OBJECTIVES: Respiratory syncytial virus (RSV) is one of the most dangerous respiratory diseases for infants and young children. Due to the nonpharmaceutical intervention (NPI) imposed in the COVID-19 outbreak, the seasonal transmission pattern of RSV has been discontinued in 2020 and then shifted months ahead in 2021 in the northern hemisphere. It is critical to understand how COVID-19 impacts RSV and build predictive algorithms to forecast the timing and intensity of RSV reemergence in post-COVID-19 seasons.

METHODS: We propose a deep coupled tensor factorization machine, dubbed as DeCom, for post COVID-19 RSV prediction. DeCom leverages tensor factorization and residual modeling. It enables us to learn the disrupted RSV transmission reliably under COVID-19 by taking both the regular seasonal RSV transmission pattern and the NPI into consideration.

RESULTS: The dataset used in the experiments is a US state-level RSV dataset that contains weekly RSV cases for 52 US states, including Puerto Rico and Washington, D.C., between 2015 to 2022. We compare the predictive performance of DeCom with deep tensor factorization model (DeTensor), LSTM, ARIMA, and Seasonal ARIMA (SARIMA). Experimental results on a real RSV dataset show that DeCom is more accurate than the state-of-the-art RSV prediction algorithms and achieves up to 46% lower root mean square error (RMSE) and 49% lower mean absolute error (MAE) for country-level prediction compared to the baselines.

CONCLUSIONS: In this research, we introduce DeCom, a deep coupled tensor factorization machine that captures both the seasonality of RSV and the influence of COVID-19 NPIs on RSV transmission. Computer results showcase that DeCom outperforms several time-series prediction baselines in predicting the timing and magnitude of RSV outbreaks under the influence of COVID-19. As a consequence, we believe that DeCom can be an effective model to forecast and assess post COVID-19 RSV outbreaks.