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

Accurate prediction of daily mortality is vital for effective early warning systems and the strategic allocation of medical resources, particularly given the rising public health challenges posed by climate change. This study presents a comparative analysis of epidemiological time series models and machine learning (ML) models for predicting daily all-cause mortality across 89 districts in Zhejiang Province, China (2009-2020), utilizing temperature and fine particulate matter (PM_2.5) data. The study first evaluated the predictive performance of five temperature indices-Daily Mean Temperature (T), Apparent Temperature (AT), Humidex (HD), Discomfort Index (DI), and Heat Index (HI)-by applying a generalized additive mixed model (GAMM) with a distributed lag non-linear model (DLNM). The HD was identified as the optimal predictor, highlighting the critical role of moisture-driven heat stress in subtropical monsoon climates. Subsequently, the performance of the GAMM was compared with that of time series models (Autoregressive Integrated Moving Average [ARIMA], Prophet) and ML models (eXtreme Gradient Boosting [XGBoost], Random Forest [RF], Support Vector Machine [SVM], Multi-Layer Perceptron [MLP], and Long Short-Term Memory [LSTM]). While the RF model showed slightly higher accuracy (R² = 0.672 full year; R² = 0.674 cold season; R² = 0.641 warm season) than the GAMM (R² = 0.643 full year; R² = 0.645 cold season; R² = 0.626 warm season), this difference was negligible given the stochastic variability inherent in mortality time series data. Crucially, the GAMM provides essential mechanistic transparency in quantifying lagged exposure-response associations. Through explicit assessment of the mortality effects of environmental stressors, the GAMM serves as a superior tool for early warning systems and for guiding environmental health interventions and climate adaptation policies. These findings also provide crucial methodological guidance for selecting models that not only enable real-time mortality forecasting using weather and air pollution data, but also effectively support the early detection of abnormal mortality events in predictive health surveillance.