Objective Mortality surveillance and early warning in population based on death cause surveillance data has great public health significance. This study explores spatiotemporal modeling for predicting daily all-cause deaths at the county level using death cause surveillance data, aiming to provide methodological support for early warning models and new insights for active surveillance systems.

Methods This study used daily, county-level and all-cause deaths data in Jiangsu province during 2015-2019, which were obtained from the National Mortality Surveillance System. Based on this dataset, a hybrid deep learning model that integrates a Bidirectional Long Short-Term Memory network (BiLSTM) with a Spatio-Temporal Graph Convolutional Network (STGCN) was developed to predict daily all-cause deaths. Data from 92 counties were incorporated into the model. Meanwhile, this model was compared with a Bayesian hierarchical spatiotemporal model and a STGCN model, with performance assessed via root mean squared error (RMSE), mean absolute error (MAE), and symmetric mean absolute percentage error (SMAPE). The models were developed by using R 4.5.1 and Python 3.10.

Results The BiLSTM-STGCN model demonstrated a RMSE value of 4.81, an MAE value of 3.64, and an SMAPE value of 27.42% in prediction of daily all-cause deaths across all counties in Jiangsu from July 1, 2019, to December 31, 2019. The model exhibited good prediction performance and outperformed the Bayesian hierarchical spatiotemporal model and STGCN model.

Conclusion The BiLSTM-STGCN model exhibits strong prediction performance and, compared with the classical Bayesian hierarchical spatiotemporal model and single STGCN model, is more suitable for the prediction daily deaths at the county level based on death cause surveillance data.