Meteorological Drivers of Influenza A and B Positivity in a Subtropical Chinese City: A Six-Year Surveillance Study Integrating Distributed Lag Non-Linear Models and Deep Learning

Influenza transmission in subtropical regions is shaped by complex, non-linear meteorological conditions, while surveillance-based forecasting has been further complicated by fluctuating testing intensity and non-pharmaceutical interventions during the COVID-19 period. Existing approaches often either characterize lagged environmental associations without forecasting capacity or apply deep learning without clear epidemiological interpretability. Using six years (2018-2023) of multi-site influenza surveillance data from Putian, a subtropical coastal city in southeastern China, we developed a two-stage framework to characterize meteorological associations and improve short-horizon forecasting. Influenza positivity rates were used as the primary outcome to mitigate testing-related surveillance bias. Distributed lag non-linear models (DLNMs) were constructed to estimate subtype-specific, lagged associations between meteorological variables and influenza positivity, and a Bayesian-optimized long short-term memory (LSTM) network integrating meteorological, autoregressive, and socio-behavioral covariates was developed to forecast influenza A and B activity. Influenza A positivity was associated with warmer conditions, peaking at 15 (relative risk [RR] = 3.01, 95% CI: 1.03-8.78) and narrow diurnal temperature ranges, whereas influenza B positivity was associated with colder (8; RR = 26.12, 95% CI: 7.79-87.61), more humid (91%; RR = 1.52, 95% CI: 1.00-2.29), and lower-radiation conditions. The LSTM achieved lower forecast error than covariate-matched baselines (multivariate Autoregressive Integrated Moving Average [ARIMA] and eXtreme Gradient Boosting [XGBoost]) for both subtypes (Mean Absolute Error [MAE] for influenza A: 0.009 vs. 0.136 [ARIMA] and 0.138 [XGBoost]; for influenza B: 0.002 vs. 0.049 [ARIMA] and 0.070 [XGBoost]; Symmetric Mean Absolute Percentage Error [SMAPE] for influenza A: 0.521 vs. 1.212 and 1.081; for influenza B: 0.484 vs. 0.810 and 0.737, respectively). Covariate-ablation and SHapley Additive exPlanations (SHAP) analyses suggested that historical positivity and testing-related variables contributed materially to forecast stability under surveillance variability. External evaluation in Sanming provided preliminary evidence of model portability. This integrated DLNM-LSTM framework offers an interpretable approach for characterizing subtype-specific meteorological associations and improving influenza forecasting in subtropical settings.