Objective To compare the effectiveness of three influenza surveillance and early warning methods recommended by the World Health Organization (WHO) in influenza surveillance in China, and provide evidence for the improvement of influenza surveillance and early warning in China.

Methods The etiological surveillance data of influenza from sentinel hospitals in China between week 22 of 2010 and week 13 of 2025 were collected, and the data in summer (weeks 22–39) and winter-spring (weeks 40–21) were used for modeling respectively. Early warning models were developed by using a rolling modeling approach based on Peak Alignment Method (PAM), Moving Epidemic Method (MEM), and Autoregressive Integrated Moving Average (ARIMA) model. The sensitivity, specificity, Youden’s index, and kappa coefficient were calculated to evaluate the accuracy and consistency of the three methods for early warning. The weeks ahead of incidence season by early warning was used to evaluate the timeliness.

Results Overall, the sensitivity, specificity, and Youden’s index of PAM, MEM, and ARIMA were 0.86, 0.93, 0.79; 0.88, 0.88, 0.76; and 0.88, 0.93, 0.81, respectively. The kappa coefficients for PAM (0.79, Z=15.57, P<0.000), MEM (0.76, Z=15.03, P<0.000), and ARIMA (0.81, Z=16.06, P<0.000) all showed high consistency with the gold standard results. In the summer season, the sensitivity, specificity, and Youden’s index of PAM, MEM, and ARIMA were 0.65, 0.86, 0.50; 0.69, 0.69, 0.38; and 0.65, 0.93, 0.57, respectively. In the winter-spring season, the sensitivity, specificity, and Youden’s index of PAM, MEM, and ARIMA were 0.93, 0.95, 0.88; 0.94, 0.93, 0.88; and 0.96, 0.93, 0.89, respectively. In the summers of 2015, 2017, and 2022, The PAM, MEM, and ARIMA were all could be used for early warnings before the incidence peaks by 3 - 8 weeks. During the winter–spring seasons from 2015 to 2024, all the three models also made early warnings at 3 - 10 weeks ahead of the incidence peak.

Conclusion All the three methods had good performance in the influenza surveillance and early warning system in China. ARIMA was the most optimal method, followed by PAM, and MEM was least effective. The models all had good performances during winter-spring season, but they were less effective during summer season, with ARIMA performing relatively better. Given its overall advantage, ARIMA is recommended as the routine model for influenza early warning in China, with PAM serving as a supplementary approach.