Introduction
Floods are the most frequently occurring form of natural disaster [
1], affecting an estimated 2.8 billion people between 1980 and 2013 [
2]. Floods are predicted to become more frequent in smaller catchments globally given more intense rainfall events. Population exposure to flooding is predicted to increase, particularly in Asia, Africa, Central and South America as intense tropical cyclones become more frequent [
3]. With 37,600 dams higher than 15 m worldwide in 2014 and a further 3,700 planned or under construction [
4], fluvial flood regimes (where rivers overflow their banks) are increasingly mediated by dams. Whilst some dams are constructed explicitly to mitigate flood risk [
5], the effects of the majority are more complex and may profoundly affect the livelihoods of downstream populations [
6]. Given this context, it is important to understand their impacts on public health, so as to inform flood preparedness and mitigation efforts, particularly downstream of dams.
Flooding has often been found to increase diarrhoeal disease risk alongside that for other waterborne diseases, but there is less evidence on its impact on healthcare utilisation. In a systematic review of flooding and diarrhoea disease risk, 19 out of 25 quantitative analyses of the relationship reported a significant positive association, with plausible dose–response relationships observed in several studies [
7]. Positive relationships have been identified in studies in low and middle income countries for cholera, rotavirus, cryptosporidiosis, but also diarrhoea not attributable to a specific pathogen [
8]. Many water-borne disease outbreaks associated with extreme weather events such as flooding result in deaths [
9]. Flooding can mobilise pathogens in soils, animal or human faeces, and sediments, with pathogens associated with resuspended sediments. For example, increased concentrations of enteric viruses have been observed in surface waters during extreme flood events [
10]. Flooding can also contaminate groundwaters, both directly and as the subsurface becomes saturated, facilitating pathogen transport [
7]. Alongside greater potential for food contamination during floods, sanitation and water infra-structure may also be compromised by flooding, with backflows contaminating water systems. Through disruption to travel, flooding may also affect health facility utilisation. However, despite several studies of diarrhoea risk from flooding relying on outpatient records [
11,
12], no studies of flooding’s impact on attendance for routine or preventative healthcare were identified in a recent systematic review [
13]. Subsequently, a Cambodian study found that flooding had no impact on attendance for childbirth and a moderate impact on outpatient attendance in some districts only [
14].
For ecological studies seeking to quantify health risks from flooding, a key issue is how to assess population flood exposure status. One approach is to use time series of flood imagery from satellite remote sensing for exposure assessment. Several such data sets are routinely produced, including daily near-real time global flood mapping from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite sensor [
15]. Whilst there is general agreement between these different products for large flood events, detailed spatio-temporal patterns vary in data-sparse regions such as Africa [
16]. The impact of MODIS-derived flooding on health outcomes and healthcare utilisation has previously been assessed in Cambodia [
14], but it is unclear whether the observed strength of relationships could be affected by the choice of flood data product. Although flood exposure misclassification has attracted little attention in related systematic reviews [
13], it has consequent potential implications for quantifying health risks associated with flooding via epidemiological studies.
A systematic mapping of flooding’s health impacts [
17] also identified a paucity of mixed methods studies, recommending that such studies be used to deepen understanding of impacts. Although there have subsequently been mixed methods studies of flooding’s impacts on integrated community case management in Bangladesh [
18] and its long-term wellbeing impacts in the UK [
19], such studies remain scarce.
In northern Ghana, rainfall is associated with diarrhoea in sludge applying communities [
20], implying that floods could have a greater impact. Furthermore, extreme weather events limit client’s ability to reach health facilities in urban northern Ghana [
21] but the perception of service providers and their capacity to render services in floods is less understood. There is a lack of studies linking dam-mediated flood events and routine health data with contextual explanations from the communities impacted. Thus, observing flood events from satellite imagery in the study area [
22], and the availability of routine health data provides the opportunity for a mixed methods study.
The aim of this study is therefore firstly to assess the relationship between dam-mediated flooding as detected via the MODIS versus Landsat satellite sensors and reported outpatient attendance at health facilities. Secondly, we aim to assess the effect of dam-mediated flooding as detected via the two sensors on monthly acute diarrhoea cases reported by health facilities. Thirdly, via qualitative fieldwork, we aim to explain flooding’s impact on healthcare provision and utilisation. In doing so, we aim to develop a methodology that integrates remotely sensed data concerning floods, routinely collected outpatient data, and related geospatial data on attending populations and travel.
Discussion
To our knowledge, our study is the first to explore how choice of satellite-based flood product affects analyses of flood impacts on healthcare utilisation and diarrhoea. For both the NRT MODIS and Landsat-derived flood products, we find a significant reduction in reported outpatient attendance during flooding (Tables
2 and
3). Qualitative interviews with health professionals in both districts suggest that this decline in outpatient attendance may partly have been associated with staff travel disruptions during flooding, limiting the number of outpatients seen during flood periods. Although flood-related disruption to patient travel to facilities has been reported via key informant interviews in urban Ghana [
21], disrupted journeys by healthcare staff to rural facilities has not. Both KIIs and FGDs suggested some healthcare staff lived in towns distant from rural health facilities to avoid flooding at these rural sites. As a consequence, both patients and service providers were unable to access health facilities during flooding.
After controlling for seasonality, monthly reported diarrhoea cases significantly increased in relation to the proportion of flood-affected area per catchment (Table
5), suggesting a positive exposure–response relationship. This effect was however somewhat mediated by the flood-affected population within each catchment. Consistent with this finding, the majority of included studies in two systematic reviews of diarrhoea risk from flooding found a positive relationship [
7,
13].Given the known risk of food and water contamination [
42], alongside pathogen transport, population displacement inhibiting safe water access, and damage to water and sanitation infrastructure during flooding [
7], the observed increase in reported diarrhoea cases during flood months is epidemiologically plausible.
Our study has implications for using satellite-derived flood products for health risk assessment. To date, few studies have used satellite remote sensing to assess flood exposure for health risk assessment. Studies of diarrhoea risk included in a systematic review [
7] mostly used other methods to assess exposure, such as river levels [
43]. Some subsequent studies of health impacts have used geospatial disaster databases [
44] and NRT MODIS satellite imagery [
14] to assess flood status. Satellite-derived flood maps constitute an objectively defined measure of flood exposure that is internationally consistent and available in data-sparse regions, which could thus address a known lack of comparable exposure assessment metrics in many studies of flooding’s health impacts [
9]. However, our study suggests that the performance of models predicting healthcare utilisation and health outcomes is sensitive to the choice of satellite-derived flood product, since we find better performing models of both healthcare utilisation and reported diarrhoea based on a LandSat product compared to NRT MODIS (Tables
2 and
4). This likely reflects underlying differences in the two flood products, noted in other studies [
16]. The finer spatial resolution of LandSat (30 × 30 m) relative to MODIS (300 × 300 m) could reduce mixed pixel problems (classification difficulties arising from different land surfaces being present in the same grid square), thereby enabling detection of smaller flooding patches. However, whilst MODIS has a two day revisit time (the time elapsed between repeat observations of the same area), Landsat 8 has a 16 day revisit time [
16], so LandSat imagery may not capture peak flood extent and so under-estimate exposure. Both products are also subject to common limitations. Neither product measures floodwater depth, which is important for assessing impassability of roads [
45] and thereby patient travel disruption. Both optical sensors are affected by cloud cover [
46], which may lead to under-estimation of flood exposure. Since recent studies [
47] have integrated LandSat with synthetic aperture radar imagery from the Sentinel-1 sensor to map flooding despite cloud cover, we suggest that this approach would be appropriate in future studies of flooding’s public health impacts in data-sparse regions.
Given increased dam construction worldwide [
4], Our study also provides evidence of the healthcare utilisation impacts of dam-mediated flooding on downstream populations [
6]. However, whilst flooding in northern Ghana has increased despite declining rainfall since 1980, since the dam releases are one of several interacting drivers of flooding such as land cover change [
25], it is difficult to attribute impacts to the Bagre Dam release schedule. Nonetheless, a hydrological modelling study [
48] suggested that dam release could exacerbate flooding, raising White Volta water levels by 75 cm at 100-150 km downstream of relative to levels without dam operation. Management of dam water release schedules involves trade-offs between different impacts of such operations. Typically, trade-offs between power generation and water availability for irrigation are modelled in planning dam release schedules [
49], but the impacts on health outcomes and healthcare utilisation have not been quantified when assessing trade-offs. In principle, health or healthcare utilisation impacts could be incorporated into trade-off modelling for dam release schedule management, but our study suggests that establishing health risk attributable to dam release would be complex and highly uncertain, limiting usefulness of such an extended modelling framework.
Aside from the issues affecting flood exposure assessment via satellite remote sensing outlined above, our study is subject to several limitations. The coarse monthly temporal resolution of DHIS2 data could have limited our ability to detect flooding’s effect on outpatient attendance and diarrhoeal disease. Weekly health facility reports could have enabled shorter lagged effects to be detected, but weekly reports are seldom completed for routine outpatient attendance since notifiable disease reporting takes priority. Similarly, our study findings may be affected by the accuracy and completeness of DHIS2 reporting, although DHIS2 reporting completeness is high in Ghana [
28]. Whilst total outpatient attendance counts have been previously used to assess flooding impacts on healthcare utilisation, the risk of some diseases included within these counts could increase during flooding. For example, alongside diarrhoea risk, injury risk could increase during flood events [
13], resulting in potential under-estimation of reduced outpatient attendance from flooding. Previous studies have used health management information systems data for conditions unrelated to flooding alongside outpatient counts, most notably childbirth deliveries at facilities [
14]. However, healthcare-seeking behaviour for childbirth may differ from that for other health conditions. In relating population exposure to flooding with facility-level outpatient data, we did not model by-passing of health facilities, which is widespread for mothers’ journeys to give birth elsewhere in Ghana [
50]. Given its ecological design, our study did not consider household or individual-level characteristics such as water source type, sanitation access, and socio-economic status [
8,
51] that could moderate diarrhoeal disease risk from flooding.
In principle, the methods and workflow in this study can be generalised and replicated in settings where health management information system data are available. It could also be applied to other health conditions affected by flooding and recorded via DHIS2, such as injuries, skin or respiratory infections, and malaria [
13]. Since the District Health Information System (DHIS2) platform deployed in Ghana as DHIMS2 is used in 60 countries at national level and 14 at pilot stage or sub-nationally [
52], potentially these countries could replicate our study’s workflow, since the flood and population products we used are also global. However, despite the widespread international coverage of DHIS2 and increasing completeness in many countries such as Ghana [
28], data quality remains problematic, varying between countries [
53] and between system components, with for example lower accuracy for acute respiratory infection data compared to antenatal data in Malawi [
54].
Future research
Future work could further explore the uncertainties affecting the two flood exposure metrics we used and examine the relationship between alternative flood exposure measures and healthcare utilisation or health outcomes. Firstly, we examined residential population exposure to flooding using the WorldPop gridded population map layer. However, the WorldPop surface is just one of several available modelled gridded population layers. Since the WorldPop and LandScan surfaces distribute more population onto floodplains, they generate a much higher estimate of flood-affected population than a third gridded population layer (HRSL), which assumes populations avoid floodplains [
55]. Future studies should therefore assess residential population flood exposure using multiple gridded population surfaces. Secondly, we assessed inundation of healthcare facility sites since flood damage to facilities disrupts healthcare delivery [
21,
56]. However, flooding also disrupts electricity supply with outages sometimes affecting sites beyond the inundated area [
56]. There would thus be potential to use new night-time satellite remote sensing products that enable detection of power outages to represent flood impacts on healthcare delivery, such as NASA’s Black Marble Night-time Light product [
57].
Our FGDs and KIIs suggest further pathways by which flooding affects health and healthcare utilisation, which we did not measure in our study. These pathways entail disruption to referrals, patient travel to facilities, and staff travel to facilities. All three could be represented via satellite-derived flood products and explored via future studies. A Mozambican study recently incorporated flood extent from NRT MODIS into impedance surfaces, enabling modelling of daily variation in travel time to nearest healthcare facility [
58]. It would be possible to use this approach to model patient travel disruption from flooding and thereby assess its impact on healthcare utilisation. Since key informants in our study noted the impact of flooding on non-resident healthcare staff’s ability to travel to facility workplaces or to deliver community-based services such as immunisation or post-natal home visits, these service delivery outcomes should also be considered in relation to healthcare staff travel disruption in future work. Healthcare staff at primary facilities have also reported that flooding disrupts their ability to refer patients to secondary care because of travel difficulties [
56], so there would be potential for future studies to assess disruption to patient referrals resulting from flooding.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.