How advocacy affects Twitter migraine conversations: A pilot cross-sectional survey of Northeast American “migraine” landscape on Twitter from May to June 2020

Data gathering methods

We designed a Twitter robot that records all unique tweets containing the word “migraine” from May 8th, 2020 to June 23rd, 2020 and within 400 km of New Brunswick, a city in the state of New Jersey in the United States. (New Brunswick is the location of the authors’ home institution, Rutgers University.) We have decided on a 400 km radius so as to include major cites between Boston to Washington DC and therefore a number of established Headache Centers in the Northeastern United States. We posted the number of tweets, as well as a rolling average of these tweets, on @MigraineTweets. This robot is implemented through proprietary software designed by the author with Python, a programming language and Tweepy, a Python library. ¹² It is implemented continuously on a Raspberry Pi 3 system. While actual source codes of our Twitter bot is proprietary, implementations of similar Twitter bots are not uncommon and can be found at a number of online tutorials. ^13,14

Data analysis methods

Once data recording has been completed, all tweets are made into lower case and filtered to have only alphanumeric characters as well as space and “#.” Each tweet is then made to occupy a single line in our database. This resultant dataset is then used for data processing. A similar filtering technique is done for websites and hashtags; if a hashtag and a website is mentioned together in a single tweet then the two are documented as connected in our database.

Twitter hashtags are often used together. As such, co-occurrence of hashtags can be interpreted as forming a social/linguistic content network. ¹⁵ To answer the question “What do people tweet about when they tweet about migraine/headaches?” we model this content network as a network graph: each hashtag in our data set can be interpreted as a node, if one hashtags is on the same tweet as another then an edge exists between them.

To assess how topics changes or remain constant despite major advocacy movements, we generated two network graphs according to the method described above, one for May and the other for June. Once the two models were generated, we calculated the most important 30 hashtags for each model through traditional network analysis methods of centrality measurements. (In a network graph, centrality measures are mathematical methods of measuring which nodes are the most important. ⁶

We used three commonly utilized centrality measurements to assess hashtags relationships on a network scale: Degree centrality allows us to rank the most “popular” hashtags. Between-ness centrality allows us to find the “brokers” and key “bridges” in a network. (In network theory, the concept of “brokers” and “bridges” are near synonymous terms used to describe actors or nodes in a network that connects two otherwise separate groups within that network; these actors or nodes therefore facilitates information exchange between groups within the network. ¹⁶ Finally, closeness centrality allows us to find topics that are closest to all of the members in a network. We obtained all three centrality measurements for our models.

We are also interested in how Twitter topics of conversation cluster together. We applied Hartuv and Shamir’s highly connected subgraph (HCS) as a way of determining subgroups within the graph. ¹⁷ We hypothesize that this clustering algorithm will help us to differentiate dominant versus marginalized conversations.

Finally, we would like to know from where Twitter users obtain their sources of information. To accomplish this, we extracted all website domains from http links referenced by our tweets. We further hypothesize that individual subgroups of hashtags may contain references to different sources of websites. In other words, different subgroups obtain their sources of information differently. To test this hypothesis, we group websites in accordance with their HCS subgroups.

Formatting of the above data set for analysis as well as access of websites are done through Haskell, a programming language. Clojure, with the use of http library clj-http, is used for accessing website references. ¹⁸ Data analysis for network analysis, including HCS and clique, are done through Haskell and Python with public domain codes and also Networkx, a network analysis library. ^19,20

Ethical considerations

Our project is considered a big data project. Since individual user data is not tracked nor recorded, an Internal Review Board is therefore not necessary in accordance with big data research standards. ²¹ To ensure legal compliance, the authors of this study have also disclosed our bot to Twitter platform explicitly. ²²

Genealogy of our methodology

Our study is not the first studies of social media platform in health care. Indeed, infodemiology as coined by Gunther Eysenbach, refers to “the study of the determinants and distributions of health information and misinformation” and forms a discipline of its own popularized through peer review journals such as Journal of Medical Internet Research. ²³ A recent editorial in Cephalalgia noted that digital epidemiology may emerge as a new influential research field in headache. Indeed, recent analyses of YouTube and Google as sources of information on migraine demonstrate that there is an interest from patients and healthcare providers alike. ^{24 –26}

In regard to the study of the migraine conversation on Twitter, we know of five prior studies: 1) Callister et al. had conducted a cross sectional analysis of Twitter activities during American Headache Society (AHS) meeting. ⁸ 2) Nascimento et al. conducted an infodemological study of the migraine space in Twitter with primary focus on word frequencies and attention to descriptions of “migraine impact,” “pain descriptors,” as well as geographical and temporal patterns. ⁷ 3) Deng et al. has recently published a study of migraine Twitter feeds using natural language processing techniques including sentiment analysis and latent Dirichlet analysis. ⁹ 4) Mullins et al. have studied topics and keywords in a randomly chosen 14-day periods in 2017 in Ireland. ¹⁰ 5) Finally, in a survey of 20,301 tweets from September 30, 2019 to January 14, 2020 we showed that topics of “#CBD,” “#cannabis,” as well as “#chronicillness,” “#chronic pain” and “#spoonie” were some of the most commonly tweeted hashtags. ¹¹

Our study takes a different methodological approach from most of the above studies: Utilizing hashtag, a built-in user topic classification system, we use network analysis to analyze migraine conversations. Application of network analysis to hashtags in social media is a well-established technique, with uses ranging from political science, to marketing research, to academic research. ^{27 –29} Indeed, our study has as a methodological precursor in Tighe et al.’s analysis of Twitter hashtag on the word “pain.” In the Tighe study individual word are considered nodes and, similar to our project, edge represents co-occurring words in the same tweet. ¹⁵ Tighe et al.’s study, unfortunately, is not specific to migraine and yields no major thesis on headache advocacy. Finally, we have previously submitted preliminary results of a longitudinal study as a presentation at the AHS Virtual Meeting. ¹¹ Our current iteration improved upon our prior approach and further taken into account, as a topic of main interest, advocacy and sources of information.

Evolving centralities, clustering and its implications

In considering the answer to the question “What do people talk about when they talk about migraine?” our study appears to support some of the findings in previous study—tweets on chronic pain and chronic illness often occur alongside migraine tweets. ¹¹ This finding is again supported this study, where in both May and June, chronic pain hashtags dominate all centrality measurements besides migraine tweets.

The addition of HCS technique for subgroups analysis allows for a more nuanced answer to our question on topics of migraine discussion on Twitter. First, as evidence by Table 5, the migraine conversation appears to be largely composed of one big conversation with a number of smaller marginal conversations. The largest subgroup is not a homogenous discussion on one topic (as they often are for the smaller subgroups) but rather a cluster of closely related topics. For example, as expected, #migraine, #headaches as well as any migraine related conversations (#migrainetips, #chronicmigraine, etc. or migraine treatment such as #botox, #aimovig, #nsaids etc.) is a part of the largest conversation. (This is supported by Mullin et al.’s study, where migraine tweets can be categorized as either “awareness,” “advice,” or “research” oriented. 10) However other diseases such as other headache disorders (#clusterheadaches, #csfleak, #tensionheadache, #trigeminalneuralgia), mental disorders (#anxiety, #depression) and chronic pain tweets—even specific subtypes of chronic pain such as #lupuswarriors or #crps—are also parts of this dominant conversation. We suspect that this reflects the comorbidities associated with migraine as well as the closely considered alternative diagnoses often associated with a migraine diagnosis. ^30,31

Parts of this larger conversation, surprisingly, also includes seemingly unrelated topics such as #gradschool, #writinglife, #cartoon, #cooking, #weekend, #workout and #covid or #quarantinelife. We speculate that these topics arise organically as migraine patients comment on how the illness affects important aspects of their lives. ^30,32 The dominant conversation also includes topics that are rather technical—for example, #eeg, #pfo, #cpp, #fdaapproval, and #telemedicine—we suspect these describe headache practitioners’ perspectives.

Our project demonstrates that advocacy can successfully change the way in which hashtag topics on Twitter are interconnected; specifically, during the month of Migraine Awareness Month, advocacy related tweets increase in frequency (Table 1) and popularity (Table 2) among “migraine” tweets. Advocacy tweets, do not, however, dramatically change as brokers of ideas (Table 3). Taken together, this implies that migraine advocacy movements successfully establish themselves as dominant topics of conversations during the month of June; when people talk about migraine in the month of June, they do indeed talk about Migraine Awareness. These conclusions are, in some sense, expected and are reassuring results for migraine advocates.

Advocacy also appears to bring in new topics of conversation into the migraine space. Between the month of May and June, migraine advocacy related tweets are among the dominant subgroup (that is, the largest subgroup as judged by HCS) and the number of distinct members in this subgroup also increased from May to June. While multiple forces can be at play (for example, AHS virtual meeting, which began on June 13th and remained online until August 24th, and Black Lives Matter movements, triggered by the murder of George Floyd in May) this does seem to imply that migraine advocacy draws in new topics into the migraine conversation. ^{33 –35}

Migraine advocacy, however, do not necessarily make new topics popular with the exception of non-pharmacological treatments such as physical therapy, acupuncture, and chiropractic maneuvers: We do not see, for example, a dramatic difference between the non-advocacy related tweets in May versus June except for those related to integrative medicine. While #physicaltherapy, #acupuncutre, #chiropractic did become more popular in terms of frequency and popularity in the month of June, the rest of the top 30 have always been there: #chronicpain and #chronicillness related tweets such as #pain, #painrelief, #spoonie are among the top 30 by frequency (Table 1) and also popularity in both months (Table 2).

Migraine advocacy also appears to have diversified the source of reference material in the dominant subgroup, i.e. the dominant conversation cluster. In the month of June, the number of unique websites that has been referenced by the dominant subgroup significantly increased compared to the month of May. Indeed, nlm.nih.gov was referenced in June and not in May. Advocacy therefore has the potential of improving the source of migraine information on social media. This is further supported by Table 6, where, www.migraineheadacheawarenessmonth.org actually beats out www.instagram.com, another popular social media site, in terms of popular site of information. We should note a major confounder to our interpretation above, however—the AHS virtual meeting also occurred in June, thus likely contributing to diversification of the source of migraine conversation.

Finally, and not surprisingly, nothing does more for migraine advocacy than the popular Chicago Bull’s retrospective The Last Dance (#thelastdance), featured on ESPN, where a key player’s migraine attack during a game drastically influenced sports history. As such, #thelastdance made it as 2nd most frequently occurring tweet in our dataset; it is nevertheless not “popular” according to degree centrality. This is likely that it has never infiltrated the dominant subgroup of conversation as noted by the HCS algorithm.

Marginalized conversations

While migraine advocacy appears to have situated itself as the dominant conversation on Twitter in the month of June, it is also worthwhile to examine the marginalized conversations as interpreted by the HCS algorithm: these are clusters of conversations that seemingly forms its own small insular group. First, as described in the results section, a number of small subgroups appear to arise from conversation around specific websites. The websites choosingwisely.org (an American Board of Internal Medicine initiative), insidehealthpolicy.com (a website focusing on health care policy), and innovationorigins.com (a news site on innovation) contain clusters of hashtags that focus on very narrow topics such as pharmaceutical companies or machine learning techniques. Of specific interest is that the website whatcorinesaid.com, a popular blog, actually was able to produce a subgroups of hashtags (#bloggershutrt, #fblchat, #theclqrt, #teacupclub, #grlpowr) with signals so strong as to make it into the most frequent degree centrality measurement.

In general, we observed that marginalized conversations belong to three distinct groups:

Some of the smaller subgroups appear to be related to migraine. For example: subgroup 6 in May appears to be associated with #acupuncture, #pharmaceutical and takes sources from americanheadachesociety.org. Subgroup 23 in May appears to be related to chocolate, a commonly cited, but questionably contributing, migraine trigger. ³⁶ Subgroup #34 in May is about the Entertainment and Sports Programing Network (ESPN) show The Last Dance. We suspect that the reason these subgroups are marginalized in the migraine conversation is that these conversations are not intended to be about migraine per se. For example, those who talk about The Last Dance on ESPN likely care more about basketball than migraine; therefore, their conversation is unlikely to be about #mham or other hashtags in the dominant subgroup. As noted below in our section on “echo chamber” phenomenon, these subgroups never contain more than 2 websites as sources of information. As such, these subgroups of conversations may represent windows of opportunities for migraine education in an online forum.

“Echo chamber” phenomenon?

It is important to note that marginalized conversations in our study contain no more than 2 websites as sources of reference. Indeed, a number of these marginalized conversations do not contain any references at all and when they do, they simply reference Twitter itself. We will require more data to determine whether this model “the echo chamber” phenomenon that has often been hypothesized on social media as source of misinformation. ⁴⁰

As discussed above, one can categorize marginalized conversations into the following: those that are clearly migraine related, clearly not migraine related, and related to major social movements. For those subgroups that are not migraine related and to which migraines are perhaps simply inadvertently mentioned this “echo chamber” effect is perhaps not particularly detrimental. For example, subgroup 11 in June appears to be a cat-oriented subgroup. To demand rigor scientific references for migraine in these discussions are perhaps excessive. However, for migraine related and/or major social movement related subgroups, the lack of diverse referenced information on migraine may be a missed opportunity for migraine education/advocacy, or worse, a potential source of propagating misinformation.

Sources of information

Our article appears to be the first project to investigate the sources of Internet information on Twitter in relations to migraine. We believe this to be a particularly important topic of investigation in the era of fake news, misinformation, and disinformation. At first glance, Twitter appears to be a self-referencing project: Twitter.com remains the most commonly referenced website. However, this observation maybe biased in that “gifs” and emojis, as well as a various numbers of Twitter objects, are from the Twitter website itself. Yet despite this bias, it is important to note that other social media sites such as youtube.com, instagram.com, and whatcorinedid.com remain the most commonly cited source of information. (As noted above, that whatcorinedid.com belongs not to the dominant subgroup but belong to subgroup #17 and 7 in the month of May/June. Thus, likely it is not a website that contribute to the majority of migraine information in the dominant subgroup of conversation.) Therefore, our project suggests that a significant number of information on Twitter come from other social media. Indeed, only clinicalpainadvisor.com and migraineheadacheawarenessmonth.org rival other social media websites as sources of information (Table 6). We know that online portrayal of objective information on migraine—such as migraineous aura—are often incorrect. ⁴¹ Information on migraine treatments in Google searches are more often than not contributed by non-medical professionals or by those with commercial interests. ²⁶ Given these biases and problematic information, social media self-referencing can be a significant factor in propagating misinformation. This is especially true in the area of alternative medicine and integrative medicine, which appear to have a rising interest in our dataset. This mirrors a similar rise in interest for complementary medicine in YouTube as noted by Saffi et al. ²⁵ Unfortunately those complementary medicine contents on YouTube are often non-evidenced based and thus has the potential of becoming sources for misinformation. ²⁵ In sum, if social media self-referencing is an important phenomenon, then it becomes more vital for practitioners and headache physicians to influence social media conversations from within.