CEEMDAN-decomposed time series forecasting of reported hepatitis B cases using KOA-optimised deep learning: a nationwide study in mainland China (2004-2027).

Background: Hepatitis B remains a leading notifiable infection in mainland China, with a persistent burden shaped by chronic reservoirs, varied immunity, and shifting surveillance practices. Reliable short- to medium-term forecasts of reported hepatitis B cases are therefore valuable for planning diagnostics, care pathways, antiviral supply, and targeted prevention. Methods: We compiled a national monthly series of hepatitis B notifications from January 2004 to December 2025 and applied Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to isolate multiscale temporal components. Four modelling approaches - gated recurrent units (GRU), convolutional neural networks (CNN), support vector machines (SVM), and a Transformer encoder - were trained on CEEMDAN-derived features using a sliding 12-month window and recursively extended to 24-month horizons. Hyperparameters were optimised via the Kepler Optimization Algorithm (KOA), while performance was assessed through R Results: All models captured dominant trends and seasonality; on the held-out test split, Transformer again delivered the best out-of-sample fit (MAE = 3105.508, MAPE = 0.024, RMSE = 4071.901, and R Conclusions: The findings demonstrate that integrating CEEMDAN with KOA optimised machine learning models delivers highly reliable short-term forecasts. This hybrid framework establishes a resilient continuum from epidemiological surveillance to predictive modelling and public health decision making, thereby providing essential foresight to guide optimal resource allocation and proactive intervention strategies for Hepatitis B control.