CEEMDAN-decomposed time series forecasting of reported hepatitis B cases using KOA-optimised deep learning: a nationwide study in mainland China (2004-2027)
Synopsis
This study compiled a national monthly series of hepatitis B notifications in mainland China from January 2004 to December 2025, applied CEEMDAN to isolate multiscale temporal components, and trained four models (GRU, CNN, SVM, and a Transformer encoder) with KOA-optimised hyperparameters on a sliding 12-month window recursively extended to 24-month horizons; the Transformer delivered the best out-of-sample fit on the held-out test split (MAE = 3105.508, MAPE = 0.024, RMSE = 4071.901, R² = 0.928), SVM ranked second, CNN outperformed GRU but not SVM, and forecasts for 2026-2027 remain elevated, signalling little improvement and even possible resurgence.
Interpretation
It combines CEEMDAN multiscale decomposition with KOA hyperparameter optimisation for short- to medium-term forecasting of the national monthly hepatitis B notification series. Relative to single-model or undecomposed time series modelling, the framework first isolates multiscale temporal components and then trains four model families with unified KOA tuning. Based on the national monthly series from 2004 to 2025, using a sliding 12-month window and recursive 24-month extension, evaluated across training, validation, and test splits.
On the held-out test split, the Transformer encoder achieved the best out-of-sample fit, SVM ranked second, and CNN performed better than GRU but worse than SVM. It provides a comparable ranking of four model families under the same decomposed features and the same optimisation procedure, rather than reporting a single model result. Transformer: MAE = 3105.508, MAPE = 0.024, RMSE = 4071.901, R² = 0.928; SVM's MAE, MAPE, and RMSE were 11.317%, 11.615%, and 9.964% higher than the Transformer's; CNN reduced MAE, MAPE, and RMSE by 21.985%, 22.178%, and 14.944% relative to GRU, with a 6.158% larger R².
Forecasts for 2026-2027 show reported cases remaining elevated, signalling little improvement and even possible resurgence. It extends the modelling output two years forward, offering forward-looking reference for planning diagnostics, care pathways, antiviral supply, and targeted prevention. Based on a recursively extended 24-month forecast horizon; these are model outputs rather than observed data.
Perspective
The results apply to short- to medium-term forecasting of national monthly reported hepatitis B cases in mainland China, serving resource arrangements such as diagnostics planning, care pathways, antiviral supply, and targeted prevention; the framework's approach can also serve as a reference for surveillance forecasting of other notifiable infections.
The 2026-2027 forecasts are recursive extrapolations, and their elevated trajectory depends on whether reporting practices, chronic reservoirs, and immunity contexts continue as assumed; readers may watch whether subsequent observations align with the forecast range and whether the model ranking remains stable under different decomposition and optimisation settings.