Stage-aware deep learning with adversarial data synthesis enables prospective season-scale A/H1N1 influenza...

Accurate influenza forecasting is essential for public health preparedness, yet many models require future covariates, provide limited interpretability, and degrade under post-pandemic regime shifts. We propose a Stage-Aware Multimodal Neural Network (SAMNN) that integrates multimodal temporal features using configurations adapted to heterogeneous transmission regimes and supports season-ahead scenario forecasting using adversarially synthesized covariates. Using 13 years of national A/H1N1 surveillance data from mainland China (2011∼2024), SAMNN was evaluated across three epidemiologically distinct seasons spanning ∼20-fold differences in peak intensity. SAMNN achieved R ² values of 0.82-0.95, 0.82-0.97, and 0.55-0.97 for 1-week-, 2-week-, and 4-week-ahead forecasting, respectively, and generally maintained competitive performance relative to five baseline models across forecast horizons. SHAP attribution showed that epidemiological signals dominated predictions, with context-dependent contributions from climatic and social-context features. To support prospective scenario forecasting, we used an adversarial synthetic covariate-generation pipeline to produce season-ahead forecasts for 2024/25 using information available at the prespecified forecast origin.