Cumulative Query Method for Influenza Surveillance Using Search Engine Data

Background Internet search queries have become an important data source in syndromic surveillance system. However, there is currently no syndromic surveillance system using Internet search query data in South Korea. Objectives The objective of this study was to examine correlations between our cumulative query method and national influenza surveillance data. Methods Our study was based on the local search engine, Daum (approximately 25% market share), and influenza-like illness (ILI) data from the Korea Centers for Disease Control and Prevention. A quota sampling survey was conducted with 200 participants to obtain popular queries. We divided the study period into two sets: Set 1 (the 2009/10 epidemiological year for development set 1 and 2010/11 for validation set 1) and Set 2 (2010/11 for development Set 2 and 2011/12 for validation Set 2). Pearson’s correlation coefficients were calculated between the Daum data and the ILI data for the development set. We selected the combined queries for which the correlation coefficients were .7 or higher and listed them in descending order. Then, we created a cumulative query method n representing the number of cumulative combined queries in descending order of the correlation coefficient. Results In validation set 1, 13 cumulative query methods were applied, and 8 had higher correlation coefficients (min=.916, max=.943) than that of the highest single combined query. Further, 11 of 13 cumulative query methods had an r value of ≥.7, but 4 of 13 combined queries had an r value of ≥.7. In validation set 2, 8 of 15 cumulative query methods showed higher correlation coefficients (min=.975, max=.987) than that of the highest single combined query. All 15 cumulative query methods had an r value of ≥.7, but 6 of 15 combined queries had an r value of ≥.7. Conclusions Cumulative query method showed relatively higher correlation with national influenza surveillance data than combined queries in the development and validation set.


Introduction
Syndromic surveillance may alert public health care providers in the early phases of an outbreak, allowing them to decrease morbidity and mortality resulting from the outbreak [1][2][3][4][5]. Syndromic surveillance is defined as the real-time or near real-time collection, analysis, interpretation, and dissemination of health-related data to enable the early identification of the impact of potential human or veterinary public health threats that require effective public health action [1,3]. The 2009 H1N1 influenza pandemic highlighted the need for syndromic surveillance to inform policy and plan for effective responses.
Because conventional syndromic surveillance of indicators such as influenza-like illness (ILI) depends on case reporting to report disease activity, time delays in reporting and case confirmation can interfere with the early detection of outbreaks or increases in influenza cases in the community. Thus, researchers have been investigating alternative data sources for the detection of outbreaks. For example, over-the-counter sales of medications and school absenteeism data have been used for earlier detection of outbreaks [6][7][8][9][10][11][12].
Internet search queries have become an important data source in recent years [13][14][15][16][17][18][19][20][21][22]. Internet search engines allow billions of people to have instant access to a vast amount of information online. New syndromic surveillance sources, such as Google Flu Trends (GFT), provide the potential to identify influenza outbreaks in real time [23]. Several studies have reported that GFT is highly correlated with conventional ILI surveillance data [23][24][25][26][27][28]. GFT has now been applied in many countries, but neither GFT nor other search query-based tools for disease surveillance are available in South Korea [24,[27][28][29][30]. Generally, Google's market share is dominant in the countries where GFT is available [24][25][26][27][28][29]31], but not in South Korea [32]. Studies using Google Trends for influenza surveillance show that it can be used as a complementary source of data but that its performance is insufficient for use as a model for prediction [33,34]. It is difficult to find queries that show high correlations for consecutive years because Internet searching behavior may change over time [33,34]. To reduce the effects from changes in search queries, we used a combination of queries and cumulation of combined queries from the search engine Daum. Daum is the second largest Web portal service provider in daily visits in South Korea (approximately 25% of the market share) [32,35]. Daum offers many Internet services to Web users, including email, messaging service, forums, shopping, and news. The main language is Korean.
In South Korea, influenza is generally seasonal, with most activity occurring during winter. The 2009/10 epidemiological year, called the Influenza A (H1N1) pandemic period, was an exceptional situation (see Multimedia Appendix 1). The primary objective of this study was to examine correlations between our cumulative query method and national influenza surveillance data.

Study Period
The study period was September 6, 2009 (week 36), through September 1, 2012 (week 34)-156 weeks of data for 3 consecutive epidemiological years. We divided the study period into two sets: Set 1 (the 2009/10 epidemiological year for development set 1 and 2010/11 for validation set 1) and Set 2 (2010/11 for development set 2 and 2011/12 for validation set 2).

Collection of Influenza-Like Illness Data
We collected the ILI data from the Korea Centers for Disease Control and Prevention (KCDC) as a gold standard. KCDC ILI data were available from the KCDC website; we downloaded the ILI data for the study period from this site [36]. A KCDC case of ILI was defined as a person with a fever of 38°C with a cough and/or a sore throat [36]. ILI surveillance consisted of 850 sentinel clinics in South Korea, and the clinics reported weekly percentages of outpatients who met the case definition of ILI [36].

Survey for Obtaining Queries
To obtain population search queries related to influenza, we conducted a survey from quota sampling based on sex and age in September 2012. The quotas were based on address of resident registry, age, and sex. There were five quota groups by age: 20-29 years, 30-39 years, 40-49 years, 50-59 years, 60 years or older. Half of each quota group were female. We randomly selected the addresses from the residence registry in Seoul, and then if interviewees living at the address of residence registry met the criteria, we included the oldest interviewee. We then conducted face-to-face interviews. The survey included searching history for influenza and typed queries. The survey was performed anonymously. A KCDC definition of ILI was a person with a fever (발열 in Korean) of 38°C with a cough (기 침) and/or a sore throat (인후통). These three queries from the definitions of ILI were included in the queries for the following operations, regardless of the survey result. In the case of queries originally submitted in English only, we translated them to Korean and added them as new queries.

Combination of Queries
We believe that people typically search for things of interest on the Internet using one or more queries at a time. To reflect people's searching behavior and include as many queries as possible, we used a combination of queries. Queries from the survey results and the definition of KCDC ILI were divided into groups as follows: query group 1 consisted of queries specific to influenza (eg, "H1N1", "Influenza"), and query group 2 contained queries not specific to influenza (eg, "Treatment", "Symptom"). Then, we combined query groups 1 and 2. Combined queries consisted of query group 1 alone and a combination of query groups 1 and 2 (eg, "H1N1", "H1N1 Treatment", "H1N1 Symptom", "Influenza", "Influenza Treatment", "Influenza Symptom").

Collection of Data from Search Engine
We sent the combined queries and the queries that belonged to query group 1 (because these queries were searchable by themselves) to Daum and received proportional data in weekly form. Proportional data for these combined queries were extracted from the Daum search engine during development sets 1 and 2. Proportional data from the Daum search engine were calculated by dividing the number of each combined query by the total number of search queries for 1 week.

Creating Cumulative Query Methods and Data Analysis
Pearson's correlation coefficients were calculated between the Daum data for the combined queries and the KCDC ILI data in development sets 1 and 2. We selected the combined queries for which the correlation coefficients were .7 or higher and listed them in descending order. To see the change of correlation coefficients over time, we also calculated correlation coefficients of the combined queries in subsequent epidemiological years. We then created a cumulative query method n representing the number of cumulative combined queries in descending order of the correlation coefficient. For example, cumulative query method 4 consisted of a summation of the proportional data from the 1st, 2nd, 3rd, and 4th highest combined queries on the correlation coefficient list. In validation sets 1 and 2, Pearson's correlation coefficients were calculated between the cumulative query method n and the KCDC ILI data. Specifically in validation set 2, we analyzed the cumulative query methods from development set 2 as well as development set 1. Useful cumulative query methods in the validation sets were defined as having higher correlation coefficienst than the highest correlation coefficient of a single combined query in the same development set. Analysis was performed using IBM SPSS Statistics software, version 20. Significance was set at P<.05.

Institutional Review Board
This study was approved by the Institutional Review Board of Asan Medical Center (Seoul, Korea).

Survey for Obtaining Queries
We contacted 322 people and included 200 participants older than 20 years who lived in Seoul, Korea. Over a quarter (56/200, 28%) answered "Yes" to the question of searching history for influenza and provided search queries (Table 1). Since the Influenza A (H1N1) pandemic period, media began to use "New flu (신종플루)" to distinguish the H1N1 influenza and previous influenzas in Korea. In 2010, KCDC announced that the official term was "Influenza (인플루엔자)". But "New flu (신종플루)" and "Bad cold (독감)" are still more popular terms than "Flu (플루)" or "Influenza (인플루엔자)" in Korea. "Bad cold (독감)" in Korean has two meanings: one is influenza and the other, a severe common cold. b The query was originally submitted in English.

Combination of Queries From the Survey
Query group 1 contained 14 queries that were specific to influenza, and query group 2 had 14 queries that were not specific to influenza (Table 2). A total of 210 combined queries were submitted to Daum. Full data of combined queries are presented in Multimedia Appendix 2. Decease Tamiflu (English) c a Query group 1 consisted of queries specific to or related to influenza. Query group 2 contained queries not specific to influenza. b Abbreviation. c The query was originally submitted in English. d "Bad cold (독감)" in Korean has two meanings: one is influenza and the other, a severe common cold. "Flu" in query group 1 is "플루" which is the English pronunciation written in Korean. In Korea, "Bad cold (독감)" is a more popular term than "Flu (플루)" or "Influenza (인플루엔자)". e PCR: polymerase chain reaction.

Collection of Data From Search Engine
Correlation analysis was performed between the Daum data for combined queries and the KCDC ILI data in development sets 1 and 2 (Table 3). In development set 1, "New flu (abbr.)" had the highest correlation coefficient (r=.894, P<.001), and 13 combined queries had correlation coefficient r values of ≥.7. Among these 13 combined queries, the number of the combined queries that had correlation coefficient r values of ≥.7 was reduced to 4 in validation set 1 and to 2 in validation set 2. In development set 2, "Bad cold + Symptom" had the highest correlation coefficient (r=.969, P<.001), and a total of 15 combined queries had an r value of ≥.7. Among these 15 combined queries, the number of the combined queries that had correlation coefficient r values of ≥.7 was reduced to 6 in validation set 2. Only "Tamiflu" and "New flu + Symptom" showed correlation coefficients r values of ≥.7 for 3 consecutive years (Figure 1). The change of correlation coefficients for all combined queries over time are presented in Multimedia Appendix 2.

Creating Cumulative Query Methods
A total of 13 cumulative query methods were created in development set 1 (see Table 4). In each development set, cumulative query methods had a higher correlation coefficient than combined queries (see Tables 5 and  6). After 1 year, 11 of 13 cumulative query methods had an r value of ≥.7, but 4 of 13 combined queries had an r value of ≥.7 in validation set 1 (see Table 5 and Figure 3). All 15 cumulative query methods had an r value of ≥.7, but 6 of 15 combined queries had an r value of ≥.7 in validation set 2 (see Table 6 and Figure 4).  a We selected the combined queries for which the correlation coefficients were ≥.7 and listed them in descending order. We then created a cumulative query method n representing the number of cumulative combined queries in descending order of the correlation coefficients. b P<.05. c Correlation of cumulative query method 1 in validation set 2 from development set 1 cannot be computed because it has a constant value in that period (see Multimedia Appendix 2). d Useful cumulative query method in the validation set was defined as having higher correlation coefficient than the highest correlation coefficient of a single combined query in the same development set. e P=.20.

Principal Findings
In this study, the cumulative query method showed relatively higher correlation with national influenza surveillance data than combined queries in the development and validation set.
Many people use Internet searches for health information before visiting a doctor [18,20,23,33]. Hence, search query trends can reflect actual disease progression earlier than conventional surveillance. Queries used prior to this study only reflected the authors' opinions [13] or were obtained from databases [13,14,23,37]. To obtain population search queries, we carried out a study survey.
Search queries may vary from country to country. In Korea, "Bad cold (독감)" in Korean has two meanings: one is influenza and the other, a severe common cold. Since the 2009/10 epidemiological season, the Influenza A (H1N1) pandemic period, the media began to use "New flu (신종플루)" in order to distinguish H1N1 influenza and previous influenzas. In 2010, KCDC announced that the official term was "Influenza (인플 루엔자)" [36]. But "New flu (신종플루)" and "Bad cold (독 감)" are still more popular terms than "Flu (플루)" or "Influenza (인플루엔자)" in Korea (Table 1).
For the 2009/10 epidemiological year (development set 1), 13 combined queries had correlation coefficient r values ≥.7. However, only 4 of these combined queries ("New flu + Cough", "Flu + Symptom", "New flu + Symptom", and "Tamiflu") had correlation coefficient r values ≥.7 in the 2010/11 epidemiological year (validation set 1) ( Table 3). But 11 of 13 cumulative query methods had an r value of ≥.7 in validation set 1 (see Table 5 and Figure 3). Among 15 combined queries of development set 2, the number of the combined queries that had correlation coefficient r values of ≥.7 reduced to 6 in validation set 2. But all 15 cumulative query methods had an r value of ≥.7 in validation set 2. We think that the cumulative query method is more robust with time, and this factor is helpful for improving surveillance performance using search queries. Since Internet searching behavior may change over time, this could have affected the performance of Web query-based surveillance model [31]. In this study, 20 out of 210 combined queries had correlation coefficients for all 3 years. And only "Tamiflu" and "New flu + Symptom" showed correlation coefficients r values of ≥.7 for 3 consecutive years (see Figure  1 and Multimedia Appendix 2). Recently, a study using Google Trends for influenza surveillance showed that Google Trends can be used as a complementary source of data [33]; however, its performance is insufficient for use as a model for prediction because its maximum correlation coefficient was .82 for only one query, "Fever", in 2009, and the coefficient decreased to .64 in 2011 [33].
It is difficult to predict the change of search queries in the future.
To reduce the effects from changes in search queries, we used a combination of queries and cumulation of combined queries to construct our method. Additionally, the method we wanted to develop was meaningful only when the cumulative query method had a higher correlation coefficient than the highest single combined query. In each validation set, 8 useful cumulative query methods were developed. The useful cumulative query methods in each validation set had a high correlation coefficient (Table 4). In validation set 2, the range of correlation coefficients of the useful cumulative query methods was from .975 to .987. These values are similar to or higher than those reported elsewhere [13,14,23,26,27,29,31]. In Europe, correlation coefficients of .716 to .940 were reported for GFT [27], and coefficients of .82 to .99 were reported in the United States [23,26,31]. In the 2009/10 epidemiological year, called the Influenza A (H1N1) pandemic period, the proportional data of queries were likely to have been different compared to the other epidemiological year. It might affect performance of cumulative query methods in set 1. The performance of the cumulative query method in set 1 was decreased with time ( Table 4). It is thought to be related to the changes of queries (see Table 3). For some cumulative query methods, the correlation did not increase by adding queries. The added query did not give extra value in the cumulative query methods 6, 8, and 10 in validation set 1 (see Table 4 and Multimedia Appendix 2). Combined queries 6, 8, and 10 from development set 1 in validation set 1 have a constant value 0 (see Table 3 and Multimedia Appendix 2). The added queries were relatively too small compare to the previous queries in the cumulative query methods 5, 9, 10 in validation sets 1, 3, 4, 7, 8, 10, 12, and 14 in validation set 2 (see Table 4 and Multimedia Appendix 2).
We used proportional data from Daum, a non-dominant local search engine (approximately 25% of the market share) in South Korea [32]. Our cumulative query methods showed a strong correlation with KCDC ILI data. Generally, Google's market share is dominant in countries where GFT is available [27,28].
Our study showed the possibility of developing a surveillance model using a non-dominant local search engine.

Limitations
There are several limitations to this study. The survey of our study is not a representative sample. Because respondents were asked to provide typed queries without mention of the influenza pandemic of 2009/10, recent search queries were more likely to have been included in this study because the survey was conducted recently. This might affect performance of the cumulative query method. Further, the data from the influenza pandemic of 2009/10 might affect the outcome of this study. In this study, we did not combine queries from the same query group. Although important, the performance of using symptoms in the definition of KCDC ILI was not tested. The learning effect from the influenza pandemic of 2009/10, news reports, outbreak briefs, health information from the Internet, and changing search behavior stemming from the diffusion of smartphones might have affected the outcome of this study. We did not determine the extent to which these factors affected the searching behavior. More data for subsequent years are required in order to know the life of the cumulative query method.

Conclusion
We presented a cumulative query method using search engine data. We conducted a survey to obtain population search queries.
To reduce the effects from changes in search queries, we used a combination of queries and cumulation of combined queries.