made at Kilifi district hospital where rigorous screening of paediatric admissions prevailed throughout the period of observation [16] and accounted for 27% of all diagnoses made in children aged less than 15 years in the 1999 to 2008 series included here. Assuming malaria is a significant contributor to admissions to hospital among children aged less than 15 years, it is hard to clinically define, and assuming all else being equal after adjusting for increasing populations in the catchment areas and inter-annual changes in seasonal rainfall, one would anticipate a decline in all-cause paediatric admission rates from 1999. This was the rationale behind the large-scale, community randomized trials of insecticide treated nets during the 1990s where all-cause childhood mortality was the principal clinical outcome in Phase III trials [50]. Under these assumptions the only significant declines in all-cause admission rates were seen to any great extent at Homa Bay, Kisii, Kilifi, Malindi, Msambweni and Hola with an overall reduction in admission rates of 0.0050 or more per month since January 1999. These observations would suggest that apparent reductions observed for malaria admissions in these six sites are true reductions and was not an artifact of misdiagnosis.

We have applied a rigorous set of time-series models to an extended series of hospital admissions in an attempt to provide some externality to earlier observations of declining malaria admissions at one hospital (Kilifi [16]) and at three hospitals (Kilifi, Msambweni and Malindi [15]) on the Kenyan coast using more covariates against standardized populations. Providing more data from more sites is important to validate and test general patterns of health transition. In doing so we have shown that, despite the difficulties in interpreting time-series hospital data, patterns observed on the Kenyan coast are not replicable across the country. In general terms malaria admissions along the coast had begun to decline before the wide-scale introduction of prevention strategies using ITN and the introduction of new, effective therapeutics into the government health service. In other areas of Kenya malaria admission rates had risen from 1999 to 2006 but showed a variable decline after 2006. Whether this was due to scaled intervention coverage or not remains uncertain and would require a much longer post-2006 time series to be able to rule out short-term periodicity in malaria transmission. Several sites showed plausible evidence of increasing malaria admission risks across the entire time-series and thus do not support the hypothesis of health impact following scaled ITN delivery or the switch from failing to effective treatment. Interestingly there was some suggestion that areas with more pronounced declines in malaria admission rates were areas with lower intensity parasite transmission at the end of the surveillance period compared to areas with higher ending transmission intensity that showed no change or a rise in cases of malaria during the interval. While our study did not examine changing infection prevalence with time, theory does suggest that the clinical impact of interventions targeted at reducing parasite exposure may be more immediate in areas of lower transmission intensity [51, 52]. With a higher baseline, starting transmission intensity the impact on parasite exposure, and subsequent disease outcomes will be slower and will initially transition through a period of higher burden as older children become increasingly at risk due to changing age-specific immune profiles [51, 52] and will only change when transmission intensity has reached an ill-defined point. We are exploring this hypothesis separately with better data on infection prevalence transitions and changing intervention coverage in selected sites included here.

Because Kenya has made significant progress at a national level on changing its drug policy to support effective case-management [43] and represents one of only a few countries in Africa that have massively increased ITN coverage since 2006 [1] we had anticipated a replication of observations made on the Kenyan coast [15, 16] at other hospital sites across the country. This was not the case and our data are an important reminder that all is not equal and single, spatially constrained observations within a country may be misleading. In this regard, it is important to highlight the fact that efforts geared towards improved intervention coverage may not have been implemented uniformly countrywide.

A country-wide assessment of trends in malaria hospitalizations indicates that all is not equal, important variations exist in the temporal pattern of malaria admissions between sites and these differences require more detailed investigation to understand what is required to promote a clinical transition across Africa.