Optimising passive surveillance of a neglected tropical disease in the era of elimination: A modelling study

Background Surveillance is an essential component of global programs to eliminate infectious diseases and avert epidemics of (re-)emerging diseases. As the numbers of cases decline, costs of treatment and control diminish but those for surveillance remain high even after the ‘last’ case. Reducing surveillance may risk missing persistent or (re-)emerging foci of disease. Here, we use a simulation-based approach to determine the minimal number of passive surveillance sites required to ensure maximum coverage of a population at-risk (PAR) of an infectious disease. Methodology and principal findings For this study, we use Gambian human African trypanosomiasis (g-HAT) in north-western Uganda, a neglected tropical disease (NTD) which has been reduced to historically low levels (<1000 cases/year globally), as an example. To quantify travel time to diagnostic facilities, a proxy for surveillance coverage, we produced a high spatial-resolution resistance surface and performed cost-distance analyses. We simulated travel time for the PAR with different numbers (1–170) and locations (170,000 total placement combinations) of diagnostic facilities, quantifying the percentage of the PAR within 1h and 5h travel of the facilities, as per in-country targets. Our simulations indicate that a 70% reduction (51/170) in diagnostic centres still exceeded minimal targets of coverage even for remote populations, with >95% of a total PAR of ~3million individuals living ≤1h from a diagnostic centre, and we demonstrate an approach to best place these facilities, informing a minimal impact scale back. Conclusions Our results highlight that surveillance of g-HAT in north-western Uganda can be scaled back without substantially reducing coverage of the PAR. The methodology described can contribute to cost-effective and equable strategies for the surveillance of NTDs and other infectious diseases approaching elimination or (re-)emergence.


Introduction
Quantifying the spatial and temporal distribution of cases is essential for the development, implementation and monitoring of programmes to control or eliminate infectious diseases [1][2][3]. In general, costs for treatment and control of a disease scale with the number of cases, but those for surveillance can be high even when disease incidence is low [4]. Programmes aiming to eliminate or eradicate a disease maintain high levels of surveillance, as illustrated by the global programmes against smallpox and polio [5,6]. Similarly, surveillance systems to detect (re-)emerging diseases may also have relatively large surveillance costs in relation to the numbers of cases [7]. Maintaining adequate surveillance for diseases approaching elimination or with potential for emergence can be prohibitively costly for national health systems, particularly for resource-poor countries in the tropics [8,9], and as such, approaches for cost-effective surveillance are required.
Neglected tropical diseases (NTDs) cause high levels of mortality and morbidity in some of the world's poorest countries [10,11]. A global program aims to eliminate or eradicate eleven NTDs by 2030 and as the elimination of several NTDs approaches, globally and/or nationally, there is a pressing need for cost-effective strategies to detect the last remaining cases [12,13].
The new surveillance strategies will need to be adapted to the disease itself, local health systems and the population at risk (PAR) [1].
Disease surveillance systems include active, passive, sentinel, and statistical approaches, but the majority of reportable disease surveillance is conducted through passive surveillance [14,15]. Passive surveillance is based primarily on compilation of case reports submitted by healthcare facilities, produced when infected individuals present to the facility for diagnosis and treatment. One approach to improving cost-effectiveness of passive surveillance is to quantify the minimal number of surveillance sites that provides maximal coverage of the PAR. Towards this aim, we combined estimates of travel time to passive surveillance sites with a simulation-based approach to determine the optimal number of sites required to ensure that 50% and 95% of a PAR of Gambian human African trypanosomiasis (g-HAT) are within 1-hour and 5-hours of a diagnostic facility, respectively. The aforementioned criteria come from incountry targets [16], however, the approach described here can be adapted for a range of NTDs in an elimination setting, with varying coverage and travel time requirements determined by national governments, NGOs or other disease surveillance operators.

Scale back in the context of human African trypanosomiasis
Human African trypanosomiasis (HAT, also called 'sleeping sickness') is an NTD occurring across sub-Saharan Africa, caused by sub-species of Trypanosoma brucei transmitted by tsetse flies. Most (>95%) cases of HAT are caused by T. b. gambiense (g-HAT), for which humans are the primary host; the remaining~5% of cases are caused by T. b. rhodesiense (r-HAT) for which the primary reservoirs are animal hosts [17]. No prophylaxis or vaccine exists for either form of HAT. Therapeutic drugs are available [18], but the toxic nature of chemotherapy requires that infection status is confirmed before the patient is treated. Disease prevention and control efforts focus on the detection and treatment of existing cases and control of the tsetse vector [19].
The World Health Organization (WHO) aims to eliminate g-HAT as a public health problem by 2030. The definitions of elimination include fewer than 2,000 cases reported per annum globally, a 90% reduction in areas reporting >1 case in 10,000 per annum compared to 2000-2004, with a country-level indicator of elimination defined as fewer than 1 reported case per 10,000 people, per annum (averaged over a 5-year period) in each health district, in conjunction with adequate, functional control and surveillance [20]. Uganda is on track to achieve this goal with a reduction from 2,757 (range, 310-948 cases/year) cases reported between 2000-2004 to 18 (range, 0-9 cases/year) cases in 2014-18 [21]. These reductions are mostly attributable to active screening and treatment of the population [22,23], and, more recently, vector control [24]. For g-HAT, active screening has been the cornerstone of surveillance in most foci during periods of high case numbers. As case numbers decline this has become increasingly less cost-effective and surveillance has switched to a passive system.
Plans to scale back the number of passive surveillance sites operating RDTs are ongoing in Uganda, with the target of ensuring that by 2030, �50% of at-risk populations should live within 1-hour of a health facility with HAT diagnostics and �95% should live within 5-hours [16]. The number of health facilities using HAT RDTs were reviewed in July 2014 and September 2015, in response to the evolving epidemiological situation of g-HAT in Uganda [25]. Previous analyses utilised the Euclidean distance to facilities as opposed to estimates of travel time. The number of facilities in operation may be further reviewed to better quantify accessibility, identify the minimal number required to ensure sufficient coverage, and to maximise cost-effectiveness.
Optimisation of passive screening capacity in NW Uganda must consider two key elements. First, there must be evidence of reduced transmission, be this supported by active screening or reduced abundance of tsetse [23,24]. Second, g-HAT has a very long interval between infection and detection of a case, and cases can remain undetected for some time [26], therefore we are required to consider the long tail of this case distribution and any reservoirs of infection [27]. Consequently, the resource review proposed here should target a wider area that has been historically at risk in order to detect any residual infections in the area. The streamlining or optimisation of the number of facilities might be considered a step towards establishment of a network of sentinel screening sites to monitor post-elimination [17].
The work described here aims to utilise information on the location of operational HAT diagnostic facilities, alongside estimates of travel time to said facilities, to inform an analysis designed to identify the optimal number, and placement of, surveillance facilities required to meet the in-country coverage targets defined above.

Study setting
The focal area of this study is north-western Uganda where we focus on seven districts within the West Nile region which form Uganda's g-HAT endemic area. These endemic districts are Arua (2,100 km 2 ), Maracha (693 km 2 ), Koboko (862 km 2 ), Amuru (3,625km 2 ), Adjumani (3,030km 2 ), Moyo (1,800km 2 ), and Yumbe (1,524 km 2 ). Collectively, these districts have a population of approximately 3.5 million people [28]. For the purpose of this study, from this point onwards, we refer to these seven districts as "north-western Uganda".

High-resolution resistance surface
To quantify travel time to diagnostic facilities, we first sought to generate a resistance surface for north-western Uganda at a 30m × 30m spatial resolution. Resistance surfaces (also termed 'friction surfaces') contain estimates of associated travel cost for gridded cells within a Cartesian plane and are used within cost-distance analyses to quantify the effort required to travel between each cell and an origin point [29]. To construct the resistance surface, we collated data from a variety of sources. We obtained comprehensive road network data from Open-StreetMap [30], and assigned travel speeds along differing road classes based off observed speeds from GPS tracks obtained during February-April 2018 [31]. Speeds were representative of motorcycle travel, in agreement with the most commonly used mode of transportation in Uganda [32]. Estimates of off-road travel were generated utilising remotely sensed Landsat-8 data [33], and a normalised difference vegetation index (NDVI) [34,35], paired with information from studies detailing travel time through varying vegetation densities [36,37]. Finally, we combined off-road and on-road travel estimates to produce a comprehensive resistance surface detailing the time taken to traverse through each 30m × 30m gridded cell. We evaluated the accuracy of the resistance surface by comparing predicted travel time along 1,000 routes with times derived from Google Maps within a linear regression [38]. An in-depth description of the process is provided within S1 File.

Health facility data
In 2013, a new passive surveillance strategy was implemented in north-western Uganda, which saw the number of facilities participating in passive surveillance increase from four to 212, which was subsequently adjusted to 170 in 2017 [25]. Each facility utilises either RDTs alone [39], or RDTs and microscopy techniques [40], and is supplemented by mini-anion exchange centrifugation technique (mAECT) [41], with three of these second group facilities also equipped with loop-mediated isothermal amplification (LAMP) for identifying suspected g-HAT cases ( Fig 1A) [42]. For facilities with available data, the number of RDTs used during 2017 is shown within Fig 1B. Facilities operating LAMP and LED diagnostics serve as referral sites for RDT positive individuals, and are representative of larger facilities with higher quality services [25].

Human population density data
To detail the number of individuals within HAT risk areas who live within 1-hour and 5-hours travel time of a HAT diagnostic facility, we utilised a high-resolution population density surface [28]. This surface matched the spatial resolution of the resistance surface generated above (30m × 30m) and was re-sampled to ensure gridded cells aligned. A crude estimate of the PAR of g-HAT within the study region was determined by creating a 10km radius buffer around reported cases between 2000 and 2018 (Fig 2A) [21,25], representative of moderate risk ( Fig  2B). Case data were obtained from the WHO HAT atlas [21], and the occurrence locations represent either the residence of the individual (data collected through active screening), or the location in which a case was diagnosed (if diagnosed during passive screening). Unfortunately, there is uncertainty surrounding the location of infection with both collection methods, and one of the advantages of buffering case data is that this approach accounts for some of the uncertainty surrounding the true location of infection. Several papers quantifying g-HAT risk similarly use a buffer around case data, albeit a more conservative distance of 30km (see Franco et al. 2014 [19], 2017 [17], 2018 [44] and 2020 [20]). The population-at-risk surfaces within Franco et al. are generated for the whole continent and we believe that, in the absence of a HAT-risk model, this approach is a good quantification of those at risk within north-western Uganda. Buffering case data identified 3,025,801 individuals living within at-risk areas (PAR provided as S1 Data). The PAR surface used in this analysis captures the historic distribution of g-HAT within north-western Uganda and is therefore greater than PAR surfaces generated using only contemporary data (e.g. Simarro et al. [45]). The inclusion of historic data (2000-2018) provides higher confidence in the detection of disease and residual cases from across the area where transmission has occurred previously.

Simulation of accessibility to diagnostic centres under differing placements
Information on the location of existing g-HAT diagnostic facilities may be paired with the resistance surface to answer any of the following questions: 1. What is the smallest collection of the existing facilities that would meet the coverage requirements (i.e. �50% of PAR within 1h; �95% within 5h)?
2. What reductions in health facility numbers can be made without significantly reducing current levels of coverage with respect to 1h and 5h travel time targets?
3. Given a fixed number of facilities (for example, informed by funding availability or from answering questions one or two), what is the best spatial placement of said facilities to ensure maximum coverage of the PAR?

Deriving the optimal number of facilities
To derive the optimal number of facilities required within north-western Uganda to ensure that �50% of the PAR live less than 1 hour, and �95% of the PAR live less than 5 hours from a health facility with HAT diagnostics respectively, we performed a simulation study. Suppose S n is the set of all possible combinations of n facilities from the 170 currently available. We wish to obtain an estimate of the average PAR living within 1 or 5 hours of a facility for all possible values of n, but once n>1, the total number of combinations of n facilities is very large and computationally prohibitive. As such, we adopted a simulation approach, such that for n = 2,..,170 we generated S n which consisted of 1,000 random samples of n facilities. For each sample, i = 1,. . .,1000 we calculated the size of the PAR within 1, (A i n ) or 5, (B i n ) hours of any of the n facilities using a costdistance analysis implemented within R (version 3.5.1). From this we can obtain M n , the average number of people within 1 hour, and V n , the average number of people within 5 hours of any of the n facilities (Eqs 1 and 2). Q n represents the proportion of simulations for which more than 50% of the PAR, (D), are within 1 hours travel to any of the n facilities (Eq 3), and Z n represents the proportion of simulations for which more than 95% of the PAR are within 5 hours travel to any of the n facilities (Eq 4). To avoid spatial clustering within samples, we utilised an inhibitory sampling approach when generatingŜ n , ensuring a minimum distance of 4km between sampled facilities [46]. The optimal facility number simulation is outlined in Eqs 1-4: The minimal number of facilities, N min , was identified as the lowest value of n for which both Q n and Z n > 0.

Deriving the optimal placement of facilities when the facility number is known
To demonstrate how our approach can be used to determine the optimal placement of a fixed number of facilities when the target number to retain is known (i.e. informed by the above simulation, or when predefined by disease surveillance budgets), we implemented a further simulation analysis. For this simulation, when given a fixed number of facilities, n, we wish to identify the optimal subset of n facilities, from the 170 available which results in the highest percentage of the PAR within 1-hour and 5-hour travel. N'dungu et al. [47], detail in their work reporting on the Trypa-NO! project within north-western Uganda, that there is funding available to retain a maximum of 51 g-HAT diagnostic facilities from the 170 facilities currently operational. Here, using 51 facilities as an example, we demonstrate how the optimal placement of these facilities can be determined from an existing distribution using a cost-distance simulation. It should be noted that the number of available combinations (C(170, 51)) is computationally prohibitive, therefore we employed a simulation approach utilising 10,000 random combinations of 51 facilities. For each sample, the total population within 1-hour and 5-hour of a facility was determined utilising a cost-distance analysis (A j 51 and B j 51 respectively, j = 1,. . .,10000). For these simulations, we did not employ an inhibitory sampling approach, with no minimal distance enforced between selected facilities within each sample. The optimal selection of the 51 facilities was the combination that resulted in the largest number of the PAR being within the pre-defined criteria, as determined by a cumulative rank assessing percentage coverage within both 1-hour and 5-hour thresholds.
As a baseline for which to assess the impact of a scale back on accessibility of the PAR, we generated a cost-distance surface including all 170 facilities, representative of the situation in 2017. We then compared the predicted travel time for each 30 × 30m cell under the optimal scale back scenario, with the predicted travel time under the scenario representative of 2017, to determine which geographic locations would be most affected.

High-resolution resistance surface
The high-resolution resistance surface generated showed good agreement when compared with freely obtained estimates of travel time from Google Maps, for 1,000 randomly generated validation routes across north-western Uganda (R 2 = 0.902, p<2e −16 ) (Fig 3A). The spatial distribution of the origin and destination locations used within the validation process is shown as Fig 3B. Whilst there is strong correlation, modelled travel times are predicted to be slightly slower when compared to Google (root-mean-square error = 14.98), this is likely due to our estimates relating to motorbike based travel, vs Google times being representative of car travel. Nonetheless, this validation process provided confidence in the resistance surface, and subsequently in our estimates of travel time to diagnostic facilities. To facilitate reproducibility and additional research applications within the region, the resulting resistance surface is provided as S2 Data.

Deriving the minimal number of facilities
Results from the simulation experiment defined in Eqs 1-4 indicate that within north-western Uganda, a minimum of three facilities are required to ensure that 50% of the PAR live within 1-hours travel of a HAT diagnostic facility, and a minimum of one facility is required to ensure and 95% of the PAR live within 5-hours travel (Fig 4), when considering a subset (1000 simulations) of all possible combinations of n facilities. These results, however, may be biased toward highly populated areas (i.e. urban centres) and it is unlikely that an initial scale back of this magnitude would be adopted by a national program due to a range of political and contextual factors. We present a plot showing the median percentage of the population living outside of 1-hours travel time as Fig 4B. This plot demonstrates that to maintain similar levels of coverage to that achieved by all 170 facilities in operation, scaling back to 51 facilities would ensure that 95.25% of the PAR would be within 1-hours travel of a facility, opposed to~50% of the PAR with the much reduced scale back to three facilities. We focus on this 1-hour threshold, as the 5-hour threshold is easily attained across our study area. Although we demonstrate that the �50% within 1-hour and �95% within 5-hour coverage targets can be achieved with very few facilities, our results question the effects of such a dramatic scale back on access inequality. Coverage targets can be easily met by a low number of facilities if such facilities are located within urban centres; this approach, however, may further isolate remote rural populations if care isn't taken when planning a scale back.

Deriving the optimal placement of facilities when the facility number is known
As a baseline for which to assess the impact of a scale back on the accessibility of the PAR, we first generated estimates of accessibility to all 170 facilities which were operational in 2017 ( Fig  5A and 5B). The results from the cost-distance analysis utilising all 170 facilities indicate that 96.15% and 99.77% of the population at risk of HAT live within 1-hour and 5-hours travel time of a HAT diagnostic facility in 2017 respectively.    (30m x 30m) show the realised distribution of the population at risk, where each pixel represents one or more individuals living inside of the mask for the buffered case data (mask shown as dark grey). The black line running through the centre of our study area represents the Albert Nile river. Maps produced in QGIS [43]. To demonstrate the ability of this approach to identify the optimal placement of facilities when the facility number is known, either as derived from the above simulation, or predefined, we generated 10,000 estimates of travel time to facilities under varying selection combinations of 51 facilities, the number operating from late 2018 onwards [47]. For each simulation, we evaluated the percentage of individuals living within each coverage category (i.e. within 1-hour and 5-hours travel of a facility). The optimal placement of the 51 facilities, from the available distribution of 170 operational facilities resulted in no significant difference in the proportion of the PAR living within 1-hours travel of a facility, when compared to the estimate utilising all 170 facilities (P = 0.91; all 170 facilities = 96.15%, 51 facilities = 95.25% coverage). As expected, the coverage of the PAR within 1-hour travel, when quantified using the optimal placement of these 51 facilities far exceeds the in-country coverage target of 50%. Although our 10,000 simulations were only a subset of the total possible combination of 51 facilities from a set of 170, (C(170, 51)), we show that multiple combinations results in coverage estimates with negligible differences, particularly in reference to the 5-hour threshold ( Fig  6). Of our 10,000 simulations, 9920 (99.2%) are estimated to have a 5-hour threshold coverage above 99.7% (meeting the coverage value if all 170 facilities are considered). Of these, 9028 (90.28%) have a 1-hour threshold coverage above 90%, whereas the 1-hour coverage is 96.15% if all 170 facilities are considered. We know that the maximum possible coverage cannot exceed that obtained using all 170 facilities (96.15%). As a result, while there may exist an excluded combination that exceeds the coverage obtained by the best identified set in our simulation (95.25%), any improvement would be marginal and negligible (at most 0.9%).
A comparison showing the accessibility to diagnostic facilities is given in Fig 5, detailing accessibility to all 170 facilities in operation during 2017 (Fig 5A and 5B) and the scaled back number of 51 facilities operating from late 2018 [47], with the optimal placement determined through our approach (Fig 5C and 5D). To aid interpretation, we visualise the cost-distance surfaces on a continuous scale (Fig 5A and 5C), and on a categorical scale, with travel time binned by 30 minute increments (Fig 5B and 5D). The difference in travel time from each 30m x 30m gridded cell, between all 170 facilities and 51 facilities is given in Fig 5E. A map detailing the optimal placement of the scaled back facilities is provided as Fig 5F, and a table listing the name, district and parish of facilities to preserve is provided (Table B in S1 File).
Areas of greatest travel time to a facility include those locations which have a low population density, such as Maaji Parish in South-West Adjumani; Waka Parish in Obongi County, Moyo District and Kei Fr Parish in Aringa county, Yumbe District South-West, among others. Additionally, large areas of Southern Arua have populations living �1-hour of a diagnostic facility, however, these populations are not generally considered to be living within areas of g-HAT risk so do not affect the population coverage estimates generated here. Maps showing Fig  5 masked by the population at risk are provided as Fig B in S1 File.

Discussion
Twenty years of a global programme for active screening and treatment of g-HAT has driven the global incidence to a historic low [17]. Building on this achievement, recent advances in methods for detecting and treating g-HAT [18,42], combined with more cost-effective methods of tsetse control [24], are providing the platform for achieving the WHO goals of eliminating g-HAT as a public health problem by 2020 and interrupting transmission completely by 2030. There will be a need for technical guidance on how to monitor the distribution and incidence of g-HAT in this elimination setting. Establishing cost-effective surveillance systems that can be maintained over the next decade will be crucial for the achievement of these ambitious goals.
Using a resistance surface informed by motorbike based travel, and a simulation-based analysis, we determined that the number of operational g-HAT diagnostic facilities within north-western Uganda could be reduced from 170 to three (Fig 4), whilst ensuring that percentage coverage targets of �50% of the PAR living within 1-hours travel and �95% 5-hours travel are met. Ensuring that �95% of the PAR lives within 5-hours travel of a facility can be achieved by retaining only one well-placed facility. This implies that such targets are easily attainable within the geographic extent of our study area. Scaling back to this degree however poses concerns regarding equity of access [48]. For a disease like g-HAT, which predominantly affects rural populations, severe scaling back to achieve moderate targets could leave remote communities without coverage. Ultimately, the degree of scale back should result from discussions with programme implementers and stakeholders, utilising empirical evidence from analyses such as those performed here in addition to the consideration of other operational and contextual factors. Given that it is known that a small number of facilities is enough to achieve the coverage targets, this methodology allows more ambitious targets of coverage to be considered. Specifically, we demonstrate that the number of facilities can be reduced by 70% with minimal loss of coverage. Through the scale back of the number of facilities, we can improve cost-effectiveness whilst preserving coverage of the entire PAR, including remote populations. These simulations can be easily repeated as the spatial distribution of the PAR changes over time, through substitution of new surfaces detailing contracting or expanding PAR in response to case detection. This will allow a dynamic and rapid response to changing circumstances. The PAR used in this analysis captures the historic distribution of g-HAT within north-western Uganda and is therefore greater than PAR surfaces generated using only contemporary data (e.g. Simarro et al. [45]). Our results could however be applied to more recent infection periods.
In this study, national borders, such as those with the neighbouring countries of South Sudan and the Democratic Republic of the Congo, are presumed to prevent individuals who are citizens of either of those two countries from seeking diagnosis within Uganda. Conversely, sub-national borders, (e.g., districts) are presumed not to prevent movement of individuals seeking treatment. As the population density surface used to construct the PAR surface pertains to individuals living within Uganda only, we currently do not capture international movement from neighbouring countries. Our friction surface would need to be expanded to include areas within DRC or South Sudan if we wish to include these populations in our simulations. The methodology described here, although applied to the surveillance of HAT, can also be adapted for a range of elimination scenarios and can also be utilised for the monitoring of emerging diseases as part of national early-warning systems.
Accessibility to a healthcare facility does not correlate precisely with Euclidean distance [49]. The use of a validated resistance surface accounts for varying land cover and road network accessibility, factors ultimately affecting travel time, especially for remote and rural communities. Due to data limitations, associated mostly with the clarity of remotely sensed imagery for characterising off-road travel, our resistance surface is representative of travel during the dry season only. Travel to diagnostic facilities will be more difficult during the wet season, particularly along untarmacked roads. By producing our resistance surface at a high spatial resolution, we also provide a more accurate quantification of travel time to a facility than existing surfaces at coarser granularity [50], giving higher precision when quantifying populations within each travel time category. Whilst the time taken to access a facility likely influences an individual's decision to seek treatment, additional factors are also at play. Future efforts should investigate the use of spatial interaction models to quantify the influence of other metrics on the probability of treatment seeking at a specific location, i.e. gravity models weighted by staff capacity and their levels of training, status of the laboratory and available materials for instance [51,52]. The construction and validation of such models requires empirical measurements of treatment-seeking behaviour [53]. Our analysis investigated the optimal retention of facilities in a specific spatial setting. The quality of services available at some facilities will be different to others, however, our approach is sufficiently flexible to allow users to retain required sites (for example, specialist referral centres), and there is scope for national health systems to reallocate resources based on the spatial selection identified.
Although the travel time targets utilised within this analysis were pre-defined [16], further analysis is required to determine the effects of these targets on treatment seeking behaviours. For instance, the cost of continuous travel to a facility which is 1-hour or more from an individual's residence may present significant barriers to treatment [32], and our analysis shows, that within north-western Uganda, �95% of the population are within 5-hours travel to just one well-placed facility, questioning the value of this target for this area. Reducing the number of accessible facilities may also influence factors such as compliance with post-treatment follow up, with a previous study identifying costs for transportation to a facility as a reason for noncompletion of treatment [54]. Alongside any adopted scale-back approaches, there is a need for sensitisation and information campaigns to inform the PAR of the reduced availability of diagnostics as this information may influence treatment seeking behaviours [55]; this campaign can be used alongside a referral system to address such issues.
Scale back of disease surveillance should be integrated within a national development plan and combined with the scale back of intervention programmes. In the context of Uganda, work is ongoing to investigate the effect of reducing the geographical extent of tsetse control operations to areas where g-HAT persists. As disease prevalence changes within a geographic area, approaches are required to ensure that appropriate surveillance and treatment systems are in place. The method described here aims to ensure maximal coverage of surveillance under the most cost-effective scenarios. For national programmes to adopt an optimised approach and integrate this approach within health care delivery systems, it is important to demonstrate that the PAR remains adequately covered, health facilities are accessible, and that a functional referral system is maintained. Providing strategic health facilities with a full range of tests for disease confirmation is critical for elimination.
For other NTDs, plans to scale back surveillance require careful consideration of the country and disease in question, alongside additional context-specific factors [1]. As such, the altering of disease surveillance systems should be an iterative process requiring regular reassessment of objectives and methods [9]. The approach described here provides a robust, reproducible framework to ensure maximal coverage of the PAR through the minimal number of sites within a passive disease surveillance system.

Acknowledgments
The data presented in this analysis are based on HAT control activities implemented by the Ugandan Ministry of Health, and supported by the Foundation for Innovative New Diagnostics (FIND) with support from the Bill & Melinda Gates Foundation, the Republic and Canton of Geneva, the German Federal Ministry of Education and Research through KfW, the Swiss Government, UK aid from the UK government and the Medical Research Council. We are grateful to all health facility staff, district leadership and district coordinators in Uganda that enabled this project. The World Health Organization (WHO) have supported targeted active screening in affected villages and kindly provided case data through the human African trypanosomiasis atlas. We also wish to thank Dr Simon Wagstaff and Mr Andrew Bennett (both LSTM) for providing the computational resources for this analysis.