Integrating distributed lag non-linear models and deep learning reveals meteorological drivers of influenza A and B positivity in a 6-year subtropical surveillance study
Subtropical influenza forecasting is complicated by non-linear meteorological dynamics and surveillance bias stemming from pandemic-era testing fluctuations. Existing approaches either characterize lagged environmental associations without forecasting capacity or apply deep learning without epidemiological interpretability. We analyzed multisite influenza surveillance data from Putian,…
In a study spanning six years in subtropical Putian, southeastern China, researchers explored the meteorological drivers of influenza A and B positivity using distributed lag non-linear models and deep learning techniques. Existing methods either focused on lagged environmental associations without forecasting capabilities or applied deep learning without offering interpretability.
The study analyzed influenza surveillance data from multiple sites, using test positivity as the primary outcome to avoid bias from testing fluctuations. Distributed lag nonlinear models identified type-specific associations between eight meteorological variables and positivity, with a window of 15 days. Simultaneously, a Bayesian-optimized long short-term memory (LSTM) network, which included meteorological, autoregressive, and socio-behavioral covariates, was compared to multivariate autoregressive integrated moving average (ARIMA) and extreme gradient boosting (XGBoost) models.
For influenza A, the research found that a temperature of 15°C (relative risk = 3.01, 95% confidence interval: 1.03–8.78) led to peak relative risk, while narrow diurnal temperature ranges influenced positivity. Conversely, for influenza B, colder conditions at 8°C resulted in higher relative risk (RR = 26.12, 95% CI: 7.79–87.61), along with higher humidity (91% relative risk = 1.52, 95% CI: 1.00–2.29) and lower solar radiation.
The LSTM model demonstrated superior performance, with lower mean absolute error (0.009 for influenza A, 0.002 for influenza B) compared to ARIMA (0.136) and XGBoost (0.138) models. Additionally, the LSTM exhibited reduced symmetric mean absolute percentage error (0.521 for influenza A, 0.484 for influenza B), outperforming the other models (0.810 for influenza A, 0.737 for influenza B).
Furthermore, covariate ablation and Shapley additive explanations analyses revealed that historical positivity and testing-related variables contributed to forecast stability. External validation conducted in Sanming provided preliminary evidence that the framework could be applicable across different regions. Overall, this integrated approach combines interpretable meteorological associations with short-horizon forecasting, accommodating variable surveillance conditions.
Rewritten by urgent.news; may contain errors — check the original.