Web search logs have been effectively applied to help monitor influenza activities in many developed countries 
. Carneiro and Mylonakis suggest that using Google Trends for disease surveillance is better suited in developed countries 
, which have large populations of internet search users. However, despite China's status as a developing country, it has nearly 400 million internet users 
. In Guangdong province over 40 million people have access to the Internet, accounting for 40% of the total population. Such a large population of web users should provide reliable data for search-term influenza surveillance in the province.
The statistical analysis undertaken indicates temporal correlations between some Google Trends in Chinese language and influenza epidemics. Both ILI surveillance and influenza virus surveillance in Guangdong are correlated with Google Trends data statistically, but the statistical significance for the search terms are different. At a provincial level, the GT search term data for fever and cough are significantly correlated with conventional surveillance data. Interestingly, GT for fever is more highly correlated with ILI surveillance while Cough and Common cold are more associated with virological surveillance. Fever and cough are the two most common manifestations in influenza cases. Those patients with ILI are usually aware of these two presenting complaints at the pre-diagnosis stage of influenza. Thus these two search terms should be sensitive to influenza epidemics.
The local habits of web searchers can be influenced by users' level of education and their cultural and language backgrounds 
. Pelat et al 
reported that the Google search for Influenza
in French is highly correlated with ILI surveillance data (ρ
<0.001). Valdivia and Monge-Corella 
also found that queries for Influenza
in Spanish show a significant correlation (ρ
0.70) with national ILI surveillance data. In contrast, the GT term for Flu
in Chinese was not statistically correlated with the official surveillance data, suggesting that this professional term in Chinese hardly reflects the influenza activity. This study also found that Google Trends have a lower correlation with weekly positive rates for influenza virus test than they do with ILI percentages which is consistent with Ortiz et al 
who argued that Google Flu Trends does not do a good job in estimating laboratory-confirmed influenza cases. As a type of syndromic surveillance, the ILI surveillance system is designed to collect data from likely influenza cases in order to signal actual influenza activities. Neither fever nor cough is a specific symptom caused only by influenza virus infection. Our results indicate that search terms describing symptoms of influenza rather than those professionally used key words can better reflect actual influenza epidemics in the south of China.
Guangdong reported 9896 laboratory-confirmed cases of H1N1 and 36 deaths in 2009 
. The high prevalence of H1N1 influenza increased online health-seeking activity. However, health care seeking behavior and internet search behavior might be different and change over time during a pandemic period. GT in 2009 were more strongly associated with surveillance data than those in the other years. It appears that those affected patients typed in some topical words for this pandemic period, like Fever
and Influenza A
, to search health information on the web. The increasing public concern and media interest also raised the level of internet searching. Taking H1N1
for example, as a search term, it became a hot word for internet search in 2009. However the yearly correlation coefficient shows the GF term H1N1
had no association with influenza during the pandemic period, which is in contrast to the overall coefficient. In Guangdong, GT for H1N1
spiked in May 2009, but the actual local H1N1 incidence peaked in November. The H1N1
search trends did not reflect the actual H1N1 pandemic activity. A possible reason for the 5 months gap between the two peaks is that the mass media started to report H1N1 events when this novel influenza virus was first imported into Guangdong in May 2009. At that time, continuing news reports, outbreak briefs and health publications on the web heavily influenced H1N1
search trends. Hence the May peak was associated with fear and information gathering in the community and the November peak of actual cases was the result of the actual progression of the spread of the disease within the province.
It is believed that increasingly patients are using web searches for health information prior to seeing a doctor 
. Hence internet search trends can reflect actual epidemics earlier than conventional surveillance. One advantage of Google Trends is that data can be obtained earlier, more easily and at little cost, while the CDC published surveillance reports usually need one to two weeks for laboratory tests and data analyses.
However, publicly reporting official surveillance data can also raise awareness of health risks and increase internet searches. Possibly, because of this interaction, our study did not find any significant improvements in correlations between Google Trends and influenza surveillance with time lags, in contrast to previously reported findings 
There are several limitations of this study. In terms of correlation coefficients, our results are lower than those in prior studies that compared Google Flu Trends data to traditional surveillance data 
. First, our study obtained only 4 years of Google Trends data, and for some search terms, there was insufficient data collected over the study period. The correlation between influenza virus surveillance data and GT data was limited to only two years of data beginning in 2009. A sore throat is also a common symptom in ILI cases, but Google fails to calculate its trends in Chinese due to insufficient proportions of this term in the total searches. Additionally, users may enter synonyms that we did not collect. More search terms might need to be investigated and correlated with standard surveillance data. Another limitation is that our analyses likely over- and underestimated some correlations. It is difficult to identify to what extent search trends are generated by true cases. Cook et al 
suggested that search data may work well for diseases with less media exposure as media reports will probably drive more non-patients to increase their web search, which can influence search trends but not reflect the actual disease activity. The correlations found in this study between CDC surveillance data and GT for H1N1
in 2009 are consistent with this finding.
In conclusion, this study has shown Google Trends data using Chinese search terms are generally well correlated with conventional methods of surveillance. Google Trends, especially those related to ILI symptoms could be used as a complementary source of data for influenza surveillance in south China. However care should be taken when there is high media reporting of a particular influenza illness, which can bias internet search trends. The development of search-term based surveillance is still in its early phase. While considering the impacts of publicity, research in the future should develop new tools using search trends in Chinese language to estimate local disease activity as well as assist in detecting early signals of outbreaks.