Comparison of ARIMA and FB-Prophet time series models for the prediction of national and regional malaria incidence in Uganda.

BACKGROUND

Within sub-Saharan Africa, Uganda carries the third largest burden of malaria with 5% of global cases. Due to the stochastic nature of malaria incidence, resource allocation of preventive measures, rapid diagnostic tests, and chemotherapeutics is a significant challenge. To better identify areas at risk and address the challenge of resource allocation, this study aimed to: (1) characterize national and regional malaria incidence in Uganda, and (2) compare the performance of time series models in predicting malaria incidence at national and regional levels.

METHODS

Aggregated data from District Health Information Software 2 (DHIS2), was used to assess national and regional malaria incidence in Uganda from 2020 through 2023. Auto-regressive moving average (ARIMA) models of national and regional malaria incidence were then created. The same data was applied to FB-Prophet, an open source generalized additive time series model. Training and validation datasets were created for each model, which ran for 41 and 6 months, respectively. Model performance was then evaluated via key performance indicators including mean average error (MAE), root mean square error (RMSE), and mean average percentage error (MAPE).

RESULTS

The incidence of malaria within Uganda increased from 200.5 cases per 1000 persons annually in 2021 to 265.4 cases per 1000 persons annually in 2022. The northern regions of West Nile and Acholi, along with Busoga region in the east, experienced the highest burden and incidence of malaria. The mean regional MAE, MAPE, and RMSE was 0.007, 31.2, and 0.01, respectively for ARIMA, and 0.01, 47.8, and 0.01, respectively for FB-Prophet. The ARIMA model outperformed the FB-Prophet model at the national level and in 14 of 15 regions.

CONCLUSIONS

Time series models accurately predicted malaria incidence on a national and regional scale in Uganda. Both the ARIMA and FB-Prophet models have the potential to guide malaria resource allocation and response efforts among other malaria control interventions deployed in Uganda and possibly in other malaria endemic settings.