Infectious disease forecasting to support public health: use of readily available methods to predict malaria and diarrhoeal diseases in Mozambique

Abstract

Background: Mozambique faces a high burden of infectious diseases but currently has limited capacity for forecasting disease incidence. Recent improvements in disease surveillance through the National Monitoring and Evaluation System now provide weekly reports of disease incidence across the country's districts. This study focuses on using these records, specifically for malaria and diarrhoeal diseases, which together account for approximately 40% of deaths among children under five, to develop statistical forecasts and evaluate their accuracy.

Methods: We utilised a Python library for time series forecasting called Darts, which includes a variety of statistical forecasting models. Three models were selected for this analysis: Exponential Smoothing (a classical statistical model), Light Gradient Boosting Machine (a machine-learning model), and Neural Hierarchical Interpolation for Time Series (a neural network-based model). Retrospective forecasts were generated and compared across multiple forecast horizons. We evaluated both point and probabilistic forecast accuracy for individual models and two types of model ensembles, comparing the results to forecasts based on historical expectance.

Results: All models consistently outperformed forecasts based on historical expectance for both malaria and diarrhoeal disease across forecast horizons of up to eight weeks, with comparable or better performance at 16 weeks. The most accurate forecasts were achieved using a weighted ensemble of the models.

Conclusions: This study highlights the potential of using a readily available tool for generating accurate disease forecasts. It represents a step toward scalable and accessible forecasting solutions that can enhance disease surveillance and public health responses, not only in Mozambique but also in other low- and middle-income countries with similar challenges.

Figure 1

Forecast trajectories obtained through the rolling-origin evaluation procedure for the case of malaria incidence at Sofala province with a four-week forecasting horizon. Each subplot presents the forecasts of one of the tested models (Exponential Smoothing, LightGBM and N-HiTS) or model ensembles (mean ensemble and WIS-weighted ensemble). In each subplot the black line represents the data, coloured line shows the mean predictions, and the darker and lighter coloured regions show the 50% and 95% prediction intervals respectively. Predictions start from July 2018. The legend of each subplot contains prediction scores using the different metrics (RMSE, MAPE, SMAPE, WIS). Similar figures for all tested cases (malaria and diarrhoeal diseases incidence forecasts for all provinces and all forecast horizons) can be found online at: https://github.com/ramiyaari/Forecasting_Malaria_Diarrheal_Diseases_In_Mozambique/tree/main/figures/230624. LightGBM – Light Gradient Boosting Machine, MAPE – mean absolute percentage error, N-HiTS – Neural Hierarchical Interpolation for Time Series, RMSE – root mean squared error, SMAPE – symmetric mean absolute percentage error, WIS – weighted interval score.

Figure 2

Summary of the forecasting models accuracy over all the tested cases using the different evaluation metrics. The variance depicted by the boxplots corresponds to the different provinces for which forecasts were made. Note that the WIS metric is not shown for the historical expectance forecasts as probabilistic forecasts were not derived for this model. WIS – weighted interval score.

Figure 3

Forecast trajectories and evaluation of the WIS-weighted ensemble. Panel A. Incidence rates per province for malaria and diarrhoeal diseases (black) together with forecast trajectories of the WIS-weighted ensemble for the various forecast horizons (blue: two weeks, green: four weeks, magenta: eight weeks, orange: 16 weeks). Panel B. The calculated MAPE value for the forecast trajectories in Panel A. MAPE – mean absolute percentage error, WIS – weighted interval score.