Using Social Media to Track Geographic Variability in Language About Diabetes: Infodemiology Analysis

Introduction

Background

Diabetes affects 30 million people in the United States, and its prevalence varies by geographic region. A better understanding of the regional differences concerning diabetes could allow for better public health messaging. The colloquial person-to-person communication about diabetes might inform that understanding, but word-of-mouth communication has been hard to measure until social media created the possibility of listening in.

Social media platforms such as Twitter, Facebook, and Instagram have emerged as high-volume, real-time data sources to study and observe communications, including health-related communications, from broad population segments [1-5]. Web-based communities are often far reaching, offering various types of communication including person-to-person communication, information seeking and dissemination, social support, and broadcasting of ideas and opinions. In addition, these communities can have similar location-specific characteristics. The content and characteristics of social media posts are associated with the regional epidemiology of disease [6-8]. For example, Instagram users residing in areas with low access to grocery stores (food deserts) posted about and consumed foods higher in fat and cholesterol compared with users residing in areas with greater access to grocery stores [3]. Thus, a better understanding of how people talk about diabetes via social media could provide insights about how to provide better targeted disease management and treatment.

Objective

In this study, we sought to characterize language about diabetes on Twitter and examine the correlation between this language and the prevalence of diabetes.

Methods

Data Source and Sample

This was a retrospective study of data extracted from Twitter about diabetes. Using natural language processing methodology, we found diabetes-specific terms, grouped them into clusters, and then quantified associations with the prevalence of diabetes. This study was approved by the Institutional Review Board of the University of Pennsylvania.

Tweets are brief status updates (no more than 140 characters during the duration of this study) containing information about emotions, thoughts, behaviors, and other personally salient information. Twitter users are broadly represented across age, geography, and social distributions [9-11]. African Americans, Latinos, and those in urban areas are overrepresented on Twitter relative to the general population [12].

For this study, we examined a random 10.00% (3,700,000/37,000,000) sample of all tweets between July 2009 and February 2015 (37 billion total tweets). We then extracted all tweets in English language with the keyword diabetes that originated in the United States, with GPS coordinates or other identifying information sufficient for linking to a US county (such as direct reference to a named county within a state, such as Philadelphia County, Pennsylvania). Approximately 21% of Twitter users provide their location information [5].

Twitter Topic Generation

We first limited our analysis to diabetes-specific language by finding those words and phrases that had a significant association with posts mentioning diabetes. Specifically, we used a random sample of 25,000 tweets including the word diabetes and 25,000 tweets without the word diabetes, and out of the 5000 most frequently used words, we kept those that were used significantly more frequently in the diabetes-related messages according to a logistic regression (Benjamini-Hochberg corrected P<.05 [13]). This removed nondiabetes-related words such as the or like. We then grouped diabetes-specific vocabulary in topics (clusters of semantically related words) using Latent Dirichlet Allocation (LDA). LDA is an automated machine learning process by which frequently co-occurring words are organized into topics [14]. Topic usage is quantified on a scale, referred to as topic probability, from 0 to 1 (from not used at all to exclusively used), which corresponds to the percentage of words from the given topic.

Two research assistants then independently reviewed 100 topics and categorized them into common themes based on the language within the topics. Any deviations between the research assistants were discussed among the research team members to reach consensus.

Relation of Diabetes Topics and Prevalence

To determine how topics on diabetes relate to diabetes prevalence, topic probabilities were individually correlated with age-adjusted county diabetes rates from the Centers for Disease Control and Prevention at the county level for 2012 [15]. In addition, topics were regressed against the 9 US Census Divisions using logistic regression controlling for language of the division.

P values were corrected for multiple testing using the Benjamini-Hochberg procedure. Pearson correlation coefficients are reported for topics, with P<.01 indicating significance.

All statistical analyses were performed with the Differential Language Analysis Toolkit version 1.1 [16] and Python 2.7.10 (Python Software Foundation).

Results

From approximately 37 billion tweets, 1.8 billion included sufficient location information to map to US counties. Of those, 1.6 billion were in English, of which 239,989 tweets (0.15%) included the term diabetes, representing 121,494 unique users.

Topics categorized into themes are displayed in Table 1. Each row of words represents 1 topic within the theme. Examples of topics that correlated with diabetes-related tweets included unhealthy food and drink-themed topics [(cupcakes, whipped, Haribo, and sundae) and (chocolate, Cinnabons, meats, and soda)] as well as a risk factors theme (body mass index, waist, drugs, alcoholic, and obese) and a fundraising theme (walk, charities, supporting, donation, and November).

Twitter users from regions with high prevalence of diabetes were more likely to tweet about unhealthy foods (candy bar, cookies, and Twinkies; r=0.088; P=.002), whereas twitter users from areas with low prevalence of diabetes were more likely to tweet about research (clinical, published, and enrolling; r=0.162; P<.001).

Themes and topics about diabetes differed in relation to overall prevalence of diabetes across US geographic divisions. Areas with high prevalence of diabetes, such as the East South Central and South Atlantic divisions, also had topics referencing unhealthy food (standardized beta range=0.073-0.146). However, research and exercise were most highly correlated with diabetes prevalence in the Northeast (standardized beta for research and exercise was .107 and .142, respectively).

Discussion

Principal Findings

This study reveals that (1) there is variation in what people post on Twitter about diabetes and (2) topics vary by county-level prevalence of diabetes. Unhealthy food–related topics were positively associated with high prevalence of diabetes; conversely, topics about research were negatively correlated with the prevalence of diabetes. The causal directions of these associations, if any, are unclear, but the results suggest opportunities to target online health messages relative to the prevalence of the disease.

This growing body of research utilizing social media platforms to explore public health topics may be helpful for targeting specific patient populations for public health messaging via appropriate language and message content. The ability to relate to different patient populations based on language can better align public health professionals and patients [17,18]. Subpopulations of patients, based on geography, disease severity, or other factors, may use different synonyms or metaphors for symptoms not known to the general public or health professionals. Local health care organizations and professionals could, for example, utilize language common to a particular geographic area with high prevalence of diabetes to target healthy messaging on social media and print media. These organizations may also utilize healthy messaging from other areas with low prevalence of diabetes to influence health behaviors. Large national organizations may also utilize regional differences in content and language to better personalize and position tweets within particular geographic contexts [19].

Content may also be enhanced by tweet modifiers (eg, hashtags and emotion) shown to impact dissemination of cardiovascular health–related Twitter posts [7]. Mining social media to find these nuances within a population posting about diabetes would be useful for outreach and message targeting. Furthermore, learning how different message types (ie, shocking or humorous) are related to gaining knowledge of serious health effects for particular health behaviors is crucial to influence behavior change [2].

Strengths and Limitations

This study has several limitations. Twitter users are not nationally representative, and tweets are not a direct proxy for all person-to-person communication. Tweets are short, and content is presumably what users are eager to share broadly (vs what they may be focused on privately). Nevertheless, tweets offer a window into public discourse about diabetes. This study also has strengths: it starts from an enormous sample of tweets, systematically addresses their content via machine learning techniques, and associates that content with disease prevalence. In doing so, it advances our understanding of public perceptions of diabetes.

Conclusions

This study demonstrates that the language used to discuss diseases is variable and complex. Systematic assessment of social media about posts on diabetes could suggest targets for promoting healthy lifestyles and behaviors.