The coronavirus disease (COVID-19) outbreak finds its roots in Wuhan with potential first observations in late 2019 [1]. On December 30, 2019, clusters of cases of pneumonia of unknown origin were reported to the China National Health Commission. On January 7, 2020, a novel coronavirus (2019-nCoV) was isolated. Two previous outbreaks have taken place since the year 2000 involving coronaviruses: (1) severe acute respiratory syndrome–related coronavirus (SARS-CoV) and (2) Middle East respiratory syndrome–related coronavirus (MERS-CoV) [2].

In the context of the global propagation of the virus, numerous initiatives have occurred mobilizing our current global technological infrastructure: (1) the internet and server farms; (2) the convergence in coding languages; and (3) the use of data, structured and unstructured.

First, the internet is used for exchanges between universities, research laboratories, or political leaders to cite a few examples. Second, the convergence in coding languages, in particular functional languages such as R (R Foundation for Statistical Computing), Python (Python Software Foundation), or Julia, has helped facilitate the communications between researchers. Reproducible research has accelerated in these past few years, with principles such as open science, open data, and open code. The use of data has been amplified by the development of new methodologies within the field of artificial intelligence (AI), allowing researchers to generate and analyze structured and unstructured data in a supervised way, as well as in a semi- or unsupervised way [3]. Data initiatives have flourished across the world, collecting firsthand data, aggregating various data sets, and developing simulation-based models. The Johns Hopkins University of Medicine has made a great effort in terms of data visualizations, which has disseminated throughout the world [4]. In doing so, it has undoubtedly helped to raise citizens’ awareness, helped policy makers inform their population and avoid some potential fake news, or helped correct some misconceptions or confusion. The Johns Hopkins initiative has been followed by several others across the world, creating a large variety and diversity of data sets. This creation process allows researchers to benefit from different levels of granularity when it comes to the data dimensions such as geography, indicators, or methodologies; the open science characteristic is also an interesting aspect of the current research contributions [1].

With better access to these new technologies and methodologies, the breadth of expertise is more extensive; epidemiologists are leading the core of the research, but data scientists, biostatisticians, researchers in humanities, or social scientists can contribute and leverage their domain expertise using converging methodologies such as decision trees, text mining, or network analysis [5].

It is with these hypotheses in mind that we propose to replicate the spirit of the Allen Institute for AI initiative and to design an R package, whose main objective is to integrate easily in a researcher's workflow. This package is named EpiBibR.

EpiBibR stands for an “epidemiology-based bibliography for R.” The R package is under the Massachusetts Institute of Technology license and, as such, is a free resource based on the open science principles (reproducible research, open data, and open code). The resource may be used by researchers whose domain is scientometrics but also by researchers from other disciplines. For instance, the scientific community in AI and data science may use this package to accelerate new research insights about COVID-19. The package follows the methodology put in place by the Allen Institute and its partners [6] to create the CORD-19 data set with some differences. The latter is accessible through downloads of subsets or a representational state transfer application programming interface. The data provide essential information such as authors, methods, data, and citations to make it easier for researchers to find relevant contributions to their research questions. Our package proposes 22 features for the 60,500 references (on June 26, 2020), and access to the data has been made as easy as possible to integrate efficiently in almost any researcher's pipeline.

Through this package, a researcher can connect the data to her research protocol based on the R language. With this workflow in mind, a researcher can save time on collecting data and can use an accessible language to perform complex analytical tasks, for instance, be it in R or Python. Indeed, it is usual that researchers use multiple languages (functional or not) to produce specific outputs. This workflow opens these data to analyses from the most extensive spectrum of potential options, enhancing multidisciplinary approaches applied to these data (biostatistics, bibliometrics, and text mining, among others).

The goal of this package in this emergency context is to limit the references to the medical domain (here the PubMed repository) but to then leverage the methodologies used across different disciplines. As we will address this point later, a further extension could be to add references from other disciplines to not only benefit from the wealth of methodologies but also their theories and concepts. For instance, to assess the spread of the disease, the literature—and theories—from researchers in demography would certainly be relevant.

In what follows, a use case about how we can use such a data set is proposed. This section is not intended to offer a complete systematic literature review of the 60,500 references. This is the purpose of another article. However, we want to illustrate some powerful techniques that can be applied to this collection of references (for instance, social network analysis) while remaining at a very high level [12].

A systematic literature review consists of mainly four stages: (1) planning, (2) conducting, (3) analysis, and (4) synthesis and reporting. In the first stage, a preliminary study aims to build a corpus of articles citing the most relevant articles in the domain. The second stage is about producing a general review of the main topics used in the corpus. The third stage here is about making a cocitation analysis of the references in each corpus article. The last stage is about proposing a keywords co-occurrence analysis [13].

In our context, we propose a slightly modified perspective on a systematic literature review. The first stage is here at a different scale since we collect not just a few relevant articles to create the corpus but a vast, almost exhaustive, list of articles on a topic. It is worth noticing that the process of data selection consists more of what we define as an “algorithmic systematic literature review,” sometimes also referred to as “automation” [14]. An algorithmic systematic literature review comes with lots of benefits. The proposed modified systematic literature review improves the more classical approach since it does not rely on a manual search and extraction, with the potential biases and limitations it might create. An algorithmic systematic literature review combines the strength of both approaches: the power of big data with the academic soundness of the systematic literature review process. As such, it does not replace the expert’s analysis of the literature. On the contrary, it should be used to augment the expert’s study [3].

The bibliometrix package allows a thorough bibliometric analysis using R. Our EpiBibR data have been designed to integrate easily with the bibliometrix package. A shiny application is also available, called biblioshiny() [11]. This package has been used extensively for various exercises mobilizing massive amounts of data [15].

Let us first propose a simple count of the references on the coronaviruses literature. In Figure 1, the historical development of research on coronaviruses can be analyzed as having three stages: exploration, initial development, and rapid development in the past year and the current year.

In 2019, a little more than 1000 papers on coronaviruses had been produced, and in the first 5 months of 2020, close to 3400 papers were written on the topic. Those papers from the past 2 years seem to be an interesting, statistically representative sample. In Figure 1, we can also highlight the most productive authors as well as the most productive countries in terms of absolute counts. The most prolific authors provide interesting statistics since it is most likely to proxy research labs. In doing so, we can find which teams are working on which aspect of the coronaviruses. To illustrate our latter point, in Figure 2, we first propose a compelling visualization, called a Sankey diagram. It links authors, keywords, and sources on a connected map. It is the first way to create groups of researchers. The results could be used by policy makers to identify areas of research on this topic.

We can also use powerful techniques such as Social Network Theory to find potential clusters of topics, clusters of researchers, and groups of country collaborations. Figure 3 is an example of the latter. The United States and China produce the bulk of the research on coronaviruses.

Figure 4 is an illustration of author collaboration networks. As previously mentioned, we remain at a very high level here. However, policy makers or public health officers, for instance, could use these techniques to find more granular networks either just within the 60,500 references or by crossing with other databases. We could even imagine crossing with unstructured data for some specific purposes [16].

Figure 5 is about finding clusters of topics. This technique can be applied to subsamples of the 60,500 references to provide a more granular analysis.

Figure 6 proposes indeed to go further on the topic dimension. For instance, we can study the evolution over time of the author’s keywords usages.

As previously mentioned, this section is not intended to cover a systematic literature review but rather to illustrate some of the potential uses of powerful techniques or highlight some of the interesting questions that can be addressed. With a data set of 60,500 references, we have access to a wealth of information. Combined with the state of the technological development we have access to today, multiple questions can be answered. To go further, we could imagine analyzing document cocitations or reference burst detections [17].