In contrast to the COVID-19 pandemic, influenza epidemics are persistent public health challenges that have occurred many times in human history. However, there is a lack of guidance on formulating NPI strategies to cope with influenza epidemics. To the best of our knowledge, the current study is the first to use a machine learning algorithm combined with an explainable artificial intelligence tool to quantify the effectiveness of COVID-19-targeted NPIs at suppressing influenza transmission. We found that social gathering limitations made the greatest contribution to suppressing influenza transmission. Additionally, we estimated the effectiveness of each intensity level of the NPIs, the intensity and time threshold for each NPI to have an effect, and the interaction between each pair of NPIs.
We found that gathering limitations, school closures, contact tracing, and health education promotion made the greatest contributions to suppressing influenza transmission. These results partially agree with those of previous studies estimating the effectiveness of NPIs against SARS-CoV-2 transmission [18, 19]. We found that school closures were more effective than workplace closures. The reason for this finding may be that the transmission of the influenza virus mainly occurs in children rather than adults. Previous studies have also found that closing schools played an important role in containing the 2009 H1N1 influenza pandemic [20,21,22].
We found that contact tracing was effective at suppressing influenza transmission, even though this NPI was specific to contact with COVID-19 patients. This may be because the disease symptoms of the two viruses are very similar (e.g., fever and cough), and both viruses are transmitted via the airborne route. Therefore, patients with a history of COVID-19 exposure may also have previously been or are currently exposed to the influenza virus, and patients with suspected COVID-19 symptoms under close tracing and monitoring may actually be infected with the influenza virus. Another reason for this finding may be the high co-infection rate of SARS-CoV-2 and influenza virus. A single-centered retrospective study of 307 COVID-19 patients conducted in Wuhan, China, reported that 57.3% of the patients were also positive for influenza viruses [23]. A recent experimental study found that influenza A virus pre-infection significantly increases susceptibility to SARS-CoV-2 infection [24]. Similarly, other NPIs, such as workplace closures and health education promotion, were also effective at suppressing influenza transmission, regardless of whether the NPI was restricted to regions suspected of having a COVID-19 outbreak or whether the NPI was implemented nationwide or at specific locations. These results may be partly explained by the fact that the two pathogens have the same transmission route, and the results further indicate the existence of interactions between the influenza virus and SARS-CoV-2. Although our results showed that health education promotion had a moderate effect and the longest lag time of the 14 NPIs, its contribution to suppressing influenza transmission in 2020 was sizeable (11.47%). This may be because health education promotion is easy to implement, and it was therefore frequently implemented by many governments globally.
For the first time, we quantified the intensity of each NPI to determine the intensity required to reach its maximum effect. For international travel restrictions, we found that only banning arrivals from certain foreign regions, rather than banning all regions or implementing a total border closure, was not effective at suppressing influenza transmission. The reason for this may be that for viruses capable of causing a global pandemic, such as influenza viruses and SARS-CoV-2, it is crucial to strictly restrict their transmission across international borders. For example, at the early stage of the COVID-19 pandemic, the US only banned travelers from China [25], even when there were emerging cases in Europe. This incomplete restriction of international movement resulted in a surge in COVID-19 cases in the US due to imported cases from Europe [26]. We found that school and workplace closures alone were not sufficient to suppress influenza transmission. This may be because leaders of schools and factories have inadequate knowledge about the threat of infectious diseases and therefore continued normal operation of their organizations. This indicates that governments need to enforce the requirements to close schools and workplaces. For health education promotion, our results showed that an official notice alone was not sufficient to significantly suppress influenza transmission, unless it was combined with a social media campaign. Therefore, the key role of social media in public health education and disease prevention should be recognized, and social media campaigns should be deployed by policymakers. However, false information and rumors spread easily via social media platforms, making the authority of an official notification indispensable.
We found that the NPIs had saturation points. Identifying saturation points may help to maximize the suppression effect, while minimizing the social and economic costs of the interventions. These findings indicate that enforcing NPIs past a certain intensity level may not provide any additional benefits. However, we lacked data regarding the public’s actual behaviors in reaction to the NPIs, which are more relevant to influenza transmission. Therefore, another possible reason for the saturation of the NPIs may be the deterrent effect of a certain intensity level. For example, restricting social gatherings to 1000 participants may also discourage people from attending social activities with less than 1000 participants. However, our analysis was restricted to the earlier stage of the pandemic, and the deterrent effect may decrease during a prolonged pandemic period, because of pandemic-policy fatigue among the public [27].
We took into consideration the lag time of each NPI. The results showed that health education had the longest lag time. This is reasonable, given that it takes a longer time for the public to receive and internalize public health information than information related to other NPIs. We also found that the lag time of some NPIs (e.g., domestic movement restrictions and testing policies) was less than 3 days, which is not reasonable in theory because the series interval (i.e., the time between two successive cases) of influenza is approximately 3 days [28]. This result may be because the effect of these NPIs was too negligible to allow the accurate calculation of the lag time. This is also supported by the low contribution rank of these NPIs.
We found a positive interaction effect when the intensity of the mask-wearing requirement and other NPIs (e.g., gathering limitations) were both high, whereas merely implementing a mask-wearing requirement at a high intensity showed a negative effect. This finding may be explained by risk compensation [29]. In other words, the public may assume that only wearing masks fully protected them from respiratory infections, thereby increasing the frequency and time of contact with others. However, influenza viruses are not only transmitted by air but also by contact (e.g., a person touches a surface with accumulated droplets from an infected person and then touches his/her face) [1]. Therefore, the number of influenza cases may have increased in those people who had close contact with others, despite wearing a mask.
A negative interaction was observed when both the intensity of international travel restrictions and the intensity of contact tracing were high, whereas a positive interaction effect was observed when the intensity of contact tracing was low. This may be because when the intensity of international travel restrictions is high enough to suppress influenza transmission, contact tracing may no longer be necessary due to a decrease in the number of cases. Additionally, during the early declining period of the influenza epidemic, we observed a positive interaction effect when the intensity of both contact tracing and testing policies were high (Additional file 1: Fig. S8C), which may indicate that other NPIs were involved.
Our study has some strengths. First, compared with previous studies that simply used the time of NPI introduction as a surrogate for the effect of the NPI, we used a more reliable, detailed, and comprehensive NPI database, which allowed us to quantify different levels of each NPI. Second, most previous studies [6, 8,9,10] simply investigated one country or region, whereas we, for the first time, included and compared the effects of various NPIs on influenza transmission across 33 countries in the Northern Hemisphere. Third, previous studies [5, 18, 19] evaluating the effectiveness of NPIs at suppressing COVID-19 spread lacked accurate case data, because at the time, COVID-19 was an emerging infectious disease and detection kits with high sensitivity and specificity were lacking. Importantly, the testing rate is very likely to be influenced by the intensity of testing policies, which vary by country. In comparison, the well-established influenza surveillance system used in this study (i.e., the GISRS) has consistently and reliably monitored influenza activity since 1952 [30]. Finally, we used a machine learning model and an explainable machine learning method to capture the complex relationship between the intensity of the NPIs and the influenza suppression index. Hence, we were able to obtain a range of more detailed and informative results.
Nevertheless, our study has several limitations. First, due to the lack of influenza incidence data in our dataset, we used the influenza positivity rate to approximate actual influenza activity. Although previous studies [31, 32] have used the influenza positivity rate or the multiple influenza-like illness rate to approximate the incidence of influenza, the majority of countries (20/33, 60.6%) in our dataset did not report the influenza-like illness rate. Second, the influenza positivity rate may have been underestimated, because in the context of the COVID-19 pandemic, the health-seeking behavior of influenza patients may have been reduced and medical resources tended to be inadequate. Nevertheless, the number of specimens tested in 2020 was similar to the number tested in the previous year (1,701,758 vs. 1,675,945, respectively) [33]. Third, the 95% confidence intervals of the Spearman’s correlation coefficients of the NPIs with different lag times overlapped (Additional file 2: Table S4). These results may partly be explained by the fact that the lag times of individual NPIs may not be a definite integer number of days, and some NPIs may, in reality, start to have an effect after 84 h (3.5 days). However, since the minimum unit of time for our NPI data was 1 day, it was not possible to specify a smaller unit of time, which may have affected the accuracy of the lag time analysis. In addition, it is possible that the efficiency of different governments at implementing NPIs and the degree of cooperation of the public may have also affected the speed of NPI implementation. Therefore, the number of lag days for the 33 countries included in our analysis may not have been accurate. In summary, we were only able to choose one approximate value that was as close to the true number of lag days as possible. Finally, we used an ecological study design, and therefore, the effectiveness of the NPIs may have been influenced by a range of uncontrolled confounding factors specific to the country in which the NPIs were implemented. These confounding factors may have included the country’s demographic structure and climate and the presence of other NPIs. Due to the same reason, we were unable to adjust for the specific implementation details of each NPI, which may have varied by country. Therefore, the results for interactions between NPIs should be interpreted carefully.
