This study was approved by the Research Ethics Committee of Southern Medical University (NFYKDX-ER2022012). The need for informed consent was waived because the data were deidentified and aggregated.

Individual data for deaths occurring between January 5, 2015, and December 30, 2018, in Guangzhou were obtained from the Guangzhou Center for Disease Control and Prevention (CDC). Individual death information included the underlying cause of death, age at death, and sex. The daily number of deaths was aggregated. Here, we considered 5 cause-of-death groupings associated with influenza virus infection: all causes (International Statistical Classification of Diseases and Related Health Problems, 10th Revision [ICD-10] codes A00-Z99), respiratory and cardiovascular diseases (ICD-10 codes I00-I99 and J00-J99), respiratory disease (ICD-10 codes J00-J99), pneumonia and influenza (ICD-10 codes J10-J18), and cardiovascular disease (ICD-10 codes I00-I99).

The Guangzhou CDC also provided influenza surveillance data, including the weekly proportions of specimens testing positive for influenza A (H1N1), A (H3N2), B/Yamagata, and B/Victoria, and the weekly proportion of consultations for influenza-like illness (ILI), that is, a body temperature ≥38 °C with cough or sore throat, among outpatient visits at sentinel hospitals in Guangzhou.

Data on annual population size were obtained from the Public Security Bureau of Guangzhou Municipality. We collected data on daily mean temperature and relative humidity from the website of the China Meteorological Data Service Center [24].

We defined a weekly influenza virus activity proxy by multiplying the weekly proportion of consultations for ILI and the proportion of specimens testing positive for different subtypes and lineages, and then multiplying the resulting value by 1000 [25,26]. Then, we converted the weekly proxy to a daily proxy with cubic smoothing splines with 1 df per week, determined with generalized cross-validation [21]. The daily population size was estimated by linear interpolation. Daily mean temperature and relative humidity data were used to calculate daily absolute humidity [17], which has been suggested to be a better predictor of influenza virus transmission and survival than relative humidity [27].

A quasi-Poisson regression model was applied to estimate the excess mortality associated with influenza for different cause-of-death groupings, age groups (<60 years, 60-79 years, ≥80 years, and all ages), and sex. The associations of death counts with influenza and temperature were determined using a DLNM [28], as follows:

where Y_t and μ_t were the observed and expected number of deaths on day t, respectively. φ is the overdispersion parameter. The logarithm of population (Pop_t), with a fixed regression coefficient of 1, was used as an offset. The influenza activity proxy variables for A (H1N1), A (H3N2), B/Victoria, and B/Yamagata on day t were Proxy_H1,t, Proxy_H3,t, Proxy_BV,t, and Proxy_BY,t, respectively. Cross-basis Tʹ_t,l(.) was constructed for the activity proxy variable of each influenza subtype/lineage, assuming a linear relationship between influenza virus activity and population mortality [3,6,13,14,21]. Cross-basis Tʹ_t,l(.) was constructed for daily temperature (Temp_t) on day t with a natural cubic spline with 4 df [29]. It should be noted that the lag-response associations for the temperature and influenza virus activity proxies were set up with natural cubic splines with 3 knots placed at equally spaced log values of lags of 30 days to fully capture possible short-term mortality displacement effects (Multimedia Appendix 1) [21,30].

In addition, we controlled the effect of absolute humidity (Hum_t) using a natural cubic spline with 4 df [22]. Additionally, we included a categorical variable for the day of the week (Dow_t) and an indicator variable for holidays (Holiday_t) in the model. Additional adjustments in the model included a natural cubic spline with 5 df for the day of the year (Doy_t; ie, 1-366) and a categorical variable for calendar year (Year_t), to control for seasonality and the time trend in death counts, respectively [21].

We present the relative risk (RR) and cumulative RR of death associated with an increase of 10 units in the influenza virus activity proxy relative to an influenza activity proxy of 0 across a lag of 0 to 30 days. The daily number of excess deaths related to influenza was estimated as the difference between the estimated number of deaths given the observed influenza virus activity and the estimate given that the influenza virus did not circulate (ie, the influenza virus activity proxy was 0) [22]. The excess mortality rates (EMRs) associated with influenza for different age groups were estimated by dividing the influenza-related excess death numbers by the corresponding population sizes and then multiplying by 100,000. It should be noted that the EMRs associated with influenza in this study were in units per 100,000 persons per year. We estimated the EMRs that incorporated the lagged effects of influenza on mortality over 0 to 30 days. Meanwhile, we applied Monte Carlo simulations to estimate the 95% empirical CI (95% eCI) of the EMRs, which accounted for autocorrelation using the Newey-West method [31-33].

To detect whether there was a sex difference in influenza-related EMRs, we divided the male EMRs by female EMRs to obtain the male-to-female excess mortality ratios (MFRs) and derived the corresponding 95% CIs using the delta method [22].

To check the robustness of the results, we conducted sensitivity analyses by (1) changing the maximum number of lag days for lag-response relationships between the death and activity proxies of each influenza subtype and lineage, as well as for the relationship between death and temperature; (2) changing the way we controlled the temporal trend in the mortality rate; (3) ignoring the classifications of influenza virus types and influenza B lineages; (4) applying different cross-basis matrices to temperature; and (5) including the influenza virus activity proxy and temperature at a lag of 7 days in the model (Multimedia Appendix 1). All analyses were performed in R (version 4.1.1; R Foundation for Statistical Computing).

In this study, we report excess mortality associated with influenza and its subtypes/lineages from 2015 to 2018 in Guangzhou, a subtropical city of China. We estimate that 1786 respiratory/cardiovascular disease deaths were attributable to influenza annually, accounting for 6.7% (7142/105,998) of all respiratory/cardiovascular disease deaths. We found that influenza-associated excess mortality was higher in the elderly, which is consistent with the conclusions of other studies [1,14,15,21,34].

Variations were observed in the estimates of influenza-associated excess mortality by subtype/lineage. The mortality burden of influenza B/Yamagata was higher than that of B/Victoria, which is consistent with previous studies [6,22,34]. The disparity between the lineages may be due to the elderly being more likely to be infected by influenza B/Yamagata than B/Victoria [35,36]. In 2018, influenza B/Yamagata was the predominant lineage in Guangzhou, and the vast majority of influenza-related excess deaths were attributed to B/Yamagata. Similarly, an outbreak dominated by influenza B/Yamagata was reported in Europe during the 2017-2018 influenza season, causing substantial excess mortality [34]. Together, the evidence implies that influenza B does not necessarily lead to mild illness in humans; the mortality burden of influenza B cannot be ignored. Furthermore, including influenza B lineages in active surveillance would help better understand the epidemiology of influenza B and reduce the excess mortality of influenza by enabling targeted interventions. Vaccination is the most effective way to prevent influenza; however, the vaccination coverage in China is low (9.4%) [23], partly because influenza vaccination is not government funded in many places [37]. During the 2018-2019 influenza season, the quadrivalent inactivated vaccine was approved for the first time in China [38]. In 2019, 30.78 million doses of influenza vaccines were distributed in China, but most of them were trivalent inactivated vaccines [37]. Compared with the trivalent influenza vaccine, the quadrivalent influenza vaccine has a higher antibody protection rate and antibody-positive seroconversion rate against the increased influenza B strain [39,40]. High-risk groups are recommended to receive quadrivalent influenza vaccines when vaccines are available, given the nonnegligible burden of influenza B.

We also noted that the mortality burden of influenza virus subtypes and lineages differed according to age group. For people aged ≥60 years, the fatality rate caused by influenza A (H3N2) was higher than that of A (H1N1), B/Victoria, and B/Yamagata. The risk of influenza virus infection and severe complications, including death, is increased by the stronger antigenic drift and virulence of influenza A (H3N2) [22,41]. However, for people aged 0 to 59 years, influenza A (H1N1) had a greater EMR than other influenza subtypes or lineages, which is in accord with previous studies [22]. The lower mortality burden of influenza A (H1N1) might be due to protection from early life exposure to this subtype [42].

Consistently with previous studies performed in Shanghai, we did not find a statistically significant difference in influenza-related excess respiratory/cardiovascular disease deaths between males and females [22]. However, males had a higher influenza-related excess respiratory mortality rate than females and a lower excess cardiovascular mortality rate. In addition, in people aged 60 to 79 years, it has been found that the influenza-related EMR for respiratory disease or cardiovascular disease in men is higher than in women, and the same situation was found in people aged 65 to 74 years in a US study [19]. However, for people aged ≥80 years, men had higher influenza-related excess respiratory mortality than women, but a lower excess cardiovascular mortality associated with influenza. In the US study, the results of a comparison of point estimates of EMR for the corresponding causes for males and females were consistent with our results, but the difference was not statistically significant [19]. Decreasing testosterone levels and higher respiratory mortality in older men [22,36], or the fact that women may have a different innate or acquired immune response than men due to discrepancies in sex steroid hormones [43], might lead to sex disparities in the severity of illness after influenza virus infection.

Our estimate of average annual influenza-associated excess all-cause mortality (25.06 per 100,000 individuals) was higher than that of previous studies in countries/cities and time periods including Greece from 2013 to 2017 (23.60) [21], Beijing from 2007 to 2013 (19.10) [14], Guangzhou from 2010 to 2012 (14.72) [15], and the United States from 1997 to 2007 (11.78) [19]. However, the estimate was relatively lower than the number of annual average influenza-related excess all-cause deaths in Europe for 2017-2018 (25.40) [34], Chongqing from 2012 to 2018 (33.50) [44], and Shanghai from 2010 to 2015 (27.66) [22]. Factors such as population structure, health status, health care capacity, circulating strains, host immunity, public health interventions (eg, influenza vaccination programs), and social distancing might contribute to these differences [20,22]. In addition to the abovementioned factors that may lead to heterogeneity in the results of different studies, some methodological differences need to be considered. The estimates of influenza-associated excess mortality varied with different analytical strategies, although the disparities were not significant in the sensitivity analysis. Influenza can be divided into different subtypes and lineages to explore the impact of influenza on the health of populations when relevant data are available, which also helps us understand the effects of different influenza subtypes and lineages on mortality.

Previous studies have estimated excess mortality based on the association between influenza activity proxies at a lag of 0, 1, or 2 weeks and mortality [20]. In this study, we used a DLNM to capture the association between influenza and mortality. A sensitivity analysis showed no significant differences between the influenza-associated excess respiratory/cardiovascular disease mortality rates estimated by a DLNM and a model that included linear terms of an influenza virus activity proxy and a natural cubic spline of temperature at a lag of 7 days. DLNMs allow us to consider a flexible lag-response relationship between influenza and death. The time course of influenza effects and potential mortality displacement can be examined with DLNMs [21,30]. We found that many influenza-related respiratory/cardiovascular disease deaths occurred during the initial exposure to influenza A (H1N1), A (H3N2), and B/Yamagata. Therefore, starting antiviral treatment as soon as possible, especially with neuraminidase inhibitors and other means of treatment for critically ill influenza patients, is likely to bring great survival benefits [35]. Taking into account the observed mortality displacement, we estimate that 743, 322, and 510 annual respiratory/cardiovascular disease deaths in Guangzhou were attributable to influenza A (H3N2), B/Victoria, and B/Yamagata, respectively; this indicates that the impact of influenza A (H3N2) and B on mortality is not limited to bringing forward deaths in the short term. Therefore, the mortality burden of influenza should not be overlooked. Efforts are needed to raise awareness of influenza as a severe disease, especially among high-risk groups.

This study had some limitations. First, we did not use age-specific influenza virus activity proxies to estimate influenza-associated excess mortality, and a proxy for all-age activity was used instead. Second, the study period was relatively short (ie, 4 years), which may have influenced the comparison of the mortality burden of different influenza subtypes and lineages. This is mainly because the virology data included samples that tested positive for influenza B, but information on lineage was unavailable before 2015. Data from 2019 onwards were not collected. Moreover, we did not consider the influence of factors such as respiratory syncytial virus and vaccination coverage on the estimate of influenza-associated excess mortality, since such data were not available.

In conclusion, after considering the observed mortality displacement, influenza was associated with substantial mortality in Guangzhou, occurring predominantly in the elderly. The mortality burden of influenza B, particularly B/Yamagata, cannot be ignored. Contrasting sex differences were found in influenza-associated excess mortality from respiratory diseases and from cardiovascular diseases, and the underlying mechanisms need to be investigated in further studies. Our findings can help us better understand the magnitude and time-course of the effect of influenza on mortality and inform targeted interventions for mitigating the influenza mortality burden, such as immunizations with quadrivalent vaccines (especially for older individuals), behavioral campaigns, and treatment strategies.