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

Respiratory syncytial virus (RSV) is a major cause of lower respiratory tract infection and hospitalization in infants, resulting in a substantial global burden on health care services. Accurate and timely forecasting is essential for the effective control and prevention of RSV infections, contributing to mitigating the substantial global disease burden. Spatiotemporal forecasting techniques have demonstrated the potential to improve predictive performance for RSV infections across multiple regions. In this study, we applied and refined the spatiotemporal graph convolutional networks (ST-GCN) model, which can extract both spatial and temporal patterns and has been widely applied in spatiotemporal forecasts in many fields. We assessed the performance of ST-GCN-based models in predicting the spatiotemporal dynamics of RSV transmission in Japan. Our results demonstrated that ST-GCN-based models outperformed baselines that do not account for spatial information, including linear regression, autoregressive integrated moving average (ARIMA), long short-term memory (LSTM) networks, and transformer-based models, by 11-39% (R²). Additionally, our results indicated that integrating a temporal-gated LSTM layer further improved the performance of ST-GCN by effectively capturing long-term temporal dependencies. These findings highlight the promise of spatiotemporal techniques in providing accurate and timely forecasts of RSV transmission.Clinical Relevance- The graph-neural-network based models can accurately forecast RSV infections for short-term horizons and be used to support the prevention or treatment of RSV infections, especially in optimizing the administration time for RSV prophylaxis plans to prevent severe RSV disease.