Daily mortality prediction under compound environmental stressors: Comparing epidemiological time series models and...

Area of Science:

• Environmental epidemiology
• Time series analysis
• Climate change adaptation

Background:

• Accurate daily mortality prediction is essential for public health surveillance and resource allocation, especially with climate change.
• Environmental factors like temperature and air pollution significantly impact mortality rates.

Purpose of the Study:

• To compare epidemiological time series and machine learning models for predicting daily all-cause mortality.
• To identify optimal temperature indices and assess model performance for public health early warning systems.

Main Methods:

• Utilized daily mortality, temperature, and PM2.5 data from Zhejiang Province, China (2009-2020).
• Applied generalized additive mixed models (GAMM) with distributed lag non-linear models (DLNM) to evaluate temperature indices.
• Compared GAMM with time series (ARIMA, Prophet) and machine learning (RF, XGBoost, SVM, MLP, LSTM) models.

Main Results:

• Humidex (HD) was identified as the optimal temperature predictor, emphasizing moisture-driven heat stress.
• Random Forest (RF) showed marginally higher predictive accuracy than GAMM, but the difference was negligible.
• GAMM provided superior mechanistic transparency in quantifying exposure-response relationships.

Conclusions:

• GAMM is a valuable tool for mortality prediction, early warning systems, and informing environmental health interventions.
• The study offers methodological guidance for selecting models that balance forecasting accuracy with mechanistic understanding.
• Effective use of weather and air pollution data supports predictive health surveillance and adaptation to climate change.