Predictive Solution for AMR Risk

We will build and implement a real-time digital surveillance platform for public health organizations and government in Zimbabwe to use to manage and forecast antimicrobial resistance (AMR) risk at neighborhood level.

Dr Matthys Potgieter

According to WHO, providing strategic information such as the surveillance of AMR and antimicrobial use, are a key priority (https://www.who.int/news-room/...).

Due to the multifactorial nature of AMR epidemiology using a systems approach is required ( https://www.thelancet.com/jour...).  Measurable indicators are needed to plan and monitor interventions as part of a theory of change approach. Machine learning approaches that have the ability to find patterns in such complex data sets, could generate indicators that are interpretable and useful.

Various databases exist within laboratories and pharmacies, but are not interconnected. Climatic drivers and contextual drivers of AMR need to be integrated in a real-time manner. Zimbabwe aims to establish an effective AMR response (https://www.afro.who.int/count...). The objective of this project  is to create a centralized geospatial dashboard for national AMR surveillance at neighbourhood level, and enable data-driven decision-making in accordance with a One Health approach, but also to search for causal relationships (also paradoxes and similarities) in 2 data sources: evolution of sales/usage of antibiotics and the evolution of AMR cases, data will include geographic location (where people reside) of patients - to make sure we can work at neighbourhood level. 

The work helps the National AMR Coordination Committee, Ministry of Health and Child Care (MoHCC) central and regional health decision makers to predict trends likely to lead to AMR spikes. We aim to capacitate them with a tool that combines data from the multi-sectoral nature of AMR (medical, human behaviour, animal and plant health determinants), to predict trends that lead to AMR spikes. This will help them tailor make community level specific interventions to detect early and prevent AMR outbreaks in communities. We will engage and understand their needs through the following: 

1.StakeHolder Mapping, co-designing the intervention through protocol development meetings and approval of the model by the Committee,

2. Sensitise, involve MoHCC and regional leadership and seek their concurrence through the phases of implementation, a surveillance task force will be formed to lead the implementation in country, while the consortium supports.

3. MoHCC will lead all support/ supervision and data collection exercises and be involved in corrective action development and execution.

4. CHAI works within our partner governments, to ensure MoHCC needs are captured and catered for, MoHCC will Co-Pi the model development and implementation.

The EPCON platform allows optimized intervention planning based on integrating data, interventions, and outcomes, to assist public health organisations reach maximum impact - in this case using cutting edge technology to identify high risk for AMR hotspots that need resource mobilisation, monitoring and evaluation. Ultimately, accurate data driven decision making will prevent future cases of AMR and assist those affected by it.

EPCONs implementation framework facilitates an inclusive approach to strengthen local capacity during the implementation of the platform. Depending on the capacity and desired level of engagement, the EPCON and CHAI teams work closely together with regional teams during the implementation of the platform. During this process, knowledge is transferred about the platform architecture and integration levels. This approach ensures a smooth transition towards autonomous capacity to use the platform. EPCON will give access to the product without further commitment to pay maintenance fees in the follow-up years, so that regional teams will be able to continue using the platform. If further maintenance by the EPCON team is required, we will do that at cost. 

On completion of the project we aim to publish the work done in Zimbabwe, in order to share knowledge with the scientific community. 

In 2019, Zimbabwe recorded 3,900 deaths directly attributable to AMR and 15,800 deaths associated with AMR. This makes Zimbabwe to be among the highest 10 countries in age-standardised mortality rate per 100,000 population associated with AMR across 204 countries (https://www.healthdata.org/sites/default/files/2023-09/Zimbabwe.pdf). Zimbabwe has been supported by the Flemming fund to develop and execute a national antimicrobial resistance (AMR) program from 2017 to 2021. This funding facilitated the establishment of essential components to address AMR, including a coordinated approach across multiple sectors, a national action plan, and a monitoring and evaluation strategy. However, the country faces significant challenges in AMR surveillance, particularly in human and animal health sectors, resulting in below-average rankings compared to global peers. The report noted the fragmented availability of various databases within the private and public sector laboratories/pharmacies which are not linked to a central dashboard, making it impossible to have a national view of AMR, preventing timely data-driven decision-making. Developing a digital platform to support linkage of the various fragmented data systems and provide precision public health interventions would augment efforts done by the government and other developmental partners in fighting against the AMR scourge.

Year 1: 

• the clearance of to receive the data at granular geographic level.
• set up the surveillance solution: uploading data, setting up dashboards, and visualising predictions. 

Year 2: 

• surveillance data will be available for the different partners and MOH in the country. 

After year 2: 

• We will share knowledge with other countries/governments where we are active, decrease the costs required for implementation for the same applications once the solution is implemented, and increase capacity for local implementation of the solution through training workshops and strengthening local partnerships.

• Optimise the expenditure of public health resources by empowering local stakeholders with real-time data and data-driven recommendations.

• A scientific publication will allow other researchers to benchmark their approaches, building on existing knowledge, and showcase real-time AMR forecasting.

• Zimbabwe is currently working on the AMR implementation strategy, and CHAI will ensure the ‘Precision for Health for AMR’ is engraved in the AMR country strategy to ensure sustainability and continuity beyond the grant. Country level buy-in involves key stakeholder mapping, stakeholder sensitization meetings, establishing the ‘Precision Public Health for AMR’ taskforce within the Coordination Committee, coordinating protocol design meetings, and getting the final intervention model approved in country in a ‘Design, Validate, Implement’ approach.

EPCON will continuously monitor, analyse, and optimise model performance, apply timely re-training, and model iterations. After a literature search, evidence based covariates will be included in  multivariate as-is burden and time series models, based on a combination of Static and Dynamic Bayesian Network (DBN) modelling. We will predict AMR rates as a whole, by separate antibiotics, and by key disease-antibiotic pairs based on mortality and morbidity rates.

We will apply K-fold cross validation to test the trained models iteratively on held-out neighbourhoods, to measure generalisability of both the as-is burden modelling approach. Following this step, a combined as-is AMR burden model will be trained using all available data.

For the time series model, a combined model will be created for testing against 12 months of unseen data, before training the model on all available data. Error metrics used will include Root Mean Squared Error (RMSE) and Mean Squared Error (MSE), and Credible Intervals (CIs) will be reported. Results will be made publicly available given data owner approval, and the platform results submitted for peer-review and publication.

AMR indicators are evaluated using the theory of change paradigm, to ensure users can link interventions to outcomes so that impact can be measured.

• Cultural challenges

  • Might be challenging to collaborate between different cultures: CHAI has a local team that has a long track-record and is well embedded in the country. 

• Financial challenges

  • Fixed budget but difficult to estimate time needed to complete full project. 

    • We will make sure we do a close follow-up of the budget using our standard tools. 

• Technical Challenges:

  • Data availability and capturing might be a challenge. Together with CHAI, we will foresee time for co-creation workshops in Zimbabwe to make sure we can train a regional team and enable them to scale post the intervention     

  • Dynamic Bayesian modelling may not yield accurate results if the local AMR time series data is chaotic or sparse, necessitating other machine learning approaches, such as reservoir computing, to be used by the platform.

• Legal challenges

  • ethical clearance  will be sort with the relevant bodies which regulate human research including the National Institute for Health Research, Medical Research Council of Zimbabwe and the Research Council of Zimbabwe

The funding will enable us to validate the approach to predict AMR, and to apply a theory of change approach in the platform to link interventions with real-time AMR risk metrics, enabling future research and applications to other diseases using the same framework. It will also enable us to integrate and validate time series prediction to our platform, which will significantly extend its value to end users that rely on accurate data driven estimates of disease risk, by optimizing active case finding, and planning interventions in a timely manner. Further, the basic data layers required to implement a solution in a given country are a main barrier to entry, but the integration of new diseases and partners once the platform is deployed, require significantly less resources to implement and maintain, essentially allowing our platform to be harnessed for future projects and clients that may not afford to  fund the whole process.

EPCON would love to collaborate more intensely with the Bill & Melinda Gates Foundation, USAID, Wellcome trust and other large funding organisations, but also with international NGOs such as ICF, FHI 360. We believe we can bring added value to large pharmaceutical companies active in LMIC such as Roche (pharma and diagnostics), Novartis and other.