Data collection occurred in three basic steps. First, health users of Twitter were identified. Lists of Twitter users created by Organized Wisdom [19], a company in the health information technology sector, were used to gather data. Organized Wisdom collates health-related Twitter users into lists that enable others to more efficiently find information online; for example, the “doctors verified” list consists of physicians who openly disclose their credentials online. Of these lists, 37 are Twitter accounts for doctors, nurses, hospitals, and other health organizations that cover seven user groups of interest: health advocates, businesses, health care organizations, government organizations, students, professionals, and other (eg, health care-related publications).

Second, from these health users, an equal number (n=200) of the most recent publicly available tweets for each user in each group, including retweets, was downloaded during the period July 31-August 3, 2012. In cases where users had fewer than 200 tweets, all available data were obtained.

Third, from this large collection of tweets from the seven user groups, a smaller collection was sampled. For each group, a user from the group was randomly selected, and a random tweet (without replacement) was selected for that user. This strategy was used to create a preliminary sample of tweets (15 random tweets from each user group, or 105 total) to develop a coding scheme and a final sample for in-depth qualitative analysis (100 random tweets from each user group, or 700 total).

The research team first hypothesized content areas of interest from their own experience and the published literature to generate a draft coding scheme. Then, using the preliminary sample of 105 tweets, 4 coauthors (JL, ZB, MD, MC) conducted content analyses of each tweet using this draft scheme. The coding scheme was then revised for the final analysis. The categories of interest are presented in Table 1. Subsequent analysis was then conducted on an additional 700 tweets as described above. Areas of disagreement between reviewer pairs were resolved by consensus in a case-by-case fashion. Reviewers were blind to the identity of the Twitter users during content analysis.

After unmasking the data, users and user groups were reviewed to verify the accuracy of categories provided by Organized Wisdom. None of the tweets reviewed were found to be authored by the research team. In several instances, the provided categories were inaccurate; for example, a self-identified hotel was miscategorized as a government agency. Therefore, two members of the research team (JL and ZB) re-categorized all users into seven new user groups based on consensus review of publicly available profiles (Table 2): users without available profiles, non-health users, health advocates, health businesses, non-provider health professionals, health providers, and health care organizations. This resulted in a different number of tweets for each of the seven groups.

Descriptive statistics were used to tabulate types of tweets for each group of users. Statistically significant differences between the proportion of tweets among user groups were assessed using the chi-square test. Stata 13 was used for the analysis.

This study was reviewed and declared exempt from further review by the Johns Hopkins Medicine Institutional Review Board.

Of all tweets, 5.0% (35/700) non-English tweets were excluded. The remaining 95.0% (665/700) English-language tweets were analyzed in depth (Figure 2). A total of 255 unique users contributed to the 665 tweets. Of the tweets in the analytic sample, 52.0% (346/665) were categorized as health-related. For example, a tweet such as “RT @Username: Tai chi improves mood, quality of life for patients with systolic heart failure” was categorized as health-related. Each health-related tweet was thereafter sub-coded into at least one of six non-mutually exclusive categories, as defined by Table 1.

Nearly one-third (31.1%, 207/665) of all tweets in the sample were personal in nature. Among personal tweets, 70.5% (146/207) were non-health related, whereas 29.5% (61/207) were health-related. This finding was statistically significant (P<.001).

The research team was most interested in health-related tweets. Just over half (53.2%, 184/346) of these tweets contained “testable” claims, defined as claims for which someone viewing the tweet might expect to be supported by medical evidence. An example of a testable tweet is shown in Table 1 (first row).

Within health-related tweets, news was the second most frequent content category; 41.0% (142/346) of these tweets contained news. Over a quarter (26.9%, 93/346) contained reference to a commercial product or service, 17.6% (61/346) relayed a personal viewpoint or experience, and 17.1% (59/346) discussed wellness.

Table 3 summarizes the distribution of tweet content areas among user groups, where content by health providers serves as the baseline for comparison. Overall, there were no significant differences in the types of content between health providers and non-health users and those without available profiles.

Health advocates tweeted about personal health matters significantly more than providers. Provider tweets were most different from those of health care businesses. Businesses were statistically more likely to tweet about news and less likely to tweet about wellness, and non-provider health professionals were statistically more likely to tweet testable claims and news stories and less about personal content. The tweets of health care organizations and providers differed as well. Organizations’ tweets were statistically more likely to include commercial content and wellness but were less likely to include personal content.

This analysis of self-identified health users and their health-related messages on Twitter identified several important findings. First, over half of the tweets in the sample included tweets with testable claims, defined for this study as content that a Twitter user might expect to be supported by medical evidence. The majority of health tweets were testable and comprised various topics. In a preliminary analysis, we found that about 40% of such testable tweets shared medical advice while 23% disseminated scientific news or research books and articles. Although the research team did not analyze this in detail, the sources behind these medical claims were more varied than scientific research (ie, “evidence-based” medicine claims), ranging from common sense (eg, advice to drink more water on a hot day) to WebMD to Dr. Oz to personal blogs. Whether the claims were actually scientifically valid was not examined and requires further study.

Second, in addition to traditional providers, a wide range of non-health providers are sending health-related tweets. Moreover, the content of tweets by health-related users on Twitter varied with user type. Health care organizations, for example, shared more health news and testable claims than individuals but also endorsed commercial services or products more than individuals. In the future, it could be important to examine whether organizations’ tweets are based more often in rigorous evidence as compared to those of individual health professionals. These findings confirm and expand upon the work of Sugawara et al [20] and Chretien et al [21], who both focused on a subset of health users. In examining the tweets of cancer patients in Japan, Sugawara found that patients tweeted about both personal (eg, greetings, chats) and medical matters (eg, information about treatments). Similarly, Chretien found that health providers (specifically, physicians) shared personal, commercial, as well as medical information over Twitter.

We found that 17% of users in this sample were non-provider individuals in the health field, together contributing 20% of health-related tweets. Their participation illustrates the wide range of non-providers engaging in information sharing and health messaging on Twitter. While this community of users represents a significant resource of health conversation and information on Twitter, these non-providers are not bound by professional guidelines on Twitter use. It is unclear how Twitter users might use or understand the tweets of this group compared, for example, to tweets by physicians and nurses. Given that trust is a significant feature of social networks (with young users potentially trusting their social network as much as physicians), this requires further examination [3,22].

Third, these findings demonstrate potential gaps between professional guidelines and online behavior [23,24]. Many medical societies, institutions, and medical schools disseminate such guidelines for clinicians and students. Often, these guidelines encourage the use of social media as a tool for endorsing medical information, discourage using the medium for medical consults, and counsel against posting personal content. Despite this, 17% of health tweets from providers and 74% from medical students included personal content. This lack of adherence to professional guidelines identifies an area where additional education or potentially consensus building within the profession is needed. Similarly, the sizable minority (14%) of providers’ tweets that included commercial products or services confirms prior findings and concerns about potential conflicts of interest on Twitter (a medium where disclosure is difficult, if not impossible) [21,25].

As the use of Twitter has increased, so has the scholarship on the use of Twitter for health-related purposes. Some researchers have used Twitter as a surveillance [7,13,26] and prediction tool for disease outbreak [27] while others have characterized the relationship between segments of health-related users on Twitter, such as medical societies [28], internal medicine physicians [29], and cancer patients [20]. Though few papers examined health-related users as a whole, or the content of their tweets in detail, ours is not the first content analysis of Twitter. For example, Chew and Eysenbach analyzed individual tweets relating to the 2009 H1N1 pandemic, finding that information from a variety of credible sources, in addition to opinions and experiences, were disseminated on Twitter [26]. Our study, a qualitative analysis of tweets from health-related users, addressed different questions and thus made use of a different approach.

This study adds to the growing literature on using Twitter for health purposes by examining the content of tweets by health-related users. In doing so, this study had several important strengths. It represents an in-depth exploration of the types of health-related users and content on Twitter. The analysis is the first to examine in detail both health users and their health-related content on Twitter, orienting researchers and clinicians to the Twitter health landscape. While previous studies focused on Twitter use among certain populations or users [8,20,21,30,31] such as cancer patients in Japan, or certain topics such as childhood obesity [9], this study explored several types of users, achieving breadth and qualitative depth simultaneously.

The study was not without limitations. Data collection occurred in the summer of 2012; this sample thus reflects a Twitter landscape that may have shifted since then. For example, adherence to professional guidelines could be expected to increase over time. This study’s findings set a baseline for future analyses. The selection of the analytic sample was limited by the absence of a simple mechanism to sample health-related users and tweets on Twitter. For this reason, it was not feasible to select a random sample of all health users and their tweets, reducing the generalizability of the study findings. Although Organized Wisdom, by organizing content into lists, seemed to offer a reasonable way to address this problem, many user category designations were inaccurate. Favoring accuracy of user type over preservation of the original categories, the research team undertook its own characterizations. Because this reduced the number of tweets in certain user groups, this reduced the ability to determine statistical significance between groups. If so, the results likely underestimate potential differences between groups. Importantly, this only reinforces the need for in-depth analysis while revealing a concern that collated “health-related user” content might not be from health-related users.

The analysis did not purposely evaluate “high impact” users (eg, Dr. Oz, Atul Gawande, Sanjay Gupta). We also did not independently verify user profiles, and users representing themselves as providers were not checked against other databases. Furthermore, while the analysis was double coded and disagreements were resolved by consensus, the subjective nature of coding limits the validity of these findings.

This study presents several opportunities for further research. While a preliminary assessment evaluated the evidence base of each testable tweet, future analysis could examine this subset of tweets (health-related tweets that contain testable claims), for example, by comparing them to accepted strength of evidence standards [32]. Cross-referencing the identities of the users would be another next step in refining the characterization of users, while including Twitter users who are verifiable, self-identified patients or providers would further enrich the analysis of health messages.

In conclusion, health-related users and the content they share on Twitter are diverse. While providers’ tweets often include testable claims, they also make use of personal statements. Non-providers also contribute to ongoing health-related messaging on Twitter. Users in health-related fields who are actively engaging in health-related conversations on Twitter, as well as those who are merely reading these users’ tweets, should be oriented to the diversity of health-related Twitter content, and, if possible, to the validity of associated tweets.