The use and applicability of Internet search queries for infectious disease surveillance in low- to middle-income countries

INTRODUCTION

Globally, the number of different emerging infectious diseases is increasing^[1]. They are defined as “diseases that are newly recognized in a population or have existed but are rapidly increasing in incidence or geographic range”^[2].

According to the World Health Organization (WHO)^[3], in low-income countries, communicable neonatal conditions and lower respiratory infections are common causes of death. In lower- to middle-income countries, communicable neonatal conditions and lower-respiratory infections each caused over one million deaths in 2019 and remain the third and fifth leading causes of death worldwide, respectively^[3].

In areas with inadequate capacity to detect and monitor local disease incidence, infectious diseases often spread unnoticed, whereas detecting disease outbreaks early and taking targeted countermeasures is a challenge for national health systems. This is particularly faced by developing regions, such as countries in Sub-Saharan Africa^[4].

Outbreaks can also pose threats to economies and livelihoods and undo decades of progress. They represent a major public health concern^[1]. Therefore, the prevention of uncontrolled outbreaks is fundamental. The more quickly disease outbreaks are recognized, the earlier investigations and control measures can be applied. Hence, surveillance systems are used to monitor timely information for outbreak detection or characterization^[5].

Apart from traditional surveillance approaches, when routine health data are available and accessible, “syndromic surveillance” tools have been incorporated to “perform statistical tests for aberrant temporal or temporal/geographical trends” ^[5] so that epidemiologists and public health experts can assess the results and implement countermeasures. Syndromic surveillance uses the disease syndromes, e.g., symptoms or other indicators of the particular disease, and is a useful application if the cause of the disease or the outbreak is yet unknown. Syndromic surveillance aims to rapidly monitor syndromic categories of diseases, usually daily or more frequently^[5]. Nevertheless, the use of health records is expensive, and therefore its use for timely outbreak detection is limited^[6]. In any case, information searching is a human trait.

In the era of digitalization, searching the Internet is the first address when somebody has a question about a certain topic. Half of the assessed participants in a cross-sectional study searched online for information before they consulted their general practitioner^[7]. In evolving situations, such as during an epidemic, researchers try to utilize this information-seeking behavior of a population by using the Internet search queries to predict potential outbreaks^[8].

In 2020, the most searched term worldwide was “coronavirus”^[9]. This “trend” is summarized by Google to show the change in online interests in time series for a selected period and region. The data, which are normalized by Google into the relative search volumes (RSVs), can be downloaded directly from the Web in a comma-separated value format. For this reason, different studies used Google trends data (GTD) to model the spread of the coronavirus in selected regions^[10-12]. However, the principle to use GTD for epidemiological surveillance was not a novel idea.

Already in 2008, Google introduced flu trends to predict doctor visits for influenza-like illness in the US^[13]. The motivation was to supplement traditional surveillance systems to help to predict outbreaks of the flu^[14]. In 2013, the GTD model predicted more than double the proportion of physician consultation than the center for disease control and prevention (CDC)^[15]. The model was overfitted, a phenomenon when prediction models, algorithms, and computational techniques are too specific^[1]. Therefore, the regression line describing the model traverses each data point, “finding” random patterns in the training data that are not part of the true model in the population it is generalized to^[16]. Google stopped the Flu Trends project in 2015; Google Dengue Trends (2004-2015) is also no longer available.

In 2009, Gunther Eysenbach introduced the terms “infodemiology” and “infoveillance”^[17]. The terms combine the word information and epidemiology or surveillance to demonstrate the symbiosis of the sciences. He defined “infodemiology” as the “science of distribution and determinants of information in an electronic medium, specifically the Internet, or in a population, with the ultimate aim to inform public health and public policy”^[17]. The corresponding use of infodemiology data was therefore called “infoveillance”^[17].

Since the introduction of the terms, Internet data have been increasingly integrated into epidemiological research^[18]. However, it is not yet known in which settings the use of GTD in epidemiological surveillance for emerging infectious diseases gives valid and reliable results. The literature suggests that unlimited Internet access is crucial to receive data that represent the information-seeking behavior of the whole population^[18].

A growing digital infrastructure, particularly in the mobile sector in regions of Sub-Saharan Africa, and technological advances offer promising opportunities for more efficient information and knowledge transfer to improve epidemiological health surveillance. The collection of information from mobile devices and the use of open data sources, e.g., freely accessible population and environmental data, hold enormous potential as a new component of traditional surveillance systems, which at the same time requires intelligent networking of qualified actors across national borders and individual disciplines.

To receive relevant information on the use and applicability of GTD in epidemiological surveillance in other low- to middle-income regions in the world, this review aims to identify relevant studies that used GTD in prediction models for epidemiological surveillance purposes regarding emerging infectious diseases such as HIV infections, SARS, COVID-19, Lyme disease, Dengue fever, E. coli, Hantavirus, West Nile virus, or Zika virus infection^[19].

METHODS

RESULTS

The search identified 265 relevant references [Figure 1]. The search date was 2 December 2021; 184 were excluded after the title and abstract screening. In total, 19 articles were eligible, and their full texts were screened. Six articles were excluded [Supplementary Material File 1]; three studies were not original research articles (e.g., the theoretical development of a mathematical model), one did not include an eligible disease, and two examined the wrong setting. Out of 13 included studies in the review, four focused on Dengue fever^[25-28], four on Zika virus^[29-32], and five on COVID-19 disease^[33-37]. No study was identified that had met the inclusion criteria and focused on HIV infections, SARS, Lyme disease, E. coli, Hantavirus, or West Nile virus. It was striking that most of the 265 identified studies were from Western and high-income countries. In the studies included in our review, most were from Latin America (Brazil, Martinique, Honduras, El Salvador, Bolivia, Colombia, and Venezuela), three were from Asia (India and Indonesia), and Sub-Saharan Africa was underrepresented with only one study (South Africa). The study by Teng et al.^[32] regarded the Zika virus globally.

Figure 1. Flow chart (According to Page et al.^[24]).

Dengue fever

Four of the included thirteen studies examined the association of GTD and Dengue fever cases. The oldest included study^[27], which was co-authored by Google employees, described the development of models that were able to estimate Dengue cases by using GTD. The regarded period was from 2003 to 2010, and higher-income countries were examined in addition to Bolivia, Brazil, India, and Indonesia. The found correlations ranged between r = 0.87 (India) and r = 0.94 (Brazil) for the overall dataset. In the dataset studied (which modeled an outbreak, irrespective of the calendar year), the correlation was lower in Bolivia (r = 0.83). The study did not provide information on the significance levels of the found correlation coefficients. In 2005, there was a major Dengue virus outbreak in a rural area of India (Kolkata, West Bengal) where using Google for Internet search was not yet common, compared to cities such as Delhi or Mumbai. The model on GTD therefore underestimated the cases, showing how dependent the fit is on Internet penetration. Nonetheless, the authors concluded that GTD can be used to adequately estimate true Dengue activity. A recently published study related to Dengue fever^[26] examined the correlation between GTD and the Indonesian national surveillance report from 2012 to 2016. Three different trends were used, the Dengue symptoms, the term “Dengue”, and an Indonesian abbreviation of Dengue. Visually, the authors found a linear time-series pattern of the GTD with the official Dengue report. In contrast to other studies, Husnayain et al.^[26] examined the overall period, included each year separately, and time lags of 1–3 weeks, respectively. When the Dengue symptoms were used for the GTD, the overall correlation was significantly high (r = 0.937, P < 0.05). In 2016, the correlation, therefore, was the highest for all years (r = 0.954, P < 0.05). With increasing time lag, the correlation decreased (time lag 0 weeks: r = 0.937, P < 0.05; time lag - 3 weeks: r = 0.517, P < 0.05). For the term “Dengue”, the overall correlation was significantly high (r = 0.93, P < 0.05), and again the highest for 2016 (r = 0.966, P < 0.05) and for zero time lag (r = 0.931, P < 0.05). When the Indonesian abbreviation of Dengue was used, the correlation was still high, but slightly lower compared to the previous keywords. Overall, the correlation was r = 0.921 (P < 0.05), higher in 2016 only (r = 0.95) and for zero time lag (r = 0.921, P < 0.05). The time lag analysis suggests that the correlation is the highest when no time lag is used. However, with one week’s time lag, the correlation remained high (r = 0.755-0.773). The authors concluded that GTD can be used for an early warning system and that the applied models can be implemented as a tool before or during a Dengue outbreak in Indonesia. Marques-Toledo et al.^[28] investigated primarily Twitter data for Dengue fever surveillance in Brazil from September 2012 to October 2016 but used GTD for comparison. The used keyword was “Dengue”, which was significantly highly correlated with officially confirmed Dengue cases in the observed period (r = 0.92, P < 0.01). The authors concluded that GTD is a useful source for surveillance and prevention of Dengue fever outbreaks in Brazil. In addition, Monnaka et al.^[25] assessed GTD and Dengue cases in Brazil but only for the region of Sao Paulo between 31 December 2017 and 30 March 2019. They found strong positive Pearson’s correlations of 0.91 (P < 0.05) and 0.79 (P < 0.05) with a time lag of - 1 week. The estimated correlation coefficients are summarized in a spider diagram [Figure 2]. In general, the reported values show a high correlation between GTD and Dengue fever cases, with the highest correlations in Bolivia from 2003 to 2010 (r = 0.94) and Indonesia from 2012 to 2016 (r = 0.937). The lowest correlation was found in India from 2003 to 2012 (r = 0.87). Indonesia was the most studied country in the included studies. The results are congruent since the values of the correlation coefficient were around 0.9. Nevertheless, the significance levels of the results were not included in the diagram and may therefore differ for each result.

Figure 2. Spider diagram of Pearson’s correlation coefficient on Dengue fever-related GTD and cases in the displayed region and period (Chan et al.^[27]; Husnayain et al.^[26]; Marques-Toledo et al.^[28], Monnaka et al.^[25]). GTD: Google Trends data.

DISCUSSION

In the past five years, especially last year, there has been a growing usage of GTD in prediction models for epidemiological surveillance purposes. Our review aims to identify infodemiology studies to assess the use and applicability of GTD for infectious disease surveillance for low- to middle-income countries. Based on our findings, GTD in general appears to be a promising tool for prediction modeling of emerging infectious disease outbreaks or early warning systems.

GTD was used in five studies on COVID-19, four studies on Dengue, and four studies on Zika. All studies utilized GTD on disease-related keywords, either the disease itself or the disease definition, treatment, or symptoms. The assessed settings were in Latin and South America, India, Indonesia, South Africa, and Syria. In terms of the applicability, overall high correlations between GTD and the disease cases were found. For COVID-19, the study results are controversial. Schnoell^[36], Sousa-Pinto^[37], and Seyyed-Hosseini^[35] found no or negative correlation, which might be due to the examined early phase of the pandemic, the great media attention^[36,37], and the crisis and war in Syria^[35]. Aragon-Ayala^[33] and Satpathy^[34] studied a longer period, later in the pandemic, and found a high correlation between GTD and COVID-19 cases. Regarding the short time frame, it is highly controversial to use Internet search queries during an emerging pandemic of a new disease. Conversely, Dengue fever is a well-known and widely spread disease. Dengue fever was discovered and studied before the Internet was widely accessible. The included studies considered already periods beginning from 2003, when the Internet penetration in the examined countries was still low. Nevertheless, a high correlation was reported over all the studies and different settings. GTD seems to be a suitable indicator for Dengue surveillance. Future studies should include time lag assessments to examine the predictive potential of GTD on Dengue surveillance. Regarding the Zika virus, GTD appeared applicable in surveillance. Since 2016, Zika virus outbreaks have occurred in many different countries and regions worldwide^[38]. The regarded periods in the included studies started from 2014, when Internet penetration and technical developments were high. Overall, high correlations between GTD and cases were reported. The predictive value of GTD was presented, and a time lag of one week was consistent among the included studies that examined time lag correlations. For Dengue fever and Zika virus surveillance, GTD appears to be a useful indicator. In 2019, the fastest growth of subscribed mobile devices (with a majority of smartphones) was seen in Sub-Saharan Africa^[39]. Since mobile devices and the Internet is accessible for the majority of the population, regardless of the economic situation of their country, more Internet search query data will become available. This can be an added value in public-health responses to outbreaks if a framework for data use is developed to ensure the systematic applicability of the data^[39]. For COVID-19, the applicability of GTD in surveillance systems might increase in a couple of years, when Internet searches are more related to the actual infections than to media events. Evaluating a suitable time lag in the regression models is necessary for future studies. The nature of the disease and the dissemination should always be considered, especially in the case of COVID-19 where we know that one can already be infectious without having symptoms or even noticing the disease. The Internet search behavior of the individual cannot function as an early predictor of the disease. Therefore, more infodemiological studies on COVID-19 are needed to allow a robust conclusion on the applicability of GTD.

There are several limitations to our review. The search was limited to one database, and no additional hand search or reference list search was added to the search strategy. Our review might have missed scientific articles from other disciplines than medical research. We did not use setting specific databases (e.g., LILACS or African Journal Online) for our literature research. The language was restricted to English, and we therefore might have missed relevant articles in other languages that regarded low- to middle-income countries. Hence, we did not find any studies investigating African countries. The quality of the included studies was assessed subjectively, and the self-developed tool is not operationalized or evaluated yet. Despite its limitations, this review has some major strengths, including the performed systematic search in Medline, with relevant search terms identified through recent literature. Since the focus was to investigate the use and applicability of GTD for epidemiological surveillance in low- to middle-income countries, it adds another insight and reveals the need for further research regarding these settings. The included quality appraisal added the necessity of standard requirements in infodemiology studies. All the included studies were ecological studies, focusing on Internet search queries and their correlation with infectious disease cases. The included studies were sufficient to show that GTD is used and applicable in Dengue and Zika surveillance. Using GTD should not replace traditional surveillance; rather, it is meant to supplement it when there is insufficient surveillance because of missing data sources. Besides, we included only studies that used GTD specifically. There might have been other useful Internet search query data utilized by studies and applied in surveillance. Therefore, this might have induced selection bias. Additionally, Chinese studies were not included because Google is not available there.

Compared to previous reviews on web-based infectious disease surveillance or infodemiology studies, this review revealed the same opportunities and challenges of using GTD. For example, Choi et al.^[20] evaluated 11 surveillance systems that used online sources, four of which used Internet search queries. They concluded that the advantages of those systems were the low costs and their intuitive and near the real-time adaptable operation. Conversely, they explained the disadvantages of inaccuracy, false predictions, and potential violation of privacy. Milinovich et al.^[6] regarded Internet-based surveillance systems for monitoring emerging infectious diseases, where they also revealed the focus on high-income countries. With regard to the emergence of infectious diseases all over the world, the authors suggested investigating the global surveillance on Internet-based resources further, as Teng et al.^[32] did already for the Zika virus. Nonetheless, surveillance systems on GTD do not provide the same capacity as traditional systems and can rather supplement traditional systems^[6]. Gianfredi et al.^[1] discussed the issue of the unmeasured influence on the search behavior of the public. In times of rapid changing always-available online news, when messages spread fast, the search behavior can be more influenced by media than by actual events. Non-evidence-based rumors and false beliefs might influence the search behavior and the resulting trends, which can have consequences for decision-making in public health^[1]. Barros et al.^[21] underlined this effect, stating that media events significantly affect the reliability of search queries data. The majority of the included studies in their recent systematic review used search queries (and social media) as Internet-based sources. They also underlined the non-generalizability of the data, since the Internet users might be higher educated and younger than the general public. The greatest limitation according to Barros et al.^[21] is the lack of transparency on how the data are obtained by Google.

GTD in epidemiological research can be useful as an additional data source to complement current disease surveillance systems. It should not be considered as the only data source, and interpretations and implementations should be handled with caution. If more research is done on the methodology of the implementation of GTD in the real-world setting, better and more precise implications can be withdrawn.

CONCLUSION

The low costs and availability of GTD remain promising for supplementing future surveillance systems in low- to middle-income countries, but there is an urgent need for a standard methodological framework for the use and application of GTD. The implementation of GTD in surveillance systems for Dengue fever and Zika can be recommended to complement existing systems. The connection of Internet search queries with zoonosis surveillance builds bridges between environmental and human health. High-quality infodemiology studies from and on low- to middle-income settings are needed to receive more relevant and applicable results in these regions.