Graph-Neural-Network Based Models for Improved Short-Term Forecasts of Respiratory Syncytial Virus Infections.

Area of Science:

• Epidemiology
• Infectious disease modeling
• Machine learning for public health

Background:

• Respiratory syncytial virus (RSV) poses a significant global health burden, particularly for infants, causing lower respiratory tract infections and hospitalizations.
• Effective control and prevention strategies for RSV rely heavily on accurate and timely forecasting of infection dynamics.
• Spatiotemporal forecasting methods offer a promising approach to enhance the predictive accuracy of infectious disease outbreaks across diverse geographical regions.

Purpose of the Study:

• To apply and refine the spatiotemporal graph convolutional networks (ST-GCN) model for predicting the spatiotemporal transmission patterns of RSV in Japan.
• To evaluate the performance of ST-GCN-based models against traditional forecasting methods and models lacking spatial considerations.
• To investigate the impact of integrating a temporal-gated LSTM layer into the ST-GCN model for improved temporal dependency capture.

Main Methods:

• Utilized spatiotemporal graph convolutional networks (ST-GCN), a deep learning model adept at extracting complex spatial and temporal patterns.
• Compared the predictive performance of ST-GCN models against baseline methods including linear regression, ARIMA, LSTM, and transformer-based models.
• Assessed model performance using R-squared (R²) values to quantify prediction accuracy for RSV transmission dynamics in Japan.

Main Results:

• ST-GCN-based models demonstrated superior performance, outperforming non-spatial baseline models by 11-39% (R²).
• The integration of a temporal-gated LSTM layer further enhanced ST-GCN performance by improving the capture of long-term temporal dependencies in RSV transmission.
• The refined ST-GCN models proved effective in predicting the spatiotemporal dynamics of RSV infections.

Conclusions:

• Spatiotemporal forecasting techniques, particularly ST-GCN, hold significant promise for accurate and timely prediction of RSV transmission.
• Graph-neural-network based models can provide reliable short-term forecasts, aiding in the optimization of RSV prevention and treatment strategies, such as prophylaxis administration.
• These advanced modeling approaches can contribute to mitigating the global burden of RSV disease through improved public health interventions.