Four interventions (Table 1) took place in highland environments, where peaks and valleys create a complex, climate-affected altitude gradient of malaria transmission and age-acquired immunity. Kisii and Gucha Districts are located in the southern highlands of Nyanza Province, Kenya, and experience low year-round transmission, with short, dramatic, and increasingly frequent outbreaks (16). Before May 1999, the last recorded epidemic had taken place from January through April 1998. The epidemic we describe also affected 10 nearby districts (17).

In Burundi's northern Kayanza Province, a 3-year time series up to September 2000 showed constant monthly caseloads of ≈10,000 outpatients/month. In 2000, MSF operated 7 of the province's 22 outpatient facilities. The September 2000–May 2001 epidemic, the largest ever recorded in Burundi, affected 9 of 16 provinces, and 3.5 million cases were reported (18).

The Ethiopian highlands experience 2 moderate transmission seasons every year (after rains in March through April and August through September). Epidemics occur in 5- to 8-year cycles; >1 million cases were recorded in 1998 (1). The 2003–2004 epidemic affected 15 million persons in 3 federal regions (19).

Finally, malaria is considered endemic in low-altitude Aweil East County (Bahr el Ghazal state, southern Sudan), although no data are available. Most cases occur from July through January after spring rains.

Findings on possible epidemic determinants are summarized in Table 1. Factors noted at all sites were drought in preepidemic years followed by above-average rainfall in the preepidemic months and elevated drug resistance. Individual sites also experienced temperature abnormalities, land pattern changes, and high malnutrition rates. Further detail on these findings is provided in the Appendix.

No early warnings were issued. In Kisii, the alert came from the media in epidemic week 5 (when the district hospital was overwhelmed with malaria cases). MSF issued alerts in Kayanza (doubling of fever cases in epidemic week 2, early exhaustion of antimalarial stocks), Aweil East (quadrupled inpatient and outpatient malaria after epidemic week 1), and Damot Gale (increased proportion of P. falciparum–positive test results among children admitted to feeding centers, Figure 1). No alert information was found for Gutten.

Formal epidemic declaration was hampered by missing data. Time series for historical comparisons were available in Kisii (12 years) and Kayanza (3 years), where, however, authorities initially suspected a typhus outbreak, until the 80% seroprevalence detected among febrile patients (epidemic week 7) pointed to P. falciparum. In Aweil East, comparison with the 2 past years was confounded by 1) a change in diagnostic strategy (from presumptive in 2001 to rapid test–based in 2002 and 2003) and 2) decreased access to treatment in 2003 because of flooding.

In Ethiopia, a malaria-specific surveillance system aimed for early outbreak detection at both the village (positivity >25% detected among slides collected by field workers in the community was considered an outbreak and theoretically led to village-level mass treatment and vector control) and woreda or zone (where weekly reports from health facilities were compiled) levels. A malaria epidemic was declared in August 2003 (6–10 weeks after probable onset) by East Wollega Zone, including Gutten. Data from this surveillance system were difficult to interpret. Only some of the reports reached the zone bureau, with a delay of 4 to 6 weeks. In Damot Gale, despite incomplete reporting, a massive increase in caseload was evident at the province level (6,500 from July through October 2003 vs. 1,233 from July through October 2002); however, by early July (epidemic onset), only reports up to April were available. Furthermore, clinics aggregated mass fever treatment and outpatient data, causing artificial incidence spikes when the former took place. Conversely, drug shortages in August gave a false impression of declining incidence. MSF had only been present in Damot Gale since April 2003 and only operated feeding centers.

Interventions occurred 3–20 weeks late (Table 2). In Kisii and Gucha, numerous agencies, including Merlin, African Medical and Relief Foundation (AMREF), Cooperative for Assistance and Research Everywhere, Inc. (CARE), World Vision, the Kenyan Army, and medical staff of Kenyatta National Hospital supported health structures. Elsewhere, MSF was the sole external agency. Everywhere, facilities were initially overwhelmed. In Kisii Hospital, bed occupancy reached 300% in the first 3 weeks. Hospitalization capacity was 0.45 beds per 10,000 people in Gutten and 0.53 beds per 10,000 in Damot Gale, where, in June and July 2003 (epidemic onset), the rate of outpatient consultations per person per year was 0.03–0.09 (0.5–1.0 is expected in such settings if access is good) (20). Although waived once epidemics were officially recognized, user fees were initially charged in Kenya, Burundi, and Ethiopia. In Aweil East, non-MSF clinics had run out of chloroquine 3 weeks into the outbreak.

All interventions included inpatient components with blood transfusion. Conversion of existing MSF nutritional structures enabled expansion of care in Aweil East and Damot Gale. To reach isolated communities, mobile clinics, consisting of teams of nurses or nursing assistants working with simple treatment algorithms, were established at each site. However, this intervention occurred late (10 weeks late in Kisii and Gucha, 7 in Kayanza, 8 in Aweil East, 13 in Gutten, and 27 in Damot Gale) and, apart from in Kayanza, after the epidemic peak (Table 2). Choice of location depended on results of a cross-sectional prevalence survey (Kisii and Gucha), distance from the nearest health center or proximity to swampy areas (Kayanza), known gathering point and greatest distance to the outpatient department (Aweil East), known gathering point near existing health posts (Gutten), and village morbidity/mortality surveillance results (Damot Gale). Diagnosis was presumptive everywhere except Ethiopia, where the P. falciparum–specific rapid diagnostic test (RDT) Paracheck (Orchid Biomedical Systems, Verna, Goa, India) was used systematically (Table 2). In Aweil East, mobile teams traveled on bicycle and canoe, spending 3–4 days in each location; because transporting patients with severe cases was impossible, more experienced teams carried injectable artemether and anticonvulsant drugs and treated 110 patients on a semi-inpatient basis (no outcome was recorded for these patients). At other sites, mobile clinics remained on site for 1 day and provided an ambulance service. Mobile teams were present in each targeted village for no more than 1–2 days a week on average and as little as once a month in Kenya and Damot Gale (Table 2). Mobile clinics treated 46,541 (9.3%) of 501,214 reported cases in Kayanza, 34,749 (68.3%) of 50,863 in Aweil East, 7,258 (19.4%) of 37,457 in Gutten, and 467 (2.8%) of 16,621 in Damot Gale (Table 3). In Damot Gale, active severe case finding was organized (no data available).

Artemisinin-based combination therapy (ACT) was deployed in Aweil East and in mobile clinics in Kenya (Table 2). Its use was not officially authorized in Burundi and in Ethiopia, where empiric evidence of poor sulfadoxine-pyrimethamine efficacy, later confirmed by in vivo studies (15), led clinicians to use quinine as first-line treatment. To ensure adherence to the 7-day regimen, high-risk patients were treated intrarectally under observation (Table 3).

In Burundi, Sudan, and Ethiopia, surveillance data were analyzed weekly. In Kayanza, RDT testing was carried out every 2–3 weeks among outpatients to monitor epidemic trends. In Aweil East and Gutten, an automated surveillance spreadsheet generated key indicators and graphs (caseload, proportionate morbidity and mortality, case-fatality, RDT confirmation of diagnosis).

The Kisii and Gucha epidemic followed a historical pattern of short dramatic peaks (Figure 2). Kisii Hospital records showed that, during the first 12 epidemic weeks, 2,669 (22.2%) of children <5 years of age in Kisii municipality (≈12,000) were hospitalized for malaria (Table 3).

The Kayanza epidemic lasted 36 weeks and roughly followed a normal distribution (Figure 3). A total of 501,214 cases were reported, for a minimum attack rate of 86.5%.

In Aweil East, a peak was reached by epidemic week 2, and a steady decline followed, which reflected percentage of confirmed malaria cases among women who came to the clinic for antenatal visits (Figure 4). Children <5 years of age (assumed to be 20% of the population) experienced attack rates of >41.2% (all malaria) and 1.1% (complicated).

In Ethiopia, the epidemic's evolution can partly be reconstructed by plotting available microscopy results from the Gutten government clinic, which yields a normal distribution (Figure 5), and percentage P. falciparum positivity among malnourished children admitted to feeding centers in Damot Gale (Figure 1). Results showed a steady rise from June, a plateau in August and September, and a new peak in late November after heavy rains.

Among uncomplicated cases, the proportion of patients <5 years of age exceeded the expected levels of 15% to 20% in southern Sudan and Ethiopia but not Kenya, where only presumptive diagnosis was used (Table 3). Patients <5 years made up half of all complicated cases in Kisii, Gucha, Gutten, and Damot Gale and almost 80% in Aweil East. Case-fatality rates were comparable across ages and sites except Gutten (footnote, Table 3). Convenience samples of outpatients or household surveys suggested high P. falciparum prevalences at or near caseload peaks (Table 3). Where MSF was involved in both outpatient and inpatient care (Kayanza, Aweil East, and Gutten), comparable proportions of patients were hospitalized (1.2%, 1.9%, and 1.4% respectively).

Our analysis did not include controls (i.e., sites where no epidemics occurred). Nevertheless, remotely sensed climate data suggest rainfall abnormalities during key preepidemic periods: relative drought in the 2 or 3 preepidemic years (with the exception of Kayanza) and above-average rainfall 1–2 months before epidemic onset. No consistent temperature pattern emerged.

The full role of such abnormalities as epidemic determinants is unclear. Furthermore, although remotely sensed environmental variables provide relatively robust and accurate estimates (21) and are becoming more publicly accessible (22), they remain only proxies for ground-based measurements. Nonetheless, we believe that our findings implicate climate abnormalities to a varying extent in all 5 epidemics and support strengthened monitoring of climate variables for early warning.

Land cover changes in Aweil East (flooding) and in Kayanza (rice paddy creation) probably favored vector breeding. Malnutrition, displacement, and drug resistance may not in themselves cause epidemics, but in our settings these factors probably exacerbated the epidemics' magnitude, duration, and case-fatality ratios. The effects of past drought and malnutrition are difficult to extricate: they are related causally, and either could result in impaired immunity (respectively, through reduced exposure to infection and nutrient deficiencies).

To our knowledge, no entomologic data were collected during any of these epidemics, which limits the strength of our findings; the role of changes in vector species or breeding habitats could have had a major role, but these factors can only be imputed from observed land pattern or climate alterations. Future studies on malaria epidemics should include detailed entomologic profiling, even during the epidemics.

In short, we believe that, given available evidence, to predicate epidemic prevention activities solely on the basis of individual risk factors (meteorologic or other) would be imprudent. Instead, appropriate decision support systems should be built that integrate all relevant data (e.g., environmental variables, food security and nutritional status, drug efficacy, health coverage, vector characteristics, population at risk) into a risk profile for each epidemic-prone population, to be updated regularly; in such a scenario, warning flags (23) resulting from detected environmental anomalies or other risk factors would result in enhanced surveillance activity and increased emergency preparedness (e.g., stockpiling drugs, ensuring insecticides and spray teams are in place), rather than leading directly to control activities.

By the time interventions were implemented, their potential effects were reduced. Mobile clinics were deployed to expand health access and detect severe cases. Implementation of clinics understandably varied according to local conditions, but apart from in Aweil East, probably had limited impact. Mobile clinic programs should be designed on the basis of clearly identified catchment areas and set frequencies with which communities should be visited. Although various criteria were used in our case studies, we believe that actual access to healthcare should be a key indicator for selecting target populations. Rapid methods to assess antimalarial treatment coverage thus need to be developed. How frequently communities are visited determines both the improvement in treatment coverage and the probability of preventing progression to severe disease through prompt treatment, which is likely to increase exponentially with frequency of mobile team visits; we hypothesize that frequent visits to selected sites may be more efficient than infrequent visits to a wider area. Impact monitoring should be included in future mobile clinic interventions to adjust their strategy as the situation evolves, and they should be evaluated after the fact. More generally, alternative modes of rapidly decentralizing care, such as fixed temporary health posts or training of resident community health workers (possibly equipped with artesunate suppositories to treat severely ill patients), merit further exploration. Where no clear indications exist that local health structures can cope with a large malaria epidemic, mobile clinics or other temporary treatment programs should be implemented immediately.

Case-fatality ratio among patients with complicated cases was lower than current best estimates of 10% (2) and 13% (26); however, whether all cases were severe depends on the case definition used. Treatment of uncomplicated cases relied on failing drugs everywhere but Sudan, and sulfadoxine-pyrimethamine monotherapy was probably counterproductive because the drug stimulates gametocytogenesis (27) and thus transmission. Ineffective drug use in Burundi probably limited the effect on mortality; in Ethiopia, quinine first-line administration proved challenging because of vomiting and required impractical patient monitoring.

Kayanza excluded, the increased proportion of children <5 years of age among inpatients, as previously observed in Kenya (28), suggests that children were more susceptible to symptomatic disease, which challenges classical notions of unstable, epidemic malaria. In Aweil East, the predominance of children is consistent with stable, mesoendemic to hyperendemic transmission, and this situation is probably better characterized as a severe seasonal outbreak.

Clinic-based attack rates approach 100% for all age groups when extrapolated to the entire epidemic period (Kayanza and Gutten) and are even more alarming among children <5 years of age in Kisii. Even after overdiagnosis from presumptive treatment is accounted for, these rates are likely to be gross underestimates. The vast gap in treatment coverage was evident in Aweil East, where large-scale deployment of mobile clinics greatly increased output, and in Ethiopia, where despite capturing only the declining phase of the epidemic, uninterrupted provision of free care with effective drugs resulted in far higher outpatient and inpatient department attendance. The true community incidence in these epidemics is probably much greater than represented by regular reporting systems and higher than current estimates of 0.5 episodes of malaria per person per epidemic (29). Only population-based studies can yield realistic estimates of this incidence.

We characterized climatic and ecologic conditions during each epidemic by using meteorologic indexes obtained by remote sensing (station data were incomplete or unavailable). Dekadal (10-day) rainfall estimates for July 1981 through December 2005 and normalized difference vegetation indexes for July 1995 through December 2005 were obtained at 8-km spatial resolution from the Africa Data Dissemination Service (http://igskmncnwb015.cr.usgs.gov/adds/index.php). Data for daytime and nighttime land-surface temperature (LST) were obtained through the Epidemio project (http://www.epidemio.info/index.php?section=homepage) for January 1995 through September 2005 at 2.5-minute spatial resolution. For each site, administrative boundaries were overlaid with remotely sensed datasets in a geographic information system with Arc/Info workstation software (version 8.1, ESRI, Redlands, CA, USA). We extracted summary statistics for each parameter and dekad (10-day period) and compared these with long-term averages for the same season. Our aim was to identify obvious climate anomalies either during or before the epidemics; therefore, we limited ourselves to descriptive analysis rather than establishing causality between climate and epidemic onset.

In Kisii, estimated rainfall from September 1998 through January 1999 was 59% below the expected (536 mm), according to the 10-year rainfall estimate dataset. The long rains of March and April 1999 (1 month before epidemic onset) were unusually heavy (704 mm compared with a long-term average of 401 mm). Land-surface temperature for 1998 and 1999 generally fell within the normal range from 1995 through 2005, except during the preepidemic period, November 1998 through February 1999, when average maximum land-surface temperature (33.1°C) was 14% higher than the corresponding long-term average.

In Kayanza, unusually heavy rains occurred between the second dekad of February and the first dekad of April 2000 (460 mm vs. a 1995–2005 average of 304 mm), and an additional peak occurred during the last dekad of May (137 mm vs. 22 mm), 3 months before epidemic onset. Data from neighboring Ngozi Province suggested a rise in temperature (1), but no land-surface temperature anomalies are evident for this period.

In East Wollega Region, including Gutten Locality, 2002 and 2003 show similar drought patterns from the last dekad of April to the end of May (61 mm and 33 mm, respectively, compared with an average of 192 mm from 1995 through 2005), followed by above-average rainfall throughout June (preepidemic month), especially in 2003. Annual rainfall estimate averages for 2001 through 2003 (962 mm) seem markedly drier than preceding years (1996–2000, 1,431 mm). Drought conditions in the 2003 preepidemic months are also captured by mean normalized difference vegetation index data (0.38 from the last dekad of April to end of June 2003 vs. 0.5 average for the same period [1995–2005]). No land-surface temperature anomalies were apparent.

In Damot Gale, annual rainfall estimate averages from 2001 through 2003 (940 mm) were lower than for 1996 to 2000 (average 1,204 mm). In the preepidemic months of April, May, and June, rainfall estimate was 268 mm in 2002 (preepidemic year), compared with 376 mm in 2003 (epidemic year) and the 1996–2004 mean of 416 mm. land-surface temperature for this period seemed within normal range.

In Aweil East, 2000–2002 was relatively dry, particularly in 2002 when annual rainfall estimate was 40% below the long-term average (963 mm). Rainfall totals for 2003, however, were ≈20% higher, with a particularly rainy season leading up to May (preepidemic month). Land-surface temperature data suggest that temperatures in the second half of 2002 and first part of 2003 were within normal range, but during the epidemic period itself (June–November), mean maximum land-surface temperature (34.1°C) was 3.5°C below the period average.

All sites but Kenya were affected by conflict and displacement. Burundi's and southern Sudan's conflicts had lasted 7 and 20 years, respectively; Aweil East's population had been repeatedly displaced by militia incursions. Both Gutten and Damot Gale had received immigrants resettled from drought-stricken, nonmalarious areas.

Parasite resistance to first-line treatments was high in Kayanza (chloroquine [CQ]), Sudan (CQ), and Ethiopia (sulfadoxine-pyrimethamine [SP] at the peripheral level and the combination CQ+SP in better-equipped centers). SP was prescribed as first-line in Kenya and Ethiopia (primaquine coadministration was mandated by Ethiopian guidelines, but this drug was not supplied to Gutten and Damot Gale during the epidemic).