Science Map of Cochrane Systematic Reviews Receiving the Most Altmetric Attention Score: A Network Analysis

The present study aimed to analyze and visualize the science map of Cochrane systematic reviews (CSRs) with high Altmetric attention score (AAS). On 2020-07-29, the altmetric data of the Cochrane Database of Systematic Reviews were obtained from the Altmetric database (Altmetric LLP, London, UK). Bibliometric data of the top 5% AAS of CSRs were extracted from the Web of Science. Keyword co-occurrence, co-authorship and co-citation network analyses were then employed using VOSviewer software. The random forest model was used to rank the importance of the altmetric resource. A total of 11222 CSRs with AAS were found (Total mentions: 305265), with Twitter being the most popular Altmetric resource. Consequently, the top 5% AAS (649 articles, mean AAS: 204.95, 95% conﬁdence level: 18.95, mean citations: 123.68, 95% conﬁdence level: 13.9) were included. Density mapping revealed female, adult and child as the most popular author keywords. According to network visualization, Helen V. Worthington (University of Manchester, Manchester, UK), the University of Oxford and UK had the greatest impact on the network at the author, organization and country levels respectively. AAS were weekly correlated with citations (r s =0.21) although citations were moderately correlated with policy document and blog mentions (r s =0.46 and r s =0.43). Cochrane systematic reviews received high levels of online attention, particularly in the Twittersphere and mostly from the UK. However, CSRs were rarely publicized and discussed using recently developed academic tools, such as F1000 prime, Publons and PubPeer.

Thi s v e r sio n is b ei n g m a d e a v ail a bl e in a c c o r d a n c e wit h p u blis h e r p olici e s. S e e h t t p://o r c a . cf. a c. u k/ p olici e s. h t ml fo r u s a g e p olici e s. Co py ri g h t a n d m o r al ri g h t s fo r p u blic a tio n s m a d e a v ail a bl e in ORCA a r e r e t ai n e d by t h e c o py ri g h t h ol d e r s .

INTRODUCTION
Cochrane is a British charity founded by Iain Chalmers in 1993. The organization was created specifically to manage medical research findings in order to facilitate the evidencebased choices in health interventions faced by health professionals, patients, health policymakers, as well as those interested in health to make informed decisions about health promotion. Cochrane includes 53 review groups from 130 countries. [1] Alternative metrics, abbreviated to altmetrics, is an emerging academic tool that measures the online attention surrounding scientific research outputs. [2][3][4] It complements, but does not replace, the traditional citation-based metrics. [5] Altmetric data resources include Twitter, Facebook (mentions on public pages only), Google+, Wikipedia, news stories, scientific blogs, policy documents, patents, post-publication peer reviews (Faculty of 1,000 Prime, PubPeer), Weibo, Reddit, Pinterest, YouTube, online reference managers (Mendeley and CiteULike) and sites running Stack Exchange (Q&A). [6] In comparison with traditional citation-based metrics, altmetric data resources are updated rapidly. A recent bibliometric analysis revealed that only 50% of articles were cited in the first three years after publication. [7] In contrast, several altmetric data resources are updated in real-time (e.g. Twitter and Wikipedia) or on a daily-basis (e.g. Facebook).
Research funders and charities such as the Welcome Trust and John Templeton Foundation pay attention to altmetric analysis. [8] A study on the influence of the alcohol industry on alcohol policy would be a good example of their engagement. [9] The study was supported by the Welcome Trust, which invests approximately £600 million (US$ 936 million) a year in research, and alleged that several submissions to the Scottish government misrepresented research outputs so as to support policies preferred by the alcohol industry. Three months following this publication in PLOS medicine, it remained without citation. Yet, altmetrics enabled the Welcome Trust to understand that this article had been tweeted by the key influencers, including members of the European Parliament, international non-governmental organizations and a sector manager for Health, Nutrition and Population at the World Bank to reveal its global impact on the policy sphere. [8] Objectives and Scope of the study Altmetrics research is a growing field in medical sciences. [10][11][12][13] Online attention surrounding medical articles have been assessed in many fields of medical sciences e.g. neurology and cardiovascular sciences. [14,15] It is well-known that the Cochrane Systematic Reviews (CSRs) do a great role in evidence-based medicine. Bibliometrics of CSRs have been analyzed and discussed previously. [16] To our knowledge altmetrics of CSRs has not been analyzed. Here we aimed to analyze and visualize the knowledge structure of CSRs with high Altmetric attention score (AAS) to discover the hot topics and influential researchers, institutions and countries. Furthermore, we planned to assess influential factors regarding AAS and the number of citations via machine learning.

METHODOLOGY
On 2020-07-29, the altmetric data associated with the Cochrane Database of Systematic Reviews (ISSN: 13616137, 1469493X) (CDSRs) were obtained from the Altmetric database (Altmetric LLP, London, UK). The bibliometric data of the top 5% CSRs with the highest AAS were extracted from the Web of Science using their DOI. The bibliometric data were imported as a UTF-8 encoded txt files to VOSviewer 13.6.1 software (http://www.vosviewer.com/, Centre for Science and Technology Studies, Leiden University, Leiden, the Netherlands) for author keyword co-occurrence, coauthorship, co-citation network analyses and density visualization. [17] The Spearman correlation coefficient was applied to identify any correlation between the main altmetric resources and the number of citations. The source of citation counts was Dimensions database. [18] Principal component analysis (based on a correlation matrix) was used to represent the data in smaller components than the dataset eventually consisted of. Chi-squared tests were used to test the fitness of the model.
The random forest regression (a machine learning algorithm) was conducted to rank the influential factors affecting the AAS and the number of citations. The random forest is made up of several random decision trees which work together like a forest. Each single decision tree is a flowchart-like model in which each internal node represents a "test" on an attribute (e.g. whether the number of tweets < 1100) to make classifications (Appendix 1). The concluding forecasts of the Appendix 1: Example of a decision tree. In this study to create a random forest model, 500 random decision trees were made by random sampling of training data points and random subsets of features applied when splitting nodes.
random forest are made by averaging the forecasts of each individual tree. [19] Data analysis was carried out using R 3.6.3 software (R Foundation for Statistical Computing, Vienna, Austria). Excel 2016 was used to draw graphs and data-bar visualizations.

RESULTS
A total of 11222 CSRs with AAS were found (total mentions: 305265). Twitter was the most popular altmetric resource ( Figure 1) with Tweets originating mainly from the UK (47,897 (18.7%)), Spain (22,953 (9%)) and the US (20,513 (8%)). In addition, @CochraneUK (4,886 total mentions from this Twitter user) was the most active altmetric resource, followed by @CochraneLibrary (1,952 total mentions from . Hot topics among author keywords of the top 5% Cochrane systematic reviews receiving the most altmetric attention. The lower part zoomed on central hot zones. The distance-based approach was used to create this map, which means the smaller the distance between two terms, the higher their relatedness. ( Figure 4 and 5). The co-citation network analysis revealed that the Lancet and CDSRs had the greatest influence on the network ( Figure 6).
AAS were weekly correlated with citations (r s =0.21, 95% confidence interval 0.14 to 0.29, p<0.001) (Figure 7  13.9) with the highest AAS were included in the study. The bibliometric data of 599 articles found in the Web of Science were further analyzed. "Vitamin C for preventing and treating the common cold" had the highest AAS (score: 2880) among the CSRs (Table 1). Density visualization indicated that "female", "adult" and "child" were the most popular keywords used by authors ( Figure 2). Moreover, the network analysis at author level revealed Helen V. Worthington       Results from the random forest model confirmed the importance of policy document mentions regarding citations. News and Twitter were the most influential factors regarding altmetric scores (Figure 9).

DISCUSSION
It is widely believed that CSRs are one of the most important resources in evidence-based clinical decision-making. [20] In the present study, we analyzed the social impact of this reliable medical evidence using altmetrics. The number of social media users has been estimated to escalate to 3.09 billion by 2021. [21] The primary source of altmetric for CSRs was Twitter, a microblogging service with 330 million active users [22] that can act as an important social media to broadcast and find academic information all over the globe. [23][24][25] Several well-known academic healthcare providers such as the Mayo Clinic also use social media (@MayoClinic with 1.92 million followers and 47600 tweets) to communicate with peers and patients to disseminate trusted medical information. [26] Although the geographical breakdown of Tweets demonstrated they were mainly posted from the UK, the demographic characteristics, social class and occupation of those who Tweeted about CSRs could not be obtained for technical restrictions. In addition, Facebook was ranked the second altmetric resource for CSRs; however; these reviews were not too popular in this social media. Facebook could play an effective role in the rapid distribution of research findings via acknowledgement of substantially active users as well as tools (e.g. chat) that may enable a direct interaction between researchers and target audiences such as peers and patients. [27] Despite the popularity of CSRs in the Twittersphere, they were rarely shared and discussed within the emerging academic tools such as F1000 prime, Publons and PubPeer. The post-publication peer review tools allow the scientific community the ability to quickly identify misconduct, fraud, or error in the context of research. F1000 prime recommendations also help scientists to discover key documents in the medical field. [28] Hence, considering the advantages of these new academic tools, both researchers and publishers should pay more attention to this issue. Although the English medical articles on Wikipedia received more than 2.4 billion official visits in 2017, [29,30] the overall Wikipedia mentions among the CSRs were low considering the established partnership between Wikipedia and Cochrane collaboration. [31] Of more interest, CSRs were widely cited among the policy documents, which is promising considering the principles of evidencebased policymaking have an undeviating impact on the health of communities. [32] It has been reported that journals with their own Twitter account get 34 percent more citations and 46 percent more tweets than journals without a Twitter account. [33] Of more interest, the Cochrane organization and related groups such as @cochranecollab (103.8K followers and 13.7K Tweets), @CochraneUK (54.4K followers and 37K Tweets), @ CochraneLibrary (70K followers and 7191 Tweets), @ CochraneCanada (5572 followers and 6462 Tweets) were active in the Twittersphere. Among the 11222 Cochrane systematic reviews included in this study, the total number of citations was 80272. With respect to the ratio of number  of Twitter followers (≈233772) to the number of citations (80272), the Cochrane institution must be aware of becoming a Kardashian organization (i.e. overactive in the Twittersphere). [34] Human knowledge in the medical field is rising incredibly in this information age, so scientists are fronting overloaded information (also known as infobesity, infoxication and information explosion), which needs to be handled carefully.
In addition to the enormous data and text analysis methods, science mapping makes it possible to review this abundant amount of data effortlessly. Density visualization map of CSRs revealed the newly emerging and innovative concepts, e.g. genomic medicine, nanotechnology and artificial intelligence, were not included among the hot topics. This finding may be explained by the retrospective nature of CSRs which was also reported in the field of Dentistry, [13,[35][36][37] Endodontology [38] and Iranian medical journals. [39] Citation analysis is considered an important part of scientometric research. In contrast to the conventional methods of classic citation analysis, which provide limited information, co-citation network analysis can illustrate patterns, trends and association through visualization of a set of multiplex citation data. The co-citation network visualization disclosed numerous well-respected journals in the network, in which the Lancet and CDSRs had the most influence on the network. This finding may be explained by the Lancet-Cochrane collaborate on systematic reviews. [40] Despite the fact that popular CSRs received a high citation rate, no significant correlation was observed between the number of citations and AAS. Likewise, no significant correlation was reported between the number of citations and AAS for the research articles published in high-impact general medicine journals, [41] the cardiovascular field, [42] radiology [43] and dentistry. [44,45] In contrast, a large-scale survey indicated a significant correlation between six altmetric resources (tweets, Facebook wall posts, research highlights, blog mentions, mainstream media mentions and forum posts) and citation counts. [46] Among the six specialized PloS journals, a significantly positive correlation was found between the normalized AAS and normalized citation counts. [47] Above all, the correlation between AAS and citations is a controversal issue that needs further research and meta-analysis.
In this study, the random forest method was used to find influential factors regarding AAS and the number of citations.
This machine learning approach is a newly growing concept in biomedical research. Pubmed query with the term "random forest " revealed 7,177 results. Trendline analysis showed number of articles in this regard are growing fast (Number of articles = 11.379y 2 -45679y + 5E+07, R² = 0.94, y=year). This rapid growth may be explained by advantages of this method such as handling different dimensions (features) of the data, judgements about the importance of the feature, being relatively simple to implement, quick prediction and training speed, low bias, etc. [48,49] The limiations of the present study should be considered. AAS may fluctuate over time in contrast to conventional bibliometrics. Moreover, altmetric data of some distinguished social media such as ResearchGate, Instagram and LinkedIn could not be retrieved because of technical limitations. In addition, some altmetric resources may be prone to gaming; for example, fake accounts or robots may increase the number of Tweets more than the reality. [50] Lastly, censorship and restrction of some altmetric resources, particularly Twitter, in some countries may limit the estimation of the social impact of CSRs. [13] CONCLUSION The top 5% CSRs had a high AAS. Twitter was the most popular resource. Tweets generally originated from the UK. Popular CSRs received a high citation rate; however, there was no significant correlation between the citation counts and AAS.

ACKNOWLEDGEMENT
We would like to thank Altmetric LLP (London, UK), particularly Mrs. Stacy Konkiel for giving us valued support and permitting us to have complete access to altmetric data.