The data for this study were obtained from the Chinese Influenza Surveillance Information System, specifically the influenza surveillance data for Lanzhou from January 1, 2014, to December 31, 2024. This dataset includes daily influenza virus test volumes, detection rates, and predominant circulating subtypes (Influenza viruses are classified according to the WHO naming conventions: influenza A(H1N1)pdm09 and A(H3N2), and influenza B Victoria lineage. For brevity, these viruses are hereinafter referred to as H1N1, H3N2, and B/Victoria, respectively) in Lanzhou City. The positivity rate for each influenza subtype was calculated using the following formula:

To analyze the impact of meteorological factors on the transmission of influenza subtypes, we divided the study period into three distinct phases based on the timeline of the COVID-19 pandemic in Lanzhou, China.

Meteorological data are publicly accessible after registration on the official website of the National Meteorological Science Data Center (data.cma.cn/). The meteorological factors included the weekly average temperature (°C), weekly average temperature difference (°C), weekly average air pressure (hPa), weekly average precipitation (mm), weekly average relative humidity (%), weekly average sunshine duration (h), and weekly average wind speed (m/s).

To systematically evaluate the epidemiological patterns of influenza viruses from 2014 to 2024, particularly the impact of NPIs, we conducted multidimensional statistical analyses encompassing descriptive, seasonal visualization, nonlinear, and interactive approaches. The study period was divided into three phases for intergroup comparisons: pre-COVID (2014–2019), during the COVID-19 pandemic (2020–2022), and post-COVID era (2023–2024). Both influenza surveillance and meteorological data were integrated into weekly time series for subsequent analysis. Weekly weather data were matched to the positivity rates of each influenza subtype. Before modeling, all time series underwent missing data checks. Isolated or consecutive missing values ≤2 weeks were imputed using linear interpolation to ensure consistency across datasets.

Data processing and statistical analyses were performed using the R4.5.3 software. Normality was assessed using the Shapiro-Wilk test, and homogeneity of variance was evaluated using Levene’s test. When the data were normally distributed with equal variances, ANOVA was applied; when normally distributed but with unequal variances, Welch’s ANOVA was used. The Kruskal-Wallis test was used for continuous variables that did not meet the normality assumption. Post-hoc pairwise comparisons were performed using the Wilcoxon signed-rank test for normally distributed data and the Mann-Whitney U test for non-normally distributed data. All pairwise comparisons were performed using the Bonferroni correction for multiple comparisons. Categorical variables were compared between groups using the chi-square or Fisher’s exact test. Statistical significance was set at P < 0.05. Figures were generated using GraphPad Prism software (version 8.0).

To quantify the independent contribution of meteorological factors to influenza positivity rates and analyze their spatiotemporal heterogeneity, this study employed an interpretable machine learning framework based on XGBoost and SHAP. The XGBoost algorithm excels at handling structured data and time series, demonstrating particular strengths in managing missing data and capturing complex nonlinear interactions among predictor variables (Zhang et al., 2025). Prediction modeling with XGBoost: The model employs cross-validation and early stopping to select the optimal iteration round, reporting R² and RMSE for regression scenarios. SHAP values for each observation were computed using the TreeExplainer algorithm of XGBoost, which assigns marginal contributions to model predictions based on the Shapley value theory from game theory (Wang et al., 2021). SHAP values are interpreted as follows: SHAP value > 0 indicates that the feature increases the predicted influenza positivity rate; SHAP value < 0 indicates that the feature decreases the predicted positivity rate; the absolute magnitude of the SHAP value reflects the strength of the feature’s contribution. To assess the impact of the COVID-19 pandemic on the meteorological-influenza association, independent XGBoost-SHAP models were constructed for three periods: Pre-COVID-19 (2014–2019), COVID-19 (2020–2022), and Post-COVID-19 (2023–2024). The relative importance and directional changes in meteorological factors across these periods were compared. The SHAP summary plot visualizes the global feature contribution distribution for each meteorological factor, where the point color indicates the feature value magnitude (red = high, blue = low), and the x-axis represents the SHAP values (impact direction and strength).

Our XGBoost model demonstrated relatively robust explanatory power during the early stages of the pandemic, but limited explanatory power during the pandemic and post-pandemic periods. The R² value exceeded 0.15 across all three periods, with the RMSE fluctuating between 0.1350 and 0.1600. Details are provided in Supplementary Table S1. Given this, we adopted a conservative approach in interpreting the model’s SHAP values, primarily using them to identify the relative importance and direction of influence of variables, rather than treating the absolute contribution rates reported within the model as direct evidence of the magnitude of causal effects.

Before the COVID-19 outbreak, influenza virus circulation in Lanzhou, China, exhibited typical seasonal patterns of the northern hemisphere, with annual peaks occurring from winter to the following spring. These outbreaks are usually dominated by a single subtype, despite the coexistence of multiple subtypes. The data revealed distinct antigenic drift patterns among the subtypes. A/H3N2 dominated in most years, whereas B/Victoria exhibited a relatively stable circulation intensity. Notably, A/H1N1 exhibited intermittent outbreak characteristics. However, during 2020-2021, all subtypes experienced epidemiological silence, consistent with the suppression of influenza transmission by non-pharmaceutical interventions against COVID-19. Subsequently, only the B/Victoria subtype was circulating during the 2021–2022 season. In the post-pandemic era, as national control measures were lifted and epidemic management gradually normalized, the 2023–2024 influenza season exhibited a typical, high-intensity seasonal peak occurring during the winter months from late 2023 to early 2024. The peak magnitude of positive cases indicated a significant rebound in overall influenza activity levels, returning to the pre-COVID-19 pandemic typical intensity and breaking free from the abnormal suppression observed between 2020 and 2022. The A/H3N2 subtype once again became the predominant circulating subtype, constituting the largest proportion and accounting for the vast majority of positive cases during the season. The influenza A(H1N1)pdm09 and influenza B Victoria subtypes were also detected, but their activity levels remained relatively low, maintaining a co-circulating status. As shown in Figure 1.

The results indicated that during the pre-COVID period, influenza positivity rates remained elevated, with a median of approximately 15%, reflecting typical seasonal influenza transmission patterns. During the COVID-19 period, the influenza positivity rate significantly declined to approximately 5%, indicating that NPIs exerted a strong suppressing effect on influenza transmission. In the post-COVID period, the influenza positivity rate rebounded to approximately 10%, as shown in Figure 2A. This suggests a gradual resumption of influenza virus transmission. Statistical analysis revealed highly significant differences between the periods (p < 0.001), as shown in Supplementary Table S2. Further analysis of the pre-pandemic data revealed seasonal patterns. The positivity rate exhibited a classic winter peak, with the median winter positivity rate (23.9%) significantly higher than that in spring (4.2%), summer, and autumn (0%) (p = 0.000172), confirming the strong seasonality of influenza. As shown in Figure 2B. The time-series trend clearly revealed regular winter peaks from 2014 to 2019 (Figure 2C), a sharp decline and near disappearance of epidemic intensity from 2020 to 2022, and the re-emergence of epidemic peaks in 2023–2024. The smoothing curve overlaid on the line graph highlights this overall “disruption-recovery” trend. To further investigate whether seasonal patterns changed across different periods, we plotted the phased seasonal boxplots. As shown in Figure 2D. The results indicate that interphase differences in winter positivity rates were highly statistically significant (p < 0.001), while seasonal patterns in spring and summer showed no significant changes across phases (p > 0.05), as detailed in Supplementary Table S3. This suggests that NPIs primarily influence the intensity of influenza activity during the traditional winter season rather than its fundamental transmission patterns throughout the year.

Before, during, and after the COVID-19 pandemic, we observed significant differences in the contributions of environmental factors across the different phases of the outbreak, with changes in multiple factors being particularly pronounced. To illustrate this, we constructed grouped-level bar charts showing the relative percentage contributions of each factor during the pre-pandemic, pandemic, and post-pandemic periods. As shown in Figure 3, the findings reveal a shift from a multi-factor-driven pattern in the pre-pandemic phase to overall suppression during the pandemic phase, followed by selective reconstruction in the post-pandemic phase. The succession and stability of the dominant factors suggest that temperature tended to remain one of the key contributors across all three periods. Its contribution was comparatively high in the pre-pandemic phase, while its relative contribution rate increased to nearly 40% in the post-pandemic phase, indicating a more prominent role among the environmental variables.Atmospheric pressure exhibited robust secondary dominance across periods, indicating a stable underlying association with influenza transmission that remains largely unaffected by intense social interventions. Among the other meteorological factors, relative humidity, sunshine duration, and temperature difference demonstrated moderate explanatory power in the pre-pandemic phase. However, their contributions significantly diminished during the pandemic, reflecting the marginalization of natural environmental drivers under high-intensity social interventions. This relative contribution analysis strongly demonstrates that the COVID-19 pandemic, as a global natural experiment, profoundly reshaped the strength of the epidemiological associations between environmental factors and influenza transmission (Supplementary Table S4).

To comprehensively reveal the dynamic changes in environmental factors influencing influenza transmission, we employed an XGBoost model combined with SHAP analysis. The results indicate that the pre-epidemic period was dominated by environmental drivers, exhibiting a classic environmental-driven pattern. The wide distribution range of the SHAP values across features suggests the strong explanatory power of environmental factors for influenza transmission. Temperature was the most influential predictor. Its point cloud distribution was the broadest, with SHAP values ranging from approximately -0.05 to 0.22, revealing a clear dose-response relationship. Low temperatures (blue points) were densely clustered in regions with high positive SHAP values, indicating that cold temperatures were a primary driver of influenza epidemics. Conversely, high temperatures (red points) predominantly corresponded to negative SHAP values, indicating an inhibitory effect. Sunshine Duration exhibited a synergistic effect with temperature. Shorter sunshine duration (blue points) correlated with an increased positive risk, consistent with shorter daylight hours in winter and increased indoor activity. The effects of Relative Humidity and Atmospheric Pressure are relatively complex, with scattered point cloud distributions indicating potentially nonlinear relationships with positivity rates or modulation by other factors. Precipitation, Temperature Range, and Wind Speed exhibited the weakest influence during this period, with SHAP values clustered around zero. See Figure 4A.

The COVID-19 pandemic was dominated by social interventions. The most notable feature of this phase was the dramatic narrowing of the distribution range of the SHAP values for all environmental factors, with values highly clustered around zero. Although the ranking of each feature of influence remained unchanged, its absolute impact significantly diminished. Taking the most important feature, “temperature,” as an example, its SHAP value range contracted from approximately 0.27 units to less than 0.1 units, as shown in Figures 4A, B. The color distribution pattern in the point cloud shifted from a blue-to-red gradient associated with temperature and became blurred, indicating that the previously strong correlation between environmental factors and the positivity rate was severely weakened. This pattern provides compelling evidence that under robust non-pharmaceutical interventions (e.g., mask mandates, social distancing, and travel restrictions), social behavioral factors completely overshadow the natural driving effects of environmental factors on influenza transmission. Consequently, the predictive role of meteorological conditions was insignificant, as shown in Figure 4B.

Preliminary observations in the post-pandemic period indicate that we are currently in a transitional and restructuring phase of the study. During this stage, the influence of environmental factors shows localized recovery, but their patterns have undergone significant evolution. The impact of temperature has rebounded, with its SHAP value distribution range widening considerably compared to the pandemic period (see Figures 4B, C), although it has not yet fully recovered to pre-pandemic levels (as shown in Figure 4A). This suggests that traditional seasonal factors are returning, although their intensity may be weakened by persistent behavioral changes. The relative ranking of the influence of atmospheric pressure has significantly increased, making it the second most important factor after temperature. Its point cloud distribution has broadened, with high pressure (red dots) correlating with positive SHAP values. This may reveal a novel meteorological-disease linkage mechanism under the “new normal” that warrants further investigation. Other factors, such as precipitation and wind speed, showed limited recovery in influence, with their SHAP value distributions remaining relatively concentrated (Figure 4C).