AMREF Health Africa and Amsterdam Business School (ABS) researchers Mayukh Ghosh, Chintan Amrit and Joaquim Gromicho developed a decision-support tool that combines estimates of healthcare demand with route optimisation for mobile clinics in Kenya. The tool was designed around AMREF’s COVID-19 vaccination programme with AstraZeneca, Kenya’s Ministry of Health and county governments in Kisii, Narok and Nyamira, and is intended to make limited mobile-clinic capacity easier to direct towards areas with lower healthcare access and higher expected demand.
Why is deciding where to send a mobile clinic difficult?
Mobile clinics can bring healthcare services closer to communities that have limited access to fixed facilities. But operating them requires choices about which locations to visit, in what order and how to use limited travel time and resources.
Those choices were particularly relevant to AMREF’s mobile COVID-19 vaccination programme in Kenya, where healthcare infrastructure is not distributed evenly. For example, Kisii had 46 higher-level health centres, while Narok had 13, despite similar population sizes.
The clinics also faced practical constraints. COVID-19 vaccines require controlled storage conditions, while the mobile clinic itself did not have the facilities needed for long-term vaccine storage. Routes therefore had to begin and end at an appropriate health centre where vaccines could be collected and stored.
At the same time, healthcare needs can change. Vaccination demand depends on factors such as access to existing facilities, population characteristics, vaccination rates, and willingness to be vaccinated.
The routing problem is therefore more complex than simply choosing the shortest journey.
Why is this a predictive and prescriptive analytics problem?
The project connects two analytical questions.
The first is predictive: where is demand for a mobile clinic likely to be highest?
The researchers combined information on population, vaccination status and healthcare accessibility to estimate demand. Accessibility was partly measured by looking at how many people lived within 2.5 kilometres walking distance of a health centre, a distance selected through discussions with AMREF field teams.
The second question is prescriptive: given that expected demand, which locations should the clinic actually visit and in what order?
The optimisation model compares possible routes and seeks a route that reaches areas with greater need while keeping travel costs low. Rather than requiring the clinic to visit every possible location, AMREF can set a minimum share of predicted demand that a route should cover, and the model identifies a lower-cost way of reaching that threshold.
This means prediction and decision-making are connected. The estimated demand does not sit in a separate report; it becomes part of the calculation used to determine the route.
The Decision Support Tool for Mobile Clinic Routing
What does the tool allow teams to do?
The researchers translated the model into a Microsoft Excel-based decision-support tool, with Python running the data processing and optimisation behind the interface.
Users can adjust parameters such as how much of the estimated demand a route should cover and which counties or sub-counties need to be considered. The tool then returns the selected locations, their visiting order, the route and its estimated travel cost.
Using Excel was also deliberate. The aim was to make the model accessible through software already commonly used by organisations rather than requiring practitioners to operate the optimisation code directly.
What has changed so far?
The project has produced a preliminary model and working decision-support tool that brings healthcare accessibility, predicted demand and routing decisions into one planning process.
It also created a way for AMREF teams to adjust the assumptions used by the model rather than receiving a fixed route from an external analytical system. However, the demand used in the preliminary evaluation was simulated from factors including population, vaccination status and accessibility, and the current version still requires field testing.
Why does keeping practitioners in the decision matter?
The tool is not intended to make healthcare deployment decisions independently.
AMREF teams can choose which locations must be included, determine how much demand a route should cover and update parameters as conditions change. The researchers made sure these choices were influenced by both data and the user’s domain knowledge.
That matters because quantitative demand is only part of a deployment decision. Field teams may also need to consider conditions that are difficult to capture fully in a model, including local operational knowledge and changes occurring between planning periods.
The tool therefore provides a route to consider rather than replacing the judgement of teams responsible for delivering services.
What can this mean for future mobile healthcare?
The immediate next step is to test the tool with AMREF stakeholders in field settings and use their feedback to refine both the model and interface. The researchers also plan to involve practitioners outside AMREF to assess whether the approach can transfer to other mobile-clinic settings.
Future versions could also replace simulated demand with estimates based on continually updated operational data. Information collected by mobile clinics could feed back into the system, improving later predictions and routes.
Although the first application focused on COVID-19 vaccination, the researchers identified potential applications in HPV vaccination, non-communicable disease services and pre- and post-natal care.
The wider opportunity is to move mobile-clinic planning from fixed routes or individual assessments towards a process that can be updated as healthcare access and demand change while keeping local knowledge and operational judgement part of the decision.
