Advancing outbreak detection: Hybridizing machine learning with wavelets for weekly dengue case forecasting

Traditional surveillance systems often struggle with the volatility of weekly case data, limiting timely prevention and control efforts. In the Philippines, the standard method for setting outbreak thresholds relies on historical moving averages that are highly affected by extreme values and slow to reflect recent epidemiological shifts. This study assessed the performance of hybrid discrete wavelet transform (DWT)-seasonal autoregressive moving average (SARMA) and DWT-SARMA-long short-term memory (LSTM) models in forecasting weekly case counts and explored their potential in defining dynamic alarm and epidemic thresholds in Quezon City, Philippines.

An ecologic time-trend study was conducted using weekly dengue case counts from 2012 to 2022. The data was decomposed using DWT, and SARMA was applied to the resulting approximate and detail coefficients. The DWT-SARMA model was enhanced by applying an LSTM to the SARMA residuals. The DWT-SARMA-LSTM model demonstrated superior performance, achieving a Mean Absolute Percentage Error (MAPE) of 12.4%, and successfully captured the case peaks and troughs. In contrast, the DWT-SARMA model produced a MAPE of 25.8%. Model-derived thresholds were more adaptive and context-sensitive than the traditional 3-year moving mean threshold, which was skewed by pre-pandemic data.

The hybrid DWT-SARMA-LSTM model is an accurate and robust approach for forecasting weekly dengue cases. It provides a more responsive basis for an early warning system than traditional thresholding methods, and has practical value for outbreak detection and resource planning in dynamic public health environments, particularly in resource-limited settings where timely and accurate data are critical.