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

Area of Science:

• Epidemiology
• Time Series Analysis
• Machine Learning in Public Health

Background:

• Traditional surveillance systems struggle with volatile weekly case data, hindering timely public health interventions.
• Current outbreak threshold methods in the Philippines, like moving averages, are slow and susceptible to extreme values.
• This limits the ability to rapidly respond to epidemiological shifts.

Purpose of the Study:

• To assess hybrid Discrete Wavelet Transform (DWT)-Seasonal Autoregressive Moving Average (SARMA) and DWT-SARMA-Long Short-Term Memory (LSTM) models for forecasting weekly dengue cases.
• To explore the potential of these models in establishing dynamic alarm and epidemic thresholds.
• To compare model performance against traditional thresholding methods.

Main Methods:

• An ecologic time-trend study utilized weekly dengue case data from 2012-2022.
• Data decomposition via DWT, followed by SARMA application to coefficients.
• Enhancement of DWT-SARMA with LSTM applied to residuals for the DWT-SARMA-LSTM model.

Main Results:

• The DWT-SARMA-LSTM model achieved a superior Mean Absolute Percentage Error (MAPE) of 12.4%, outperforming the DWT-SARMA model (MAPE 25.8%).
• The hybrid model effectively captured dengue case peaks and troughs.
• Model-derived thresholds proved more adaptive and context-sensitive than traditional methods.

Conclusions:

• The hybrid DWT-SARMA-LSTM model offers an accurate and robust method for forecasting weekly dengue cases.
• This approach provides a more responsive basis for dengue early warning systems compared to traditional methods.
• The model has practical value for outbreak detection and resource planning, especially in resource-limited settings.