A hybrid model for tuberculosis forecasting based on empirical mode decomposition in China

Area of Science:

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

Background:

• Pulmonary tuberculosis (TB) remains a significant global health challenge.
• Accurate predictive models are crucial for effective TB prevention and control.
• This study addresses the need for improved TB incidence forecasting.

Purpose of the Study:

• To develop and evaluate advanced models for predicting pulmonary tuberculosis incidence.
• To compare the predictive performance of various time series and machine learning models.
• To identify the optimal model for forecasting TB epidemic trends.

Main Methods:

• Utilized monthly TB incidence data from China (2008-2018).
• Developed and compared ARIMA, LSTM, EMD-SARIMA, EMD-LSTM, and EMD-ARMA-LSTM models.
• Evaluated model performance using Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).

Main Results:

• Decomposition-based models (EMD-SARIMA, EMD-LSTM) outperformed their non-decomposed counterparts.
• The EMD-ARMA-LSTM model demonstrated superior prediction accuracy, reducing error metrics significantly compared to EMD-SARIMA and EMD-LSTM.
• Model performance remained consistent across various prediction horizons (3, 6, and 9 months).

Conclusions:

• Hybrid models integrating decomposition techniques and multiple algorithms enhance predictive accuracy.
• The EMD-ARMA-LSTM model offers superior performance for forecasting pulmonary tuberculosis incidence.
• Findings provide a theoretical foundation for developing targeted TB prevention and control policies.