Towards Reliable Malaria Forecasting: A Noise-Resilient Wavelet-STL Hybrid Framework for Cameroon

Abstract

Background:

In Cameroon, malaria is a significant public health issue, with heterogenous distribution and seasonal change making control and containment harder to plan. Medium-term projections are needed to provide reliable early warning and optimal prevention resource allocation. Traditional methods, including ARIMA and SARIMA, are dogged by noisy surveillance data as well as structural non-stationarity, and the hybrids currently being applied rarely measure the uncertainty in the forecasts.

Methods:

We introduce a multi-step hybrid forecasting pipeline to combine wavelet-based denoising, robust Seasonal-Trend decomposition with Loess (STL), and state-of-the-art remainder modeling. The remaining component was decomposed by using ARIMA, SARIMA, or Bayesian Structural Time Series (BSTS) and forecasts were recreated out of all components. The analysis was conducted on monthly malaria incidence in the ten administrative regions of Cameroon and 24-month future projections were developed. RMSE, MAE, R², and information criteria were used to evaluate model performance, and uncertainty was measured using analytical intervals (ARIMA/SARIMA) and posterior predictive distributions (BSTS).

Results:

The Wavelet STL preprocessing significantly enhanced model stability, and model accuracy in all regions. ARIMA and SARIMA models performed similarly, and R² values were typically between 0.49 and 0.77, indicating the usefulness of STL in eliminating seasonal effects. In many areas BSTS was significantly higher or as large as ARIMA/SARIMA, and obtained higher R² values. Notably, BSTS offered probabilistic predictions that were calibrated, which allowed to effectively measure the forecast uncertainty. The results presented in this paper indicate that the hybrid pipeline suggested is both noisy and uncertain, and can provide forecasts of malaria cases in the villages of Cameroon over a period of 24 months with high confidence.

Conclusion:

The Wavelet-STL hybrid model is the next step in malaria prediction by combining the denoising, structural decomposition, and probabilistic models. Its deployment in the regions of Cameroon shows innovative methodological value and direct applicability to early warning systems. The method can be easily extended to other infectious diseases with seasonal spread and noisy surveillance data.