Real-time, spatial decision support to optimize malaria vector control: The case of indoor residual spraying on Bioko Island, Equatorial Guinea

Public health interventions require evidence-based decision-making to maximize impact. Spatial decision support systems (SDSS) are designed to collect, store, process and analyze data to generate knowledge and inform decisions. This paper discusses how the use of a SDSS, the Campaign Information Management System (CIMS), to support malaria control operations on Bioko Island has impacted key process indicators of indoor residual spraying (IRS): coverage, operational efficiency and productivity. We used data from the last five annual IRS rounds (2017 to 2021) to estimate these indicators. IRS coverage was calculated as the percentage of houses sprayed per unit area, represented by 100x100 m map-sectors. Optimal coverage was defined as between 80% and 85%, and under and overspraying as coverage below 80% and above 85%, respectively. Operational efficiency was defined as the fraction of map-sectors that achieved optimal coverage. Daily productivity was expressed as the number of houses sprayed per sprayer per day (h/s/d). These indicators were compared across the five rounds. Overall IRS coverage (i.e. percent of total houses sprayed against the overall denominator by round) was highest in 2017 (80.2%), yet this round showed the largest proportion of oversprayed map-sectors (36.0%). Conversely, despite producing a lower overall coverage (77.5%), the 2021 round showed the highest operational efficiency (37.7%) and the lowest proportion of oversprayed map-sectors (18.7%). In 2021, higher operational efficiency was also accompanied by marginally higher productivity. Productivity ranged from 3.3 h/s/d in 2020 to 3.9 h/s/d in 2021 (median 3.6 h/s/d). Our findings showed that the novel approach to data collection and processing proposed by the CIMS has significantly improved the operational efficiency of IRS on Bioko. High spatial granularity during planning and deployment together with closer follow-up of field teams using real-time data supported more homogeneous delivery of optimal coverage while sustaining high productivity.

operational efficiency was also accompanied by marginally higher productivity. Productivity ranged from 3.3 h/s/d in 2020 to 3.9 h/s/d in 2021 (median 3.6 h/s/d). The novel approach to data collection and processing proposed by the CIMS has significantly improved the operational efficiency of IRS on Bioko. High spatial granularity during planning and deployment together with closer follow-up of field teams using real-time data supported more homogeneous delivery of optimal coverage while sustaining high productivity. This statement will be typeset if the manuscript is accepted for publication.
Please review the submission guidelines and the instructions link below for detailed requirements and guidance.
The author(s) received no specific funding for this work

Competing Interests
On behalf of all authors, disclose any competing interests that could be perceived to bias this work.
This statement will be typeset if the manuscript is accepted for publication.
Please review the instructions link below and PLOS Digital Health's competing interests policy to determine what information must be disclosed at submission. This statement will be typeset if the manuscript is accepted for publication.
PLOS allows rare exemptions to address legal and ethical concerns. If you have legal or ethical restrictions, please use the box below to detail these in full sentences for the Journal team to consider.
Public health interventions require evidence-based decision-making to maximize impact. Spatial decision support systems (SDSS) are designed to collect, store, process and analyze data to generate knowledge and inform decisions. This paper discusses the use of a SDSS, the Campaign Information Management System (CIMS), to support malaria control operations on Bioko Island and assess its impact on indoor residual spraying (IRS) coverage, operational efficiency and productivity. We used data from the last five annual IRS rounds (2017 to 2021). IRS coverage was calculated as the percentage of houses sprayed per unit area, represented by 100x100 m map-sectors. Optimal coverage was defined as between 80% and 85% and under and overspraying as coverage below 80% and above 85%, respectively. Operational efficiency was defined as the fraction of map-sectors that achieved optimal coverage. Daily productivity was expressed as the number of houses sprayed per sprayer per day (h/s/d). These indicators were compared across the five rounds. Overall IRS coverage (i.e. percent of total houses sprayed against the overall denominator by round) was highest in 2017 (80.2%), yet this round showed the largest proportion of oversprayed map-sectors (36.0%). Conversely, despite producing a lower overall coverage (77.5%), the 2021 round showed the highest operational efficiency (37.7%) and the lowest proportion of oversprayed map-sectors (18.7%). In 2021, higher operational efficiency was also accompanied by marginally higher productivity. Productivity ranged from 3.3 h/s/d in 2020 to 3.9 h/s/d in 2021 (median 3.6 h/s/d). The novel approach to data collection and processing proposed by the CIMS has significantly improved the operational efficiency of IRS on Bioko. High spatial granularity during planning and deployment together with closer follow-up of field teams using real-time data supported more homogeneous delivery of optimal coverage while sustaining high productivity. Public health precision requires efficient and effective targeting of interventions to those 2 most in need using the best available evidence [1][2][3]. Spatial decision support systems 3 (SDSS) represent critical tools to achieve this goal by transforming disease data into 4 information and knowledge useful for decision-making [3]. The spatial component is 5 essential to enable the prioritization of resources and efficient and equitable delivery of 6 interventions. This is particularly relevant for malaria vector control interventions that 7 aim to provide community-wide protection. 8 Indoor residual spraying (IRS) is a critical component of malaria control in many 9 endemic countries [4,5]. IRS delivery is a challenging endeavour that entails the 10 simultaneous deployment of many fieldworkers. Often, malaria programmes plan and 11 deploy IRS based on a target daily productivity per sprayer [6,7]. This demands close 12 and strategic management and monitoring of activities. The ultimate goal of IRS is to 13 achieve universal coverage that assures community protection [8][9][10].
14 Canonically, the recommended threshold for universal IRS coverage has been loosely 15 defined as between 80% and 85% of houses sprayed within a given targeted area [7,11]. 16 Though the evidence supporting this recommendation is limited [12][13][14], IRS is an 17 expensive intervention [15][16][17], and there is a need to balance the community effects of 18 IRS while maximizing coverage equity. Striking the right balance could maximize the 19 overall impact of scarce resources. Operationally, this would require optimizing the use 20 of commodities and labor by maximizing productivity towards reaching optimal 21 coverage based on known denominators. 22 We use optimal operational coverage (hereafter referred to as optimal coverage) as a 23 key concept for these analyses and articulate it with a simple thought experiment (see  Heterogeneity in coverage can potentially leave gaps of unprotected populations despite 27 seemingly achieving overall adequate coverage, a phenomenon explained by the scale 28 effect and subject to concerns surrounding the modifiable areal unit problem 29 (MAUP) [18,19]. We also introduce the term operational efficiency to express the 30 frequency at which optimal coverage is achieved. It follows that maximizing operational 31 efficiency improves the even delivery of IRS, ideally maximizing its community effect. 32 This is a non-trivial undertaking that could be operationalized through the use of a 33 SDSS. 34 SDSS have long been recognized as a necessity for malaria control and elimination. Achieving optimal spray coverage through planning at higher spatial granularity.
A. An administrative division of Bioko Island, its four subdivisions and the distribution of houses within (n=2,341); overlaid are the map-sectors (n=203) and a smoothed prevalence surface. B-E. Different hypothetical scenarios of IRS, all achieving optimal coverage at the administrative division-level. In scenarios 1 to 3, 1,873 houses are sprayed to achieve exactly 80% coverage in different configurations. In scenario 1, all the houses in three subdivisions and 16.8% of houses in the fourth subdivision are sprayed. In scenario 2, 90% of houses in three subdivisions and 48.5% in the fourth are sprayed. In scenario 3, 80% of randomly selected houses across the division are sprayed. In scenario 4, spraying is deployed based on optimal coverage calculated at map-sector-level, with 1,946 houses sprayed and an overall 83.1% coverage. The bar graphs depict the proportion of map-sectors by coverage and the number of houses over and undersprayed in each scenario. The latter refer to the number of houses that were unnecessarily sprayed and those which should have been sprayed in order to reach optimal coverage, respectively.
During the Global Malaria Eradication Programme, geographical reconnaissance was 36 advocated as essential for the attack and consolidation phases to assure that 37 interventions reached every household [20]. At the time, digital tools were both incipient 38 and not readily available to complement paper maps with data. Early efforts to improve 39 intervention management using information technology were documented for southern 40 Africa [21]. This initiative used relational databases and geographic information systems 41 to replace paper-based reports, facilitating regular monitoring of spray coverage, worker 42 performance and insecticide use. Data systems enabled the resolution of common 43 operational problems during implementation. The further advancement of information 44 technology allowed SDSS to gain considerable attention in current malaria control and 45 elimination programmes, particularly regarding the delivery of vector 46 interventions [22][23][24][25][26][27][28]. Real-time data support is also increasingly acknowledged as 47 hugely beneficial for malaria monitoring and surveillance [29,30]. 48 This paper describes the use of a novel SDSS to support malaria vector control on 49 Bioko Island, with IRS as an illustrative example. The SDSS, the Campaign Information 50 Management System (CIMS), is fully described in the Supplementary Information (S1 51 File). The CIMS has been progressively developed and used to support malaria control 52 on Bioko. The grid-based mapping system underpinning the CIMS is described in detail 53 elsewhere [31] and represents the crux behind intervention deployment planning. We use 54 data from the last five annual IRS rounds on Bioko Island (2017 to 2021) generated 55 through the CIMS to track coverage and productivity. Over this period, the CIMS   Overspraying of map-sectors progressively decreased in the five-year period, from a 101 high of 36.0% in 2017 to a low of 18.7% in 2021. There was a significant reduction in 102 map-sector overspraying in 2018 (29.5%, P = 0.011) and then another in 2020 (19.9%, 103 P < 0.001). The latter could be explained by the lower overall productivity and 104 coverage in that year due to COVID-19 [33], but this drop was sustained in 2021 with a 105 further, though not significant (P = 0.268), 1.2% reduction. Cumulative density functions of coverage, by round. The grey band marks the optimal coverage range between the vertical lines at 80% and 85%. A vertical line at the maximum coverage of 100% is also drawn to highlight the level of overspraying. C. Probability density functions of relative coverage, by round. Relative coverage is calculated by the ratio of actual houses sprayed to houses needed to spray to achieve no less than 80% and no more than 85% coverage, where 1 is equivalent to optimal coverage (see main text).

113
Median productivity across all rounds was 3.6 h/s/d, ranging from 3.3 in 2020 to 3.9 114 h/s/d in 2021 (Table 1). No significant differences were observed between rounds, 115 except for 2020, when the drop in productivity was statistically significant (P < 0.001) 116 followed by a significant increase in 2021 (P < 0.001). There was no statistically 117 significant difference between productivity in 2019 and in 2021 (P <= 0.608). Fig 4B   118 September 28, 2021 5/15 illustrates the distribution of daily productivity by round. In 2017 and 2018, 119 productivity was over-dispersed, but this distribution narrowed progressively around the 120 target of 4 h/s/d in the most recent rounds. Fig 4 serves  Managing overspraying is more reliant on planning and implementation. As such, it 147 is imperative that spray teams have access to data that signal when coverage targets 148 have been reached, prompting exit from the map-sector. Evidently, this protocol has 149 improved over the last years as is reflected by the significant decrease in overspraying. 150 It is not uncommon, however, to find situations when residents request fieldworkers to 151 spray their houses even when the target has been met. This could partly explain why, 152 despite the real-time support of the CIMS, some map-sectors were oversprayed.

153
With regards to productivity, the data showed that the median was close to the 154 target of 4 h/s/d in the five rounds examined, with marginally higher productivity 155 observed in 2021 (Fig 4B). Certainly, this was possible thanks to the close monitoring of 156 field activities using the CIMS for constantly assessing map-sector coverage and daily 157 productivity. The use of real-time monitoring through online dashboards proved highly 158 beneficial in 2021 by allowing field managers to determine where adjustments and 159 corrective actions to boost productivity were needed. The system was also critical for 160 better supervision of spraying performance to identify productivity outliers. This is 161 reflected by the more constant worker output measured in 2021 compared to earlier 162 rounds. In 2017 and 2018, productivity was over dispersed due to some fieldworkers 163 reporting very high daily outputs without proper confirmation of these reports (Fig 4B). 164 The distribution of productivity in 2021 showed a more even spread around the target. 165 Factors affecting worker productivity are the subject of ongoing research.

166
A requisite for an effective SDSS is the human resource capacity to enter, use, 167 process and interpret data [22,23,35] training is a gradual process that takes time to produce the necessary human capacity 175 to improve IRS deployment. This steep learning curve was another reason why, despite 176 the progress, coverage and productivity indicators in 2021 still showed substantial room 177 for improvement. Guiding a team of over 100 sprayers across thousands of map-sectors 178 to spray thousands of houses in mostly urban areas is a far from negligible task that 179 unavoidably falters. The ultimate goal is a greater operational efficiency, no 180 underspraying and productivity higher than the minimum threshold of 4 h/s/d (i.e. a 181 picture similar to Scenario 4 in Fig 1 and Box #1). Notwithstanding this challenge, the 182 engagement of fieldworkers returned positive outcomes and the trends revealed by the 183 data promise that future rounds will see further advancements.

184
The grid-based coding system at the core of the CIMS promotes a highly spatially 185 granular approach to data collection, processing, analysis and feedback. The fine 186 balance between achieving optimal coverage and using limited resources cost-effectively 187 can be easily altered by the spatial resolution at which coverage is measured. The 188 grid-based approach helps overcome the effects of the MAUP and resource allocation 189 inefficiencies by increasing the spatial granularity of intervention planning and coverage 190 derivation [19]. As is illustrated in Fig 1, the grid-based approach is pragmatic, protection. The 80% to 85% optimal recommendation [7] is likely inherited from canons 200 of early efforts of malaria eradication [36] that are supported by limited evidence [13]. 201 Optimal IRS coverage could well be lower, as has been suggested by data from 202 Malawi [14] and similar to the more widely investigated community effects of bed 203 nets [37][38][39][40][41]. More importantly, a spillover community effect is expected around areas 204 with high intervention coverage [42] and this would also influence IRS planning and 205 deployment. For bed nets, community protective effects have been observed within a 206 300 m radius from intervened areas [43,44]. However, the evidence of the size and  Understanding spatial area effects could help give rise to IRS coverage patterns that 216 would maximize impact. Hypothetically, if spillover effects of IRS were around 200 m at 217 80% coverage, then planning deployment at high spatial granularity would allow us to 218 guide spraying along a reticular pattern of map-sectors within which 80% of houses were 219 sprayed and which would be separated from the nearest sprayed map-sectors by four 220 September 28, 2021 7/15 unsprayed map-sectors (Fig 5). This would improve protection to all the population, 221 regardless of whether their neighborhood was sprayed or not. Such a setting would save 222 considerable resources as wider areas would be sprayed using the same workforce and 223 amount of insecticide. This is only a simplified example of a plausible scenario.

224
Whatever the spatial configuration required, the high spatial granularity of the CIMS 225 provides the necessary flexibility for the implementation of field activities.
226 Fig 5. Theoretical example of IRS deployment with and without allowing for spillover.
A. All map-sectors within a target population are sprayed at optimal coverage. B.
Map-sectors are strategically targeted, taking into account a 200 m spillover effect, or the distance comprised by two map-sectors. The spillover effect plausibly wanes with growing distance from high IRS coverage, but for illustrative purposes it is assumed that everyone inhabiting the purple map-sectors is equally protected by the intervention. In A, 696 map-sectors are sprayed whereas in B only 259 are sprayed.
High spatial granularity is also essential for optimally targeting interventions, which 227 is particularly relevant in areas where disease risk is highly heterogeneous [45]. The use 228 of map-sectors (100m across) for operational planning and implementation guarantees 229 more precise targeting, tracking and monitoring of interventions (see Box). In addition, 230 the increased spatial granularity would provide a more flexible framework for temporal 231 intervention targeting and better scheduling of delivery [46]. This flexibility renders the 232 grid-based approach to data collection a powerful asset of the CIMS. budgeting to planning deployment to defining sampling frameworks for surveys, and 241 more. The ability to track individual-level performance has been used for quality 242 assurance and quality control of interventions [6], algorithms that integrate 243 entomological, parasitological and case data are constantly improved to facilitate 244 surveillance and response to malaria outbreaks [47], and outreach training and 245 supportive supervision for malaria case management in public health facilities through 246 the CIMS is being established [48]. Importantly, the CIMS offers excellent versatility to 247 adapt to public health interventions beyond malaria vector control and in settings 248 outside of Bioko Island.  included only map-sectors with denominators equal or higher than a convenient cutoff 275 of 10 houses. 276 We determined the number of houses needed to spray in each map-sector in order to 277 achieve optimal coverage. This was calculated as the ceiling of denominator * .8 and as 278 the floor of denominator * .85, with the rule that the latter would have to be equal to 279 or higher than the former. For example, for a map-sector with 25 inhabited houses, the 280 number of houses to spray would be between 20 and 21 (i.e. ceiling (25 * .8) and 281 f loor(25 * .85)). We then calculated relative coverage at each map-sector, which 282 corresponded to the ratio of houses sprayed to houses needed to spray to obtain optimal 283 coverage according to the above calculation. In the example above, relative coverage 284 was 1 if the houses sprayed in that map-sector were 20 or 21. Map-sectors were 285 classified as optimally, over or undersprayed if their relative coverage was equal to, 286 above or below 1, respectively. 287 We calculated underspraying and overspraying at two different scales. First, the 288 number of houses sprayed below and over the optimal coverage band within each 289 map-sector was aggregated for all targeted map-sectors to provide the overall number of 290 houses that were under and oversprayed per round. Second, the number of map-sectors 291 at, below and above optimal coverage represented the number of optimally sprayed, 292 undersprayed and oversprayed map-sectors. Both the number of houses and map-sectors 293 thus classified were expressed as the proportions of the total houses and map-sectors 294 sprayed, respectively, and were compared between rounds using tests of proportions in 295 R [32]. 296 Calculating productivity 297 Productivity was measured as the number of houses sprayed per spray operator per day 298 (h/s/d). According to the World Health Organization, productivity is expected to be as 299 high as 10 to 15 h/s/d in locations where houses are easily accessible and relatively 300 small, and as low as 5 h/s/d where houses are scattered or large [7]. In the landscape of 301 Bioko Island, IRS operational challenges differ by setting, which includes approximately 302 80% urban areas and 20% peri-urban and rural areas. Households in urban areas of 303 Bioko present high refusal rates, low availability rates and often reduced co-operation by 304 residents to prepare their house for spraying. This inevitably reduces the average daily 305 productivity attainable by sprayers. Based on experience across the many annual 306 rounds of IRS, minimum target productivity for planning purposes on Bioko has been 307 defined at 4 h/s/d. We reported productivity using the median and inter-quartile range 308 September 28, 2021 9/15 of the daily productivity by round and compared productivity between rounds using the 309 Wilcoxon-Mann-Whitney non-parametric test [32].

Box #1
Optimizing IRS coverage We define a band for optimal coverage as no less than 80% and no more than 85% of houses sprayed in a given map-sector, pragmatically justified on the basis of established canons rather than on existing evidence [7]. It is plausible that optimal upper and lower bounds for coverage may well be set different and likely context-specific, depending on heterogeneous transmission and programmatic goals. We use a band rather than a single cutoff because often it is practically impossible to spray an exact proportion of the denominator. Any coverage below and above this band represents under and overspraying. These concepts are motivated by operational rather than epidemiological principles, as the goal is to balance community protection against use of resources. The assumption is that resources are limited and commodities are procured to maximize cost-effectiveness while assuring community-wide protection of the entire population targeted for the intervention. Figure 1 illustrates these concepts using four hypothetical scenarios within an administrative division of Bioko Island. The details of each scenario are explained in the caption of Figure  1. In this example, we assume that sufficient insecticide and human resources are secured for spraying between 80% and 85% of houses within the administrative division, which was selected to receive IRS based on predetermined criteria.
Scenario 1 is the worst-case because it leaves a large proportion of map-sectors undersprayed (40.4%), failing to achieve the community protection objective while also overspraying 58.6% map-sectors and thus failing the cost-effectiveness objective. Scenario 2 has a lower level of over and underspraying (54.7% and 39.4% map-sectors over and undersprayed, respectively; Figure 1C). These scenarios could be expected when spray teams are guided through a deployment designed for convenience and logistical ease at the expense of coverage and resources. One explanation could be that houses in the East of the administrative division are more accessible than those in the West. In scenario 3, under and oversprayed map-sectors are interspersed throughout the administrative division. Despite this, 38.8% and 48.8% of map-sectors are under and oversprayed, respectively, though fewer houses within these map-sectors were sprayed above the required number to achieve optimal coverage ( Figure  1D).
In all three scenarios, a low proportion of map-sectors are optimally covered (1% in scenario 1, 5.9% in scenario 2 and 12.3% in scenario 3). By way of contrast, in scenario 4, 42.4% of map-sectors are adequately covered and no map-sectors have coverage under 80%. Even though 57.6% of map-sectors in scenario 4 are oversprayed, this is explained by the small number of houses within them. This translated into only 38 houses sprayed over the number required to achieve optimal coverage in these map-sectors. All of this is achieved with an overall IRS coverage of 83.1%, or 73 more houses sprayed than in the other scenarios ( Figure  1E).

Box #1, (continued)
An important consideration is that malaria transmission is highly heterogeneous [45,49,50]. Figure 1A illustrates this heterogeneity, where the highest malaria prevalence appears localized in 63 map-sectors, mainly along the West and the South (pink map-sectors in Figure 1A where Pf PR > 20%). It follows that blanket intervention deployment, such as scenarios 1, 2 and 3, misses protecting populations at the highest risk and, depending on how IRS affects mosquito ecology, could fail to maximize the community benefits in surrounding areas. This is particularly the case in scenarios 1 and 2, in which 79.4% and 68.3% of high Pf PR map-sectors have sub-optimal IRS coverage. In scenario 3, although a more even coverage is achieved, 39.7% of the high prevalence map-sectors are undersprayed. In scenario 4, universal coverage is achieved, protecting the entire population at risk of malaria, including all those living within the highest malaria risk map-sectors.
These theoretical scenarios are not accounting for the spill-over effects of interventions such as IRS, whereby populations inhabiting map-sectors adjacent to those sprayed are also protected. They also assume that optimal spraying is defined by the 80-85% band, when in reality these bounds may be lower or higher, depending on the setting. Different assumptions would change the way these results are interpreted or even how the intervention is deployed in the first place. Notwithstanding this caveat, the scenarios serve as a stark reminder that intervention deployment can be severely biased by the geographic scale at which coverage is calculated. This bias is due to the scale effect of the MAUP [18,19].