Creation, dissemination, and management of knowledge are cornerstones of safe and effective health care, but achieving these goals in functional and successful large care systems remains enormously challenging. Within this construct of health systems [1], the concept of a network is vital for improving patient care by fostering collaboration, stimulating engagement and knowledge acquisition and management, and promoting learning [2]. Recently, social media-based platforms have been proffered as educational tools [3,4], based on the critical assumption that activity in online networks can lead to the emergence of personal learning networks [5] and communities of practice (CoPs) [6].

CoPs are an anthropological concept, defined as a group of people who organize around a specific component of knowledge (eg, a profession or particular task). They create, organize, and share information and not only develop specific domains but also foster development of individual members [6]. These communities are organized around a domain, community, and practice, where the domain is a realm of knowledge that the members of the group have interest in and value as important; the community is the coalescence of interactions and negotiations between members around the domain; and practice is the results of the cooperation leading to the creation of tangible resources affecting their practice of the knowledge. The presumed natural evolution of organically organized online communities [7] is the description and identification of an online CoP (oCoP) [8]. This construct offers an enormous potential for the creation, dissemination, and management of knowledge and people using a relatively minimal amount of resources and appearing as an ideal framework for information-oriented teams. These have arisen from the Internet in general and particularly in social media platforms [9] because of their ability to efficiently connect and engage groups, provide peer-based education [10], and manage meaningful knowledge [8].

The concept of CoP has been thoroughly described in health care [11], and the critical components of an appropriate network were refined and expanded by Aveling et al in 2012 [12]. A health care CoP is defined by 8 core characteristics (described below) that are themselves derived from Wenger’s [6] pivotal components of community, domain, and practice.

Although CoPs are not a new phenomenon, social media platforms have enabled them to form rapidly and across international boundaries, with nonlinearity and on a large scale [13]. These new oCoPs create discrete and quantifiable data flow among users (nodes) of the network. Knowledge and relationships develop by these interactions (links), with functional communities arising as information transfer increases. These communities of online learning and practice offer an open digital space for information sharing, with a flat role hierarchy, strong group identity, high engagement, and rapid flow of information, content curation, and knowledge translation. In many ways, their topology represents the structure of scale-free networks [14,15]. The use of social media platforms as an educational medium is dependent on these oCoPs providing reliable and manageable information for their participants [16].

Despite the recognition of several possible candidates for CoPs and their derivatives in online platforms [8], particularly in the health care sector [11] and for peer-to-peer patient interactions [17], medical learning and practice oCoPs have been rarely described in the medical literature to date. An international movement that began within the past 5 years now aims to collaborate to create, curate, and disseminate medical knowledge with the intention of changing patterns of care; an explicit aim is to reduce the knowledge translation gap [2,15]. This movement describes the concept as “free open access medical education” (FOAM) [18,19] and organizes around the Twitter hashtag #FOAMed; the group name is synonymous with and represents its core philosophy.

The strength of oCoPs can be established if these communities are identified and their relationships described; social media education could then be validated as a viable paradigm for knowledge translation [16] that affects health care practice. The aim of this study was to show preliminary evidence of the emergence of a health care oCoP focused on knowledge translation and organized as a network architecture around the Twitter hashtag #FOAMed. By using data extracted from our #FOAMed hashtag analysis, we aimed to demonstrate that the FOAM community fulfills the core characteristics of a CoP, as described by Wenger [6] and expanded by Aveling et al [12], and is a community of individuals with different roles who belong to different organizations and places, who organize themselves around a concrete domain and specific goal, and who function to support and promote each other’s development to achieve change.

We hypothesized that the FOAM community, defined by the explicit interactions that include the #FOAMed hashtag, constitutes an oCoP. To confirm this, we believed it should be possible to recognize the emergence of the community, domain, and practice of the network, while describing information flowing between nodes and the semantic relationships between the members. Describing the structure of the network, as well as the influences of nodes outside the network (eg, spambots), requires a robust intrinsic familiarity with the community; therefore, we approached the analysis using a netnographic methodology [20,21] based on our personal participation in the FOAM movement.

The challenge of analyzing billions of interactions in social media [22,23] is a relatively recent area of activity that requires a high level of expertise and computational power. We created a Symplur Signals database (Symplur LLC, Upland, CA, USA) around the Twitter hashtag #FOAMed and interrogated it from March 1, 2013, through August 31, 2015; this dataset contained almost all data available for the hashtag because the accepted inception time for #FOAMed was March 2013 and the activity occurred explicitly around the hashtag [18]. The computational tools implemented by Symplur Signals have been validated previously for the analysis of Tweet chats [24]. There is a large amount of FOAM community activity, such as blogs, podcasts, and conference proceedings, that is not directly related to the #FOAMed hashtag or Twitter; however, these products are typically references in community activity and it is highly unlikely that such activity outside of Twitter would nullify the existence of the oCoP.

To reduce potential bias in the analysis, we submitted a protocol to the Symplur Signals team defining the a priori analytic strategy. Within this protocol, we proposed several proofs to demonstrate concordance with the Aveling-Wenger postulate for a CoP. Table 1 lists these proofs, matched to the 8 distinct core characteristics. For example, evidence of interdependent groups and individuals (A1C) would be demonstrated by showing that the #FOAMed hashtag was used by individuals who had interactions with each other and that, over time, influencers increased in number, becoming divergent rather than convergent. To do this, we would therefore need to describe #FOAMed hashtag use quantitatively. Conceptually, we aimed to analyze #FOAMed hashtag activity as a network. The hashtag [25] constitutes the core scaffold in which the identified social activity occurs. The network consists of Twitter users (nodes), and interactions (links) with specific metrics of interactions such as retweets, favorites, and engagements, which constitute the connections and distribution by which networks are typically constructed [26,27].

Given the social nature of the Twitter platform, we used the social network analysis component of the netnographic framework for analysis, as described by Kozinets [20]; this method is particularly effective at identifying external influences that are not meaningful to the community (eg, spambots, large commercial interests), as well identifying digital artefacts that are created by mistake or are otherwise inappropriate (eg, links to unrelated events associated with the hashtag but used by a completely different community). From a netnographic perspective, it is difficult to appreciate the metrics of centrality without a personal understanding of the symbolism, meaning, and consumer patterns of the groups [28]; however, as members of the community of study, we had an immersive and descriptive understanding of the FOAM movement. Our intimate understanding allowed us to determine (via a consensus view of the authors) the relevance of the top 100 shared links identified in the ethnographic analysis of the FOAM community We generally excluded broken links, links to landing pages of websites or home pages, and any other links that did not point to a webpage with a publicly visible forum for comments and feedback or had no comments posted. This study was deemed exempt from ethics board review by the English National Health Service Health Research Authority.

During the study period, we identified 49,459 active users who issued 429,606 tweets and created 1,258,692,900 impressions (ie, the number of times a user is served a Tweet in their timeline or search results); this translated to more than 8 tweets and 25,000 impressions per user involved in the conversation (Table 2). The user and hashtag activity, expressed as number of daily tweets, increased during the study period (from ~250 to ~700 tweets/day) (Figure 1). User participation increased substantially during the same time frame, rising from around 1000 users to more than 45,000; an additional 1000 users per month have participated since March 2014 (Figure 2). Multimedia Appendix 1 shows the top influencers. In terms of engagement, 27,635 users (55.71%) participated with 1 tweet (representing 6.43% of all tweets), while 2603 users (5.25%) tweeted more than 20 times and were responsible for 72.45% of all tweets (Table 3).

From a network distribution perspective, the top 200 users created 148,185 tweets, representing 34.49% of all tweets of the community. The Symplur Signals analytics algorithm created a centrality map by considering nodes (users) and their interactions (links) to assign weight and used the metrics of mentions, authority, and hub to assign the weight of centrality [29,30]. The 100 users with the highest weights also tended to have high authority and edge scores (Multimedia Appendix 2), and from a netnographic perspective, they were recognizable as community leaders. Figure 3 outlines their status in relation to the number of interactions with other participants in a graphical representation, where the size of the node represents its network centrality (relative influence) and the links depict the strength and directions of the interactions between members (nodes) of the network. Also, Figure 4 depicts a perspective of the communication between nodes on a conversation matrix showing the frequency and strength of interactions among the most active community members.

In terms of geographic distribution, the community members generally lived in anglophone Western countries, although some of them were located in Latin American and Asian countries (Table 4). However, the analytic tool was unable to consider the geographic location of more than 50% of users because they did not provide self-identified geolocation information.

Multimedia Appendix 3 details the top 150 conversations (defined by Symplur in terms of level of engagement) generated in the community during the study period. The users involved were those with the highest network weight and hub index. The conversations showed a relatively low level of branching, with most having 2 or 3 levels of interactions (replies). An automatic sentiment analysis [31] showed a 0.367 positive sentiment embedded in the tweets. In terms of knowledge sharing, an analysis of the top 100 links (Multimedia Appendix 4) shared by the group shows a mix of self-created content (blogs), referral to social platforms (eg, YouTube, Vine), and referral to traditional media (eg, website for the New England Journal of Medicine). We analyzed a total of 26 blogs (Multimedia Appendix 5) to explore the contextual meaning of the #FOAMed community. Ethnographic examination showed forums of discussion by way of blog comments, with commenters also being members of the #FOAMed Twitter community.