Figures and data
Schematic of the integrated analytical framework.
Sentinel surveillance data, daily meteorological records, and contextual covariates (weekly testing volume, weekday/weekend status, proportion of non-local residents, mask-wearing stringency indices) were jointly analysed using two complementary approaches: (i) distributed lag non-linear models (DLNM) to characterise non-linear, lagged exposure–response relationships between meteorological variables and subtype-specific influenza positivity, providing mechanistic interpretability; and (ii) a Bayesian-optimized Long Short-Term Memory (LSTM) neural network for short-term forecasting of daily positivity rates, benchmarked against a seasonal ARIMA model. The primary outcome variable for both analytical streams is the influenza positivity rate (laboratory-confirmed cases ÷ ILI specimens tested), selected to mitigate surveillance bias arising from temporal variation in testing intensity.
Temporal trends and demographic distribution of confirmed influenza cases in Putian city (2018 to 2023).
This figure presents the temporal and demographic characteristics of influenza cases in Putian City over the period from 2018 to 2023. (A) Line plots illustrating the monthly case counts of various influenza subtypes (Influenza A: H1N1, H3N2; Influenza B: Victoria, Yamagata) across the six-year study period. The transition from historical heatmaps to precise line plots enhances the visualization of distinct epidemiological surges, notably capturing the profound interruption of typical seasonal transmission during the peak COVID-19 pandemic restrictions (2020–2021) and the subsequent viral rebound. (B) A population pyramid detailing the age and sex distribution of the 3,155 laboratory-confirmed influenza cases. The visualization bifurcates cases by gender (males on the left, females on the right) across age intervals, underscoring the differential susceptibility of specific demographic cohorts, particularly highlighting the heightened vulnerability of children (0-10 years) and the elderly (60 years and above).
Time-series dynamics of meteorological factors and absolute influenza incidence in Putian city from 2018 to 2023.
This figure vertically aligns temporal fluctuations of influenza-related and key meteorological variables, January 2018-December 2023. From top to bottom: (A) average temperature, (B) humidity, (C) precipitation, (D) solar radiation, (E) daily temperature ranges, (F) ultraviolet index, (G) wind speed, and (H) wind direction, plotted continuously over the five-year surveillance period. In each panel, the left y-axis corresponds to the observed values of the respective meteorological factor (yellow lines), while the right y-axis delineates the absolute number of confirmed cases for Influenza A (red lines) and Influenza B (blue lines). These descriptive plots visually contextualize the raw seasonal correlations and climatic extremes preceding major influenza outbreaks. Note the substantial suppression of influenza activity during 2020–2022 (pandemic intervention period) and the pronounced post-restriction rebound during 2023, particularly for influenza A, which exceeded pre-pandemic peak levels.
Cumulative risk of meteorological factors on Influenza A and B positivity rates over a 15-day Lag period.
This figure illustrates the non-linear, cumulative exposure-response relationships between eight individual meteorological drivers and the 15-day cumulative relative risk of influenza A and B positivity, estimated via Distributed Lag Non-linear Models (DLNM). Models were adjusted for long-term trends, weekly detection volumes, and other fixed covariates. Panels in (A) depict the non-linear cumulative risks for Influenza A, while panels in (B) shows the corresponding risks for Influenza B. In each subplot, the x-axis represents the observed continuous range of a specific meteorological factor, and the y-axis indicates the cumulative relative risk compared to a predefined median reference value (e.g., 24.1°C for average temperature, 75.8% for humidity). The solid line reflects the estimated non-linear trend of risk variation, and the shaded regions denote the 95% confidence intervals (CIs).
Lagged temporal effects of extreme meteorological conditions on Influenza A and B positivity rates.
This figure visualizes the distribution of relative risks over a 15-day lag period following exposure to extreme meteorological conditions, derived from Distributed Lag Non-linear Model (DLNM). The analysis compares the temporal lag effects at the extreme high (95th percentile) and extreme low (5th percentile) values of selected weather variables for Influenza A (A) and Influenza B (B). In each subplot, the x-axis represents the specific lag days (from day 0 to 15), and the y-axis shows the relative risk compared to the median reference condition. Solid orange lines (for Influenza A) and yellow lines (for Influenza B) represent extreme high-value conditions (e.g., extreme heat), whereas solid blue lines (for Influenza A) and green lines (for Influenza B) represent extreme low-value conditions (e.g., extreme cold). The shaded areas surrounding the curves represent the 95% confidence intervals. These lag-response curves elucidate the dynamic delay between specific extreme weather events and the subsequent peaks in influenza test positivity.
Optimal LSTM hyperparameters for predicting influenza A and B positivity rates.
High-fidelity forecasting of the Influenza Positive Rate using the LSTM network and SHAP interpretability.
This comprehensive panel visualizes the predictive performance and feature importance of the LSTM model configured to forecast the influenza positivity rate. (C) and (D) Training loss curves: These panels show the variation in loss values during the training of the LSTM network. The x-axis represents the number of training epochs, while the y-axis indicates the loss values. The smooth decline and stabilization of the loss curves indicate effective learning by the network, with no significant signs of overfitting. This is a crucial measure of the network’s stability and effectiveness throughout the training process. (A) and (B) Internal Out-of-Time validation (2023) comparing the LSTM-predicted positive rate (blue lines) against the observed positive rate (red lines) for Influenza A and B. The x-axis represents the date, and the y-axis reflects the positivity rates. This comparison visually demonstrates the network adeptly captures the non-linear viral rebound dynamics. (E) and (F) SHAP (SHapley Additive exPlanations) feature importance rankings. The bar charts quantify the mean absolute SHAP values, identifying historical subtype positivity rates (typeA_rate / typeB_rate), testing volume (weekly_detection), average temperature (temp), and mask-wearing stringency indices (facial_coverings) as paramount predictive drivers. The x-axis denotes the mean SHAP value, while the y-axis lists the variables. (G) and (H) SHAP summary plots illustrating the distribution of impacts for each feature on the model’s output, where color indicates feature value (red = high, blue = low). This graph provides a visual understanding of how and to what extent different features influence the influenza predictions, offering critical insights into the network’s decision-making process. (I) and (J) External spatial validation utilizing 2023 surveillance data from the neighboring city of Sanming, confirming the LSTM’s robust generalizability in forecasting independent epidemic curves. The date is plotted on the x-axis and the positivity rates are plotted on the y-axis.
