A bibliometric analysis was performed to understand and explore research on big data in infectious disease modeling and surveillance. The adopted bibliometric methodology involved three main phases: data collection, data analysis, and data visualization and reporting [23].

Regarding data collection, which entails querying and exporting data from selected databases, we queried the Web of Science (WoS) core databases for publications using specific inclusion and exclusion criteria. Compared to other databases, the WoS has been shown to have better quality bibliometric information [23,24] and more excellent coverage of high-impact journals [25]. With the aid of domain knowledge experts from the fields of both big data and epidemiology, we iteratively developed a search strategy and selected the following search terms. The following search string queried all documents’ titles, abstracts, and keywords, and generated 3235 publications in the WoS collection:

AND

Documents not written in English and not peer-reviewed, including editorial materials, letters, meeting abstracts, news items, book reviews, and retracted publications, were removed from the data set given the focus on bibliometric analysis, leaving 3054 documents for the analytic sample (Figure 1).

The 3054 bibliographic data were exported into the R package bibliometrix [23] for analysis. This package was specifically used to conduct performance analysis and science mapping of big data in infectious disease epidemiology. Performance mapping evaluates the production and impact of research constituents, including authors, institutions, countries, and journals. Science mapping examines the relationships between the research constituents by analyzing the topic’s conceptual, intellectual, and social structure.

There are several metrics available for bibliometric analysis. In this study, the primary metrics used for evaluating productivity and influence were the H-index and M-index. The H-index represents the number of published papers h, such that the citation number is at least h [26]. The H-index can be computed for different bibliometric units of analysis: authors, journals, institutions, and countries. The M-index simply adjusts the H-index for the academic age (ie, the number of years since the researcher’s first publication). Other utilized performance analysis metrics were obtained from yearly research output and citation counts. These metrics also contribute to identifying the main themes and the key actors in the research area.

In terms of science mapping, network maps were constructed for some selected bibliographic units of analysis [27]. These networks exhibit frequency distributions of the involved bibliographic data over time. For instance, international collaborations can be explored by assessing same-country publications. A cocitation network analysis was also used to analyze publication references. In addition, using the Louvain clustering algorithm and a greedy optimization technique [28], a co-occurrence analysis was used to understand the conceptual structure of the research area. The basic purpose of co-occurrence analysis is to investigate the link between keywords based on the number of times they appear together in a publication. Notable research topics and over-time trends were detected by generating clusters for author-provided keywords [29]. VOSviewer [30] was used to construct the network visualizations. Each network node represents a research constituent (eg, author, country, institution, article, document source, keyword). The node’s size is proportional to the occurrence frequency of the relevant parameters. The degree of association is represented by the thickness of the link between nodes, and the various colors reflect distinct clusters.

The bibliographic data set comprises 3054 documents from 1600 sources, 14,351 authors, and 121,726 references. From the 3054 documents, 2666 (87.30%) were original research articles and the remaining 388 (12.70%) were review papers. The research output before 2009 was relatively low. The annual publication output during the 27 years (1995-2022) grew steadily, with a yearly growth rate of 26.5%. The publication growth increased steeply between 2013 and 2020 (Figure 2). Table 1 presents the summary statistics of the primary characteristics of these 3054 publications, including the time span and information about documents and authors.

As shown in Table 2, the most productive and influential sources publishing on topics related to big data and infectious diseases epidemiology were Journal of Medical Internet Research and PLoS One (H-index=18), followed by IEEE Access (H-index=13). In terms of productivity, Journal of Medical Internet Research produced a slightly higher number of publications (n=61) than the next best journal PLoS One (n=56). PLoS One had the highest number of total citations at 1893.

As shown in Table 3, the most productive and influential author was Zhang Y (H-index=17), followed by Li X (H-index=13) and Wang J (H-index=12). Wang L had the highest total citations (n=1072), which was substantially higher than the next most impactful author Wang J (total citations=861).

The aim and scope of the top 10 most influential journals, as listed in Table 2, is to publish medical research, medical informatics, or multidisciplinary studies. It can thus be inferred that major future breakthroughs regarding big data in infectious diseases epidemiology will likely appear in these journals.

Figure 3 displays the top 20 most productive institutions. Institutional contributions were assessed by affiliations with at least one author in the publication. Except for the University of California, the top three institutions, which account for 21.3% of the number of publications in the top 20, were medical schools: Harvard Medical School (7.9%) and Icahn School of Medicine at Mount Sinai (6.4%). The other institutions, each accounting for more than 6% of the total, included Columbia University and Oxford University in the top 5, whereas others in the top 20 are research universities: London School of Hygiene and Tropical Medicine focuses on global and public health, Taipei Medical University is medical-based, and Huazhong University of Science and Technology is focused on science and technology. The United States produced the majority of the top 10 most productive institutions, which were in the top 5.

The 20 most productive countries (Figure 4) are led by the United States and China, accounting for more than half (57.3%) of the total publication output. The United States alone accounted for 41.1% of the productivity in this field. The other countries in the top five were the United Kingdom (9.4%), India (4.4%), and Canada (3.3%).

Computer science was the most productive research domain in the bibliographic collection (Figure 5), accounting for 17.6% of the top 10 subject areas. In order of productivity, the other research subjects in the top 5 were public environmental and occupational health (11.4%), health care services (9.6%), medical informatics (9.0%), and engineering (8.8%).

Two major clusters of countries represent the collaboration patterns of the most productive countries (Figure 6). The network was set to include only countries with at least 10 documents, resulting in 50 productive countries. The clustering results demonstrated a demarcation of European countries from the others. For instance, cluster 1 (red) represented most countries from Europe, with England, Germany, and Spain being the core countries. Non-European countries constituted the second cluster (green). The United States and China were the core countries of this group.

Regarding collaboration strength, the United States, with a total link strength of 570, featured the highest number of partners (48), accounting for almost all 50 countries in the network (96%). China, which distantly followed the United States, featured 38 partners and a total link strength of 304. This implies that collaboration is mainly regional.

Figure 7 shows a network map of cocited references in this research area, wherein the node’s size represents the citation strength of the individual studies. The network was set to include only studies with at least 25 citations, resulting in 37 studies. Ginsberg et al [31] published the most highly cited article (185 citations). This 13-year-old study presented a method that used Google search queries to track flu-like illnesses in a population. The second most cited study by Eysenbach [9] introduced the concept of infodemiology, the science of using the internet (eg, social media, search engines, blogs, and websites) to inform public health and public policy. Table 4 further summarizes the top 15 most cited references, including the title, year of publication, number of citations, type of disease, and data source.

The 37 studies in the network map of cocited references produced four thematic clusters (Figure 7); disease monitoring and surveillance (cluster 1), utility of electronic health (or medical) records (cluster 2), methodology framework for infodemiology tools (cluster 3), and machine learning and deep learning methods (cluster 4) were the main topics discussed.

Keyword co-occurrence analysis serves as a supplement to enrich the understanding of the thematic clusters derived from the reference cocitation analysis and helps identify the core topics and contents [29]. As shown in Figure 8, the co-occurrence network displayed 100 relevant keywords after assigning a selection threshold of 10 for the number of keyword occurrences. The top 5 most frequently used keywords were COVID-19, big data, machine learning, coronavirus, and electronic health records.

The 100 author-derived keywords produced four clusters from the coword analysis (Figure 8). Cluster 1 (yellow-green) is related to public health and infectious diseases, with top keywords such as COVID-19, SARS-CoV-2, epidemiology, and epidemics. Cluster 2 (green) is related to electronic storage and delivery of health care, with top keywords including electronic health records, clinical decision support, primary care, epidemiology, and telemedicine. Cluster 3 (blue) involves infodemiology tools, with top keywords including coronavirus, google trends, social media, infodemiology, and surveillance. Cluster 4 (red) is more coherent and broadly related to big data and artificial intelligence, including top keywords big data, machine learning, artificial intelligence, deep learning, and big data analytics.

Further filtering of the top 20 papers was performed to determine if they met the following criteria: (1) addressed at least one infectious disease and (2) utilized a big data source. A review of these 20 papers (summarized in Table 5) was then performed. These selected studies were mainly characterized by papers that utilized novel data sources, including internet search engine data (Google Trends: n=11; Baidu or Weibo index: n=2; Yahoo: n=1) and social media data (Twitter: n=5). Other data sources included electronic health or medical records (n=3) and Tencent migration data (n=1). The most frequently studied diseases were COVID-19 (n=10) [35,36,39,42,45-50], followed by influenza (n=8) [37,40,43,44,51-54]. Only one study considered the Zika virus [55], and another considered the trio of meningitis, legionella pneumonia, and Ebola [56].

Novel big data streams have created interesting opportunities for infectious disease monitoring and control. The review of the top 20 papers suggests the domination of high-volume electronic health records and digital traces such as internet searches and social media. Of note is the relatively increased use of Google Trends. Most studies used Google Trends data by correlating them with official data on disease occurrence, spread, and outbreaks. Some of these studies further adopted nowcasting for disease surveillance. However, using Google Trends for forecasts and predictions in infectious diseases epidemiology fills a gap in the extant literature. Few studies have gone as far as predicting incidents and occurrences, even though data on reported cases of various health concerns and the associated Google Trends data have been correlated in many studies. Predicting the future is hard; hence, more reliable and efficient methodologies are needed for forecasting infectious disease outbreaks.

There are a few drawbacks to digital trace data that should be considered. Many of these data streams miss demographic information such as age and gender, which is essential in almost any epidemiological study. Besides, they represent a growing but still limited population segment, with infants unfeatured and fewer older adults than younger people. Geographic heterogeneity in coverage exists, with underrepresentation in developing countries, although these biases tend to fade and are arguably less pronounced than those found in traditional global surveillance systems. Further, the retrieved data are subject to spatial and temporal uncertainty. Accordingly, hybrid systems that supplement rather than replace conventional surveillance systems as well as improve prospects for accurate infectious disease models and forecasts should be developed.

Most studies, except for those in the United States and China, were conducted in the European context. Thus, more studies need to test the utility of these big data streams for infectious disease epidemiology in the context of more countries, especially in Africa. Future research questions should ask if any cross-cultural differences between countries affect the adoption and use of big data in infectious disease epidemiology.

The vast majority of infectious diseases have a global distribution. Apart from the coronavirus, influenza, Zika, and Ebola virus outbreaks that are featured in our review, the utility of these big data sources for more infectious diseases should be studied.

A few limitations were inherent in our study. First, like any bibliometric study, we are limited by the search terms and database used. This study utilized English publications from the WoS core collection; therefore, relevant publications may have been missed. However, our choice of WoS was informed by its greater coverage of high-impact journals. Second, some studies may have been published after we concluded document extraction. Accordingly, this study does not claim to be exhaustive but rather extensive.

The bibliometric study identified the United States and China as research leaders in this field, with most affiliations from the Harvard Medical School and the University of California. Top authors were Zhang Yi and Li Xingwang. Journal of Medical Internet Research and PLoS One are the most productive and influential journals in this field. Internet searches and social media data are the most utilized data sources. COVID-19 and influenza were the most studied infectious diseases. The main research themes in this area of research were disease monitoring and surveillance, utility of electronic health (or medical) records, methodology framework for infodemiology tools, and machine/deep learning. Most research papers on big data in infectious diseases epidemiology were published in outlets related to computer science, public health, and health care services.

Opportunities for future research are revealed directly from the results of this study. Integrating multiple surveillance platforms, including big data tools, are critical to better understanding pathogen spread. It is also paramount for the research needs to align with a global view of disease risk. The risk of infectious disease is globally shared in an increasingly connected world. The COVID-19 pandemic, including the rapid global circulation of evolved strains, has emphasized the need for an interdisciplinary, collaborative, global framework for infectious disease research and control. There is a need to empower epidemiologists and public health scientists to leverage insights from big data for infectious disease prevention and control.