FaXNet: a frequency-adaptive, explainable, and uncertainty-aware network for influenza forecasting

Background: Accurate and interpretable influenza forecasting is critical for public health preparedness, yet many models struggle to capture multi-scale temporal dynamics and to provide reliable uncertainty estimates. These challenges are particularly pronounced in China, where influenza seasonality differs between northern temperate and southern subtropical regions. Methods: We propose FaXNet, a frequency-adaptive, explainable, and uncertainty-aware deep learning framework that integrates data-driven spectral representation with interpretable component selection and probabilistic forecasting. We compiled weekly influenza positivity rates from the Chinese National Influenza Center and aligned them with ERA5-Land meteorological variables (temperature, dew point, and precipitation) for northern and southern China from 2011 to 2023. FaXNet was evaluated against representative statistical, machine learning, deep learning, and decomposition-based baselines for 1-4-week-ahead forecasting using standard accuracy and calibration metrics. Results: of 0.4493 (north) and 0.4960 (south). The proposed method maintained a statistically significant performance advantage against all benchmarks across varying horizons, validating the effectiveness of frequency-adaptive modeling in mitigating error accumulation. Model explanations highlighted precipitation as the dominant meteorological driver in the north and temperature as the primary factor in the south. Conclusion: FaXNet provides accurate, interpretable forecasts with calibrated prediction intervals across 1-4-week horizons, offering actionable lead time for region-specific risk assessment and resource planning. Performance may depend on surveillance data completeness and the limited set of exogenous covariates considered, motivating future extensions with additional drivers (e.g., mobility or vaccination) and broader external validation.