Integration of DHIS2 and Climatic data for disease prediction: A case of malaria occurrence in Gulu District, Uganda

Abstract

Background and Purpose: Prediction models are important tools that can assist public health in evidence based decision making. Since 1950, Uganda has had several prevention and control interventions but malaria is still the leading cause of outpatient visits certainly because disease burden estimates rely on periodic reports of routinely collected and aggregated numbers with occasional estimates of seasonal weather changes. This kind of disease incidence and prevalence monitoring makes it hard to achieve timely and equitable response. Additional efforts to predict disease occurrence incorporating other factors through modelling has the potential to estimate future occurrences. In this research we sought to explore current malaria prevention and control approaches, incorporate weather variables with their time lags to predict future estimates of malaria cases in Gulu District. Methods: A Mixed methods retrospective study utilizing locally collected and reported malaria cases data from an electronic Health Management Information System (eHMIS) powered by DHIS2 from the Ministry of Health and weather data from the Uganda National Meteorological Authority for a span of five years (2013 to 2017). The methods were utilized to explore interventions of control, prevention and treatment by the NMCP and discover the ongoing prediction process and to provide both statistical and machine learning models and visualize the outputs of the prediction models. Using STATA version 14.1, results from a multivariable Negative Binomial Model highlighting a significance of P<0.05 are presented. Machine learning methods were also used including a feed-forward network with sigmoid activation function in the hidden layers and a linear function in the output node using WEKA. Results: An average of 2689 malaria cases were observed. The average Rainfall recorded was at 28.15mm, an average of minimum and maximum temperature was 19.36OC and 30.68OC, Relative humidity at 06 and 12hrs was 75.39 and 51.19 respectively per week. The lagged model was better than the unlagged by 34.18. Weka predictions showed higher discrepancies at the beginning reducing to actual overtime with a difference of only 4.46 and prediction accuracy for the first week was at 99.56%. These results show that the integration of malaria cases data with two weeks’ historical weather data leads to better prediction of future malaria cases compared to reliance on only currently reported cases. Conclusions: DHIS2 data complemented with two weeks’ historical weather data provides better prediction of malaria cases compared to real-time data. Scaling up the data to cover overall national data and more years could greatly improve the accuracy of the models and improve equitable allocation of the scarce prevention, control and treatment resources. Visualization of the predictions provides a quicker and easier alternative to understanding the hidden patterns and simplifies meaning of the numerical values.