Modelling access to emergency obstetric care

Access to health services is influenced by economic conditions, topography and infrastructure. Modelling travel time is an effective way to measure ease of access to services such as EMOC in regions of low income, poor infrastructure, and difficult geographic terrain. For example, Bailey et al. [14] measured time taken to transfer patients needing emergency obstetric and newborn care from first-level facilities to referral facilities in the Tigray and Amhara regions in Ethiopia, a country which has sparsely populated mountain areas where road infrastructure is poor [2]. Chen et al. [15] modelled travel time to reach emergency obstetric and neonatal care in the Kigoma Region, Tanzania, where geographic access to such care is poor. Makanga et al. [4] modelled travel time to maternal health services in southern Mozambique, which has poor infrastructure; almost 90% of the roads are unpaved, of which 75% are minor unplanned roads. In the wet season, roads are difficult to traverse; when flooded, they are unusable. Myers et al. [16] modelled travel time to EMOC in Timor Tengah Selatan (TTS) district, eastern Indonesia. The terrain of the district is mainly rough, there are few roads, and those are in poor condition. In the rainy season, the roads are difficult to traverse because of flooding or landslides.

Travel was modelled using two inputs: travel time and cost. The first was based on the location of health services and the second on the cost of movement across each geographical feature [17]. Geographical features were entered as land cover raster grids that combined vegetation, transport infrastructure (roads), watercourses and elevation data [18]. Each of the land cover grids assumed different speeds of travel. For example, Okwaraji et al. [17], when modelling time to travel to health facilities in Ethiopia, assigned a walking speed of 5 km/h across level ground, and 0.1 km/h up slopes greater than 30 degrees and through water. Myers et al. [16], when modelling time of travel to EMOC in a district in eastern Indonesia, used local knowledge to define travel speed, based on type of transportation: by ambulance (car), public transport (bus), and walking through different types of land cover and weather conditions (dry and wet seasons).

In this study, travel time was generated with GIS software using raster analysis techniques. One example of raster analysis software used for modelling travel time is AccessMod [1, 4, 16, 19–21], developed by the World Health Organization (WHO), which uses cost-distance analysis to model accessibility of health services. Initially a plugin in ArcView, AccessMod is currently an extension in ArcGIS 9.3.1. Other studies have used cost analysis on IDRISI Taiga GIS software to calculate travel time to health posts [22].

However, GIS software generates only a static model of travel time, while travel time is dynamic. Myers et al. [16] stated that simplistic scenarios cannot capture the variability of changing weather; they suggested using an interactive model to capture various scenarios. Therefore, to capture a thousand various travel time scenarios, Fisher and Lassa [18] developed a geo-simulation interactive tool using GIS format in NetLogo [11]. Travel time is calculated automatically through geo-simulation. We followed the method Fisher and Lassa [18] but added a cost-benefit analysis.

Geo-simulation

Modelling travel time can be simulated using cost-distance transform and chamfer metrics [18]. This algorithm calculates distance between neighbouring cells on a grid [23]. Fisher and Lassa [19] calculated distance from the centre point using the 5 × 5 distance transform mask. Calculation of cell distance is done by scanning vertically from top left to bottom right corner, then bottom right to top left corner, then repeated as a horizontal scan. The process scans a square or rectangular lattice by placing a mask over a cell in the grid. The lattice can be a single source or destination point, or a set of points which are automatically generated. The destination points are initialised to 0 and all other points are assigned as large values, for example, 9999 [23] or 99 999 [18]. Points assigned large values were designated barriers or inaccessible areas. The scan calculated travel time at each point from the central cell to neighbouring cells, using speed values of each cell multiplied by its fractional value, 0.9866, 1.4141 or 2.2062 [18, 23]. The minimum calculated travel times of surrounding cells are added to the current cell travel time value if they are smaller than the current value. After scanning, the distance to the nearest point in the set of source points is assigned to each cell in the final lattice.

Study area

This study was undertaken in the Kupang district in the province of Nusa Tenggara Timur (NTT) in eastern Indonesia. The district was selected as a case study having high maternal deaths. The regional medium-term development plan of Kupang district planned to decrease the MMR to 225 per 100 000 live births by 2015 [24]. However, the district did not achieve this target: in 2018, it was 257 [25]. Most maternal deaths occurred at home because of difficulty of access to the closest clinic [24].

The population of Kupang district in 2019 was 380 900 people across three islands (Timor, Semau and Kera) [26]. Most people live on Timor and Semau islands. Only 82 families live on Kera island [27].

The district has five EMOCs: Puskesmas Oesao, Oekabiti, Takari, Uitao (on Semau Island) and Lelogama. There is one PONEK (rujukan obstetri dan neonatal emergensi/komplikasi) at a public hospital, RSU (Rumah Sakit Umum), which provides comprehensive emergency obstetric care (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Locations of PONED and PONEK in Kupang district in the red stars.

https://doi.org/10.1371/journal.pone.0251869.g001

Data collection

When developing the scenario modelling tool, we collected datasets as input for the model and software from 8 sources:

1. Five geographical coordinates of the location of PONEDs (EMOC) and PONEKs were taken using a handheld global positioning system (GPS) device, Garmin eTrex^® 30x.
2. Satellite imagery from Landsat 8 (Path: 111/67) [28] and DEM (digital elevation model) of Timor Island taken in 2016 [29] were used.
3. Datasets of road networks in 2016 were obtained from bappeda (district planning agencies). ‘Missing’ roads, that is, those not on the maps, were digitised using satellite imagery in QGIS [30]. The classes of roads were (1) national, (2) provincial, (3) district, (4) village.
4. Roads to the locations of PONEDs and planning meetings were plotted from Oelamasi, the capital city of Kupang district.
5. The residential map from bappeda in JPEG format was digitised into SHP format by the first author [30].
6. Maps of administration boundaries were obtained from bappeda.
7. GPS datapoints of the location of rivers and broken roads were collected.
8. GPS ground references of each land cover type (forest, savanna, ocean, farm and residential, river) were taken.

We used three free simple GIS software packages to create the travel time scenario model:

1. Saga [31]
2. QGIS: Quantum GIS [30].
3. NetLogo [11]

Development process

This section presents the development process of the scenario modelling (with cost-benefit analysis) tool for infrastructure planning. We followed three steps (Fig 2).

1. Step 1: Creating land cover grids on Saga. This required input from three datasets: roads, Landsat imagery and DEM.
   All data layers were projected onto the spatial reference frame WGS84/UTM 51S, then cropped to the administrative boundaries of each EMOC catchment area. The Landsat image was reclassified, using an unsupervised classification, to derive vegetation grid layers (forest, savanna, ocean, farm and residential). The DEM was used to produce a watercourse grid (stream classes 1–6) by implementing a Fill Sinks operation producing a rivers network vector with Strahler order [18]. The roads data was classified into four types: village, district, provincial, and national. The roads layer and rivers network layer were gridded into the raster grid layer. The EMOC point locations were gridded as EMOC layers. We assigned the ID value of 1000 for EMOC. Then the four raster layers (vegetation, roads, rivers network and EMOC) were mosaicked into one land cover grid layer consisting of 16 classes. Each class has a different travel time, calculated in seconds for each 30-metre cell (1): (1)
   TT: Travel Time in seconds; Travel speed in km/hour. Results for each class are shown in Table 1.
2. Step 2: The district residential image in JPEG format was digitised in QGIS to shapefile format, then gridded into a raster layer in Saga (Fig 3) with two classes, 1 and 0 (0 for no data).
3. Step 3: The grid layers of land cover from Step 1 and the residential grid layers from Step 2 were then converted into the ASC format used for creating the dynamic travel time scenario modelling tool on NetLogo. The PNG image of the land cover grid, which shows the location of EMOCs, was used as background of the interface of the tool. The two ASC files of land cover and residential grids were used to calculate travel time. Travel speed across each class of land cover was assigned on the NetLogo interface. The location of any new EMOC was defined by users.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Design of scenario modelling tool.

https://doi.org/10.1371/journal.pone.0251869.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Grid layer of residential areas in red.

The white dots as EMOC facilities.

https://doi.org/10.1371/journal.pone.0251869.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Travel time and speed for each class of landcover.

https://doi.org/10.1371/journal.pone.0251869.t001

To create the scenario modelling tool, we set up the interface in NetLogo based on the model design (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Interface of the travel time scenario modelling tool.

https://doi.org/10.1371/journal.pone.0251869.g004

The interface consists of 16 items; four coding operations were performed for each of these items.

1. Reset procedure. This procedure clears all data and sets the patch colour based on travel time ranges.
   if (travel-dist / 60) > = 0 and (travel-dist / 60) < = 30 [set pcolor 103];blue
   if (travel-dist / 60) > 30 and (travel-dist / 60) < = 60 [set pcolor 94]; sky
   if (travel-dist / 60) > 60 and (travel-dist / 60) < = 90[set pcolor 55]; green
   if (travel-dist / 60) > 90 and (travel-dist / 60) < = 120 [set pcolor 45]; yellow
   if (travel-dist / 60) > 120 and (travel-dist / 60) < = 9999 [set pcolor 14]; red
   if (travel-dist / 60) > 9999 and (travel-dist / 60) < = 9999999 [set pcolor 0]; black
2. Setup procedure. This procedure sets up the screen with GIS data uploaded from the ASC land cover grid file.
   ask patches [set cost 0 set dist 999999]; set cost 0 for distance 999999 from PONED
   ;;load landcover layer
   set covergis gis:load-dataset "data/LC-Takari.asc"
   resize-world 0 gis:width-of covergis 0 gis:height-of covergis
   gis:set-world-envelope
   (gis:envelope-of covergis)
   if active-gis [gis:apply-raster covergis cover set-cost]
   if active-gis [import-pcolors-rgb "data/Takari.png"]; input image map
   ask patches with [cover = 1000] [sprout-dest 1 [set shape "cylinder" set size 2 set color white]]; PONED was assigned as 1000
   ask patches with [cover = 1000] [set dist 0]
3. Simulation procedure. This procedure implemented the distance transform (DT) and chamfer scan [23] to calculate the minimum distance between each point cell. As an example, the DT formula for forward scan at patch (-1 2) is written as:
   Ifelse (([cover] of patch-at -1 2 = 1) or ([cover] of patch-at -1 2 = 2) or ([cover] of patch-at -1 2 = 3) or ([cover] of patch-at -1 2 = 4) or ([cover] of patch-at -1 2 = 5) or ([cover] of patch-at -1 2 = 101)or ([cover] of patch-at -1 2 = 102)or ([cover] of patch-at -1 2 = 103)or ([cover] of patch-at -1 2 = 104)or ([cover] of patch-at -1 2 = 105)or ([cover] of patch-at -1 2 = 106) or ([cover] of patch-at -1 2 = 201)or ([cover] of patch-at -1 2 = 202) or ([cover] of patch-at -1 2 = 203) or ([cover] of patch-at -1 2 = 204)) [set LDME1 (C1 * cost + ([dist] of patch-at -1 2))] [set LDME1 999999]
   where LDME1 is the least distance mask element 1. The fractional values for the DT are:
   set a1 0.9866 set b1 1.4141 set c1 2.2062
   Four scanning processes were applied: forward scan, back scan, top scan and bottom scan (Fig 5).
   Before scanning, the travel cost of each cell grid was set, based on land cover type (previously assigned on Saga). Travel cost of roads and Stream 3 were defined in NetLogo. Travel cost of roads was taken from the value of each road’s sliders in the Netlogo interface; these were used to assign the travel speed (in km/h) on national, provincial, district and village roads. Planned new roads were created by drawing them in using the code shown below:
   if (mouse-down?)
    [ask patch mouse-xcor mouse-ycor
     [
     set pcolor 30; change this color coding for boat’s route to white
     ;set pcolor white
     ifelse Active-GIS [
     if (Road-Type = "National") [set cover 204]
     if (Road-Type = "Provincial") [set cover 203]
     if (Road-Type = "District") [set cover 202]
     if (Road-Type = "Village") [set cover 201]]
     [set cost 2]
       ]
    ]
   Travel cost of Stream 3 varied with the weather. It is a big river that can be crossed in the dry season, but it floods in the rainy season and is impassable. Therefore, the dry season travel cost was set at 108, but at 99 9999 for the rainy season (i.e. stream number 3 is a barrier). A PONED location can be drawn by clicking the mouse, causing a white circle to appear on the map. The code is:
   if (mouse-down?)
    [ask patch mouse-xcor mouse-ycor [sprout-bar 1
     [set cover 106 ask patches in-cone 1 360 [set pcolor black] die]
     ] ]
   The barrier was drawn by giving the patch the same value as that of land cover number 106, which has cost 99 999. The code is:
   if (mouse-down?)
   [ask patch mouse-xcor mouse-ycor [sprout-dest 1
   [set dist 0 ask patches in-cone 2 360 [set pcolor white] die]]]
4. Cost-benefit procedure. Cost-benefit analysis was undertaken using cost of infrastructure and benefit of people who can travel to EMOC in less than 1 hour. Cost of road building was calculated by multiplying the road length with the price of road each metre. To get the road length, we first calculated the Euclidean distance for 30-metre pixels (Fig 6), the cell size of the patch, with Eq (2). (2)

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Scanning process.

Clockwise from top left: forward scan, back scan, top scan and bottom scan.

https://doi.org/10.1371/journal.pone.0251869.g005

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Pixel size of each patch.

https://doi.org/10.1371/journal.pone.0251869.g006

The cost of road building can be calculated by Eq (3): (3) Where inputcost is the cost of each metre of road and countroad is the total road length.

In the cost-benefit analysis, the benefit is the number of people able to travel to EMOC for each travel time zone. It was calculated using population density for each EMOC catchment area and residential grid cell. Each grid cell on the map has 30 m resolution per pixel. If the population density is 68 people/km² for each grid cell, the number of people per cell is equal to (30 x 30 x 68/1 000 000).

The residential grid cells map was gridded from a residential location map of Kupang district on Saga (Fig 6). The map was then converted into ASC format, uploaded into NetLogo, and the number of people reaching EMOC was calculated on NetLogo using the same method as the Spatial Analyst tool/Zonal/Tabulate Area tool in ArcGIS [32]. This method was used to assign residential areas for each class of travel time. Coordinates (x,y) of residential data (LC-Residential) are read and stored through the scanning process:

if ([LC-Residential] of patch-here = 1) [set list-ta lput (patch pxcor pycor) list-ta]

The list of the coordinate patch (x,y) is used to obtain areas according to the classes of travel time. The number of patches for each travel time class was calculated as follows:

set count-blue (count patches with [pcolor = 103])

set count-sky (count patches with [pcolor = 94])

set count-green (count patches with [pcolor = 55])

set count-yellow (count patches with [pcolor = 45])

set count-red (count patches with [pcolor = 14])

The estimated number of people in each of the classes of travel time was presented on a graph. For the cost-benefit analysis, the number of people who travel to EMOC in less than 1 hour was compared with the cost of building new infrastructure.

Verifying and validating the model

The purpose of verification and validation is to gain credibility for the model with decision makers [33]. They will have confidence in the model if it is shown to be a valid and accurate model of the real-world system which it represents.

We verified this model during its development to determine whether the program was working as we had planned. We also checked that the output was correct with a variety of inputs, thus establishing the accuracy of the model.

After verification, we performed validation. The model’s objective is to model scenarios of infrastructure planning to improve access to EMOCs by decreasing travel time. The validation process comprised three steps. First, we made field trips from residential locations to EMOC locations, calculating actual travel time, to check whether the model’s results were similar to the real travel time. Second, we checked the population result with the population data of Kupang district from the district census data [34]. Third, we introduced the model to the district planning officers and midwives to check whether the model fulfilled its purpose. Following verification and validation, we sought and received the consent of the research participants. First, we asked permission from the Governor of NTT province and from the Headman of Kupang district to conduct this research. Then we conducted a meeting with a bappeda (district planning agency) and then with midwives in different locations. At the bappeda meeting, we obtained written consent of participants for the research activities, including video recording. Fifteen people attended the bappeda meeting. Because of the large number of participants (60) at the midwives’ meeting, and limited time, the meeting leaders asked the researcher to read the consent form and then asked the participants directly to give their consent verbally to the research activities. Acceptance of this process indicated that consent had been given for the research activities. Each participant was not identified.

Ethical considerations

Ethics approval for the study was obtained from the university’s Human Research Ethics Committee (reference number H14016). There was no likelihood of harm or discomfort to participants in this research project. Participants’ activities were recorded using a video camera. Any possible repercussions from criticising current planning processes were avoided by keeping responses confidential and reporting them in an aggregated way so that individual responses could not be attributed. Written informed consent was obtained from bappeda participants and verbal consent from midwives.

Evidence-based planning

To determine if our scenario modelling tool is easy to use as evidence for improving access to EMOC, we introduced it to district planning officers (bappeda, badan perencanaan daerah) and to the representatives of midwives from all districts in East Nusa Tenggara Province. Our case study was done in Kupang district, but we wanted to introduce the tool to other districts as well. The midwives’ meeting was arranged by the Health Department of Nusa Tenggara Timur Province. We introduced the tool to the midwives and trained them to use it. The midwives, who face problems in helping pregnant women in their areas, engaged with the modelling tool, using their local knowledge for the best results. Now this tool is known in all districts in this province.

We also introduced the scenario modelling to bappeda, facilitators of the planning process in the district. However, unlike the midwives, they did not have local knowledge of EMOCs and did not engage with the tool.

In our previous study, we found that in the district planning meetings, called musrenbang (musyawarah rencana pembangunan), the planning proposals were submitted from musrenbangdes at village level. Midwives attend the musrenbangdes. This tool can be used by midwives to submit their suggestions about improving access to EMOC. A midwife in the current study said:

This tool is very good. I want to share to my friends at village. So, when musrenbangdes, they can convince the village’s officers about planning to improve access to EMOC [PONED]. Then the village officers can bring this planning evidence to the musrenbangkab, so this proposed planning can be accepted.

The midwife realised that she needed evidence to submit in the planning process and that she could use scenario modelling to convince the musrenbang participants to prioritise her planning proposal. She realised that she should use the scenario modelling in musrenbang at village level because she cannot attend the musrenbang at district level. Only village representatives go to the next level of musrenbang [36] and, as representatives of musrenbangdes, they can take the scenario modelling to the next level of musrenbang to be used as evidence for planning. The tool makes midwives aware of the importance of using scenario modelling as evidence in submitting planning proposals at musrenbang. Previously, midwives did not know how to submit a proposal to increase access to EMOC.

The midwife’s desire to use this tool as evidence is confirmed by comments from bappeda:

It depends on its urgent. If there is debate at musrenbangcam (planning meeting at subdistrict), we can use this modelling as one of evidence to be considered in making decision of priority planning. For example, we want to compare scoring 3 and 2 of infrastructure planning which both planning ideas are important. But our budget is limited. So, we can use this modelling to choose which planning can improve access to health service quickly. This method is applicable for us.

Both midwives and bappeda wanted to use scenario modelling as evidence in the musrenbang. Scenario modelling can be used as planning evidence from musrenbangdes (planning at village level) to musrenbangkab (planning at district level). Bappeda facilitate the musrenbang process [37]; they form a team (called tim pemandu) to guide the musrenbang process. There is usually debate about priorities when making planning decisions; with limited budgets, it is difficult to determine which planning option is the most urgent and deserves priority. To overcome this problem, bappeda wanted to use this scenario modelling as the basis of planning and decision making to improve access to health services.

An interactive tool

Scenario modelling can be used easily–like playing a game. Users can test planning scenarios such as adding new roads, adding new PONED, and drawing barriers. They can compare various scenarios of infrastructure planning to improve access to EMOC. This tool provides basic knowledge of GIS for the layperson–for example, midwives without background knowledge of GIS can easily create travel time maps, read and understand the map, and analyse the results.

Improving the model

Participants made suggestions to improve the scenario modelling:

1. The scenario modelling tool should use the location of pregnant women. A midwife explained:

   This tool is suit for advocation. Because it is not easy to ask building road for Puskesmas (clinic). It takes years. If we are only waiting the road exist, maybe many pregnant women will die. It will be good if adding the location of pregnant women on the model. So, we can focus more detail to the exact women who need help.

   
   The scenario modelling tool has demonstrated that it can be used to analyse EMOC location. We did not include the location of pregnant women in this scenario modelling, since our model is based on the PONED guidelines of the Indonesian Ministry of Health [10], which decree that a PONED should service 50 000–100 000 people, not only pregnant women. Our modelling tool can also be used as an indicator of the need for emergency obstetric care by WHO [38], which uses total population, not only pregnant women, when determining the need for EMOC in a region. Hence, the tool enables health planners to analyse the location of EMOC based on the guidelines from WHO and the Indonesian Ministry of Health.
2. The scenario modelling should include small islands in its geographic conditions. The bappeda suggested including not only the main islands but also other small islands. One bappeda said:

   We have complicated problem with health access. How the community should get better access to health services. Thanks, ibu (called the researcher) to introduce travel time modelling for planning process. I suggested that the modelling should consider geographic conditions because this Timor land consists of many islands with different characteristic.

We have presented scenario modelling showing travel between islands, such as from Timor Island (the main island) to Semau Island. However, we only modelled areas that can be captured by NetLogo, which has limited maximum capacity. The model cannot capture all areas of the Kupang district, which consists of many small islands around the main islands. These small islands do not appear on the NetLogo interface. We used Netlogo because it is a free application.

We welcome any suggestions to improve our model. For the future, we recommend adding the location of women aged 15–39 years to the scenario modelling. Spatial distribution of these “at-risk” women of reproductive age can be analysed [39] to determine the number of these women who may need emergency obstetric care in less than 1 hour. A layer for these women could be aggregated into the travel time layers to calculate the number of at-risk women in each travel time zone. We suggest using district census data to compile a dataset of women of reproductive age [24].

Action plan for the study result

We have created an interactive scenario modelling tool with example data from the Kupang district as a case study. This tool can be used directly by the midwives and bappeda in Kupang district in the musrenbang. However, to enable the results of this study to be implemented in other areas, we suggest an action plan. That is first, bappeda should create the landcover map (step 1 of the method) and residential map (step 2 of the method) for the district as an input to the tool in Netlogo. Second, bappeda can then introduce the scenario modelling tool to the district health department. Third, the health department can organise a meeting with village midwives to introduce this tool to them. Bappeda and the health department then provide training to the midwives at this meeting to use this tool. The midwives can then use this tool to propose planning to improve access to emergency obstetric care (EMOC) in the musrenbang.

Conclusions

We developed an interactive scenario modelling tool that can be used by a layperson such as a midwife without prior GIS knowledge to support evidence-based planning for musrenbang from the village level in the district in eastern Indonesia. The tool can compare planning scenarios using cost-benefit analyses: the cost is the budget of the proposed infrastructure, and the benefit (in this proof-of-concept instance) is the number of people who can travel to EMOC in less than 1 hour. This tool can model travel time influenced by difficult geographic and terrain conditions based on the travel cost of moving around different types of landcover. It enables the midwives to identify and mark poor infrastructure and model travel across the ocean in the district, taking into account the season and problems such as flooding, damage roads, or broken bridges. Using this tool, the midwives will know how to submit a proposal to increase access to EMOC in district planning. The district planning agency (bappeda) will use this tool as the basis of planning and decision making to improve access to health services.

Our study has important implications for maternal health programs to decrease maternal deaths in low-income and lower-middle-income countries by improving access to emergency obstetric care in areas with difficult geographic conditions and poor infrastructure, lack of human resources, and limited budgets. This finding could help policymakers carry out effective planning to improve maternal health care service delivery and contribute to the achievement of Sustainable Development Goal #3 (SDGs), which emphasises the decreasing of the global maternal mortality ratio to less than 70 per 100 000 live births by 2030.