Accurately forecasting the case rate of malaria would enable key decision makers to intervene months before the onset of any outbreak, potentially saving lives. Until now, methods that forecast malaria have involved complicated numerical simulations that model transmission through a community. Here we present the first data-driven malaria epidemic early warning system that can predict the 13-week case rate in a primary health facility in Burkina Faso. Using the extraordinarily high-fidelity data of infant consultations taken from the Integrated e-Diagnostic Approach (IeDA) system that has been rolled out throughout Burkina Faso, we train a combination of Gaussian Processes and Random Forest Regressors to estimate the weekly number of malaria cases over a 13 week period. We test our algorithm on historical epidemics and find that for our lowest threshold for an epidemic alert, our algorithm has 30% precision with > 99% recall at raising an alert. This rises to > 99% precision and 5% recall for the high alert threshold. Our two-tailed predictions have an average 1σ and 2σ precision of 5 cases and 30 cases respectively.

No abstract

Objective: Millions of medical consultations are conducted each year in Burkina Faso using the Electronic Register of Consultations (REC). Based on the consultation data collected, we present a method to quantify the quality of individual and ensembles of consultations conducted by frontline healthcare workers (FHWs). Methods: We focus on anthropometric measurements and vital signs (age, weight, height, mid-upper arm circumference and temperature) of children aged between two months and five years old. We compare individual and ensemble of consultations to a multivariate probability distribution defined by an external population-specific, gold standard consultation dataset. By comparing the distributions of consultations to the reference probability distribution, we define a score to rate the quality of measurements and data entry of each FHW. Findings: The defined scores allow us to detect which measurements are most problematic. They also allow us to detect potential biases in the consultation and treatment of different patient groups. No systematic gender-bias was found among FHWs. Height measurements were the most challenging; consultations with the lowest scores were associated with underestimated heights in children. Among these consultations, height was found to be even more underestimated among boys than girls. Conclusion: Our findings enable us to support capacity building of frontline healthcare workers. The REC can be enriched with real-time specific alert on errors, individual FHW can be proposed targeted trainings, and dynamic dashboards can support district managers to navigate the entire population of FHWs and understand which problems should be prioritised.