A dual-branch deep learning framework for tiered early warning of COVID-19 utilizing wastewater data

ABSTRACT Background: Wastewater offers earlier, population-level signals, yet few models integrate environmental drivers for reliable routine COVID-19 alerts. We hypothesized that combining wastewater and environmental covariates in a dual-branch deep model leveraging FFT would improve forecasting and alerting. Methods: Using weekly wastewater, meteorological, and case data from Changzhou, China (Jan 29–Dec 10, 2024), we developed a framework that forecasts case trajectories and triggers tiered yellow/red alerts at predefined thresholds. Results: On 2-week-ahead internal tests, performance was: RMSE 1.40 (1.13–1.67), MAE 1.23 (0.99–1.48), MAPE 10.44% (5.20–16.40), and R2 0.99 (0.99–0.99). On an external test, both yellow and red alerts were correctly predicted 3 weeks ahead. Versus a naive baseline and the COVID-19 Forecast Hub model, our approach reduced missed alerts, whereas the Forecast Hub model reduced false declarations. Ablation showed necessity of the dual-branch architecture and covariates. Conclusions: The novel framework delivers accurate, timely forecasts and reliable early warnings from multi-source data, supporting proactive public health response to COVID-19. It may also be a promising approach for the prediction of other infectious diseases. However, validating and adapting the approach across locations, and epidemic patterns is a key next step to establish robustness, generalizability, and operational value.