This systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [26].

Studies were included if (1) the objectives were to address the study of health misinformation on social media, search systematically for health misinformation, and explicitly discuss the impact, consequences, or purposes of misinformation; (2) the results were based on empirical results and the study used quantitative, qualitative, and computational methods; and (3) the research was specifically focused on social media platforms (eg, Twitter, Facebook, Instagram, Flickr, Sina Weibo, VK, YouTube, Reddit, Myspace, Pinterest, and WhatsApp). For comparability, we included studies written in English that were published after 2000 until March 2019.

Articles were excluded if they addressed health information quality in general or if they partially mentioned the existence of health misinformation without providing empirical findings. We did not include studies that dealt with content posted on other social media platforms. During the screening process, papers with a lack of methodological quality were also excluded.

We searched MEDLINE and PREMEDLINE in March 2019 using the PubMed search engine. Based on previous findings [16], the query searched for MeSH terms and keywords (in the entire body of the manuscript) related to the following three basic analytical dimensions that articulated our research objective: (1) social media, (2) health, and (3) misinformation. The MeSH terms were social media AND health (ie, this term included health behaviors) AND (misinformation OR information seeking behavior OR communication OR health knowledge, attitudes, practice). Based on the results obtained through this initial search, we added some keywords that (having been extracted from the articles that met the inclusion criteria) were specifically focused on the issue of health misinformation on social media. The search using MeSH terms was supplemented with the following keywords: social media (eg, “Twitter” OR “Facebook” OR “Instagram” OR “Flickr” OR “Sina Weibo” OR “YouTube” OR “Pinterest”) AND health AND misinformation (eg, “inaccurate information” OR “poor quality information” OR “misleading information” OR “seeking information” OR “rumor” OR “gossip” OR “hoax” OR “urban legend” OR “myth” OR “fallacy” OR “conspiracy theory”). This initial search retrieved 1693 records. Additionally, this search strategy was adapted for its use in Scopus (3969 records) and Web of Science (1541 records). A full description of the search terms can be found in Multimedia Appendix 1.

In total, we collected 5018 research articles. After removing duplicates, we screened 3563 articles and retrieved 226 potentially eligible articles. In the next stage, we independently carried out a full-text selection process for inclusion (k=0.89). Discrepancies were shared and resolved by mutual agreement. Finally, a total of 69 articles were included in this systematic review (Figure 1).

In the first phase, the data were extracted by VSL and then checked by VSL and JAG. In order to evaluate the quality of the selected studies and given the wide variety of methodologies and approaches found in the articles, we composed an extraction form based on previous work [27-29]. Each extraction form contained 62 items, most of which were closed questions that could be answered using predefined forms (yes/good, no/poor, partially/fair, etc). Following this coding scheme, we extracted the following four different fields of information: (1) descriptive information (27 items), (2) search strategy evaluation (eight items), (3) information evaluation (six items), and (4) the quality and rigor of methodology and reporting (15 items) for either quantitative or qualitative studies (Multimedia Appendix 1). Questions in field 2, which have been used in previous studies [27], assessed the quality of information provided to demonstrate how well reported, systematic, and comprehensive the search strategy was (S score). The items in field 3 measured how rigorous the evaluation was (E score) for health-related misinformation [27]. Field 4 contained items designed for the general evaluation of quality in the research process, whether quantitative [28] or qualitative [29]. This Q-score approach takes into account general aspects of the research and reporting, such as the study, methodology, and quality of the discussion. For each of the information fields, we calculated the raw score as the sum of each of the items by equating “yes” or “good” as 1 point, “fair” as 0.5 points, and “no” or “poor” as 0 points (Multimedia Appendix 2). The purpose of these questions is to guarantee the quality of the selected studies.

Furthermore, in order to be able to compare the methods used in the selected studies, the studies were classified into several categories. The studies classified as “content/text analysis” used methods related to textual and content analysis, emphasizing the word/topic frequency, linguistic inquiry word count, n-grams, etc. The second category “evaluating content” grouped together studies whose methods were focused on the evaluation of content and information. In general, these studies analyzed different dimensions of the information published on social media. The third category “evaluating quality” included studies that analyzed the quality of the information offered in a global way. This category considered other dimensions in addition to content, such as readability, accuracy, usefulness, and sources of information. The fourth category “sentiment analysis” included studies whose methods were focused on sentiment analysis techniques (ie, methods measuring the reactions and the general tone of the conversation on social media). Finally, the “social network analysis” category included those studies whose methods were based on social network analysis techniques. These studies focused on measuring how misinformation spreads on social media, the relationship between the quality of information and its popularity on these social platforms, the relationship between users and opinions, echochambers effects, and opinion formation.

Of the 226 studies available for full-text review, 157 were excluded for various reasons, including research topics that were not focused on health misinformation (n=133). We also excluded articles whose research was based on websites rather than social media platforms (n=16), studies that did not assess the quality of health information (n=6) or evaluated institutional communication (n=5), nonempirical studies (n=2), and research protocols (n=1). In addition, two papers were excluded because of a lack of quality requirements (Q score <50%). Finally, the protocol of this review was registered at the International Prospective Register of Systematic Reviews (PROSPERO CRD42019136694).

Ultimately, 69 studies were identified as eligible, and they covered a wide range of health topics and social media platforms, with the most common data source being Twitter (29/69, 43%), followed by YouTube (25/69, 37%) and Facebook (6/69, 9%). The less common sources were Instagram, MySpace, Pinterest, Tumblr, WhatsApp, and VK or a combination of these. Overall, 90% (61/69) of the studies were published in health science journals, and only 7% (5/69) of the studies were published in communication journals. The vast majority of articles analyzed posts written exclusively in one language (63/69, 91%). Only a small percentage assessed posts in more than one language (6/69, 10%).

Table 1 classifies the studies by topic and social media platform [30-97]. It also includes the prevalence of health misinformation posts. The topics were articulated around the following six principal categories: vaccines (22/69, 32%), drugs or smoking issues (16/69, 22%), noncommunicable diseases (13/69, 19%), pandemics (7/69, 10%), eating disorders (6/69, 9%), and medical treatments (5/69, 7%). The quality assessment results for the S score, E score, and Q score are reported in Multimedia Appendix 3.

Figure 2 shows the prevalence of health misinformation grouped by different topics and social media typology. Studies are ordered according to the percentage of health misinformation posts found in the studies selected. These works were also classified according to the type of social media under study. In this way, papers focused on Twitter, Tumblr, or Myspace were categorized as “microblogging.” Additionally, papers focused on YouTube, Pinterest, or Instagram were classified within “media sharing” platforms. Moreover, papers focused on Facebook, VK, or WhatsApp were included within the group of “social network” platforms. While all topics were present on all the different social media platforms, we found some differences in their prevalence. On one hand, vaccines, drugs, and pandemics were more prevalent topics on microblogging platforms (ie, Twitter or MySpace). On the other hand, on media sharing platforms (ie, YouTube, Instagram, or Pinterest) and social network platforms (ie, Facebook, VK, or WhatsApp), noncommunicable diseases and treatments were the most prevalent topics. More specifically, Twitter was the most used source for work on vaccines (10/69), drugs or smoking products (10/69), pandemics (4/69), and eating disorders (3/69). For studies on noncommunicable diseases (9/69) or treatments (5/69), YouTube was the most used social media platform.

Overall, health misinformation was most prevalent in studies related to smoking products, such as hookah and water pipes [33,59,71], e-cigarettes, and drugs, such as opioids and marijuana [45,70,97]. Health misinformation about vaccines was also very common. However, studies reported different levels of health misinformation depending on the type of vaccine studied, with the human papilloma virus (HPV) vaccine being the most affected [67,68]. Health misinformation related to diets or pro–eating disorder arguments were moderate in comparison to the aforementioned topics [35,93]. Studies focused on diseases (ie, noncommunicable diseases and pandemics) also reported moderate misinformation rates [56,85], especially in the case of cancer [76,96]. Finally, the lowest levels of health misinformation were observed in studies evaluating the presence of health misinformation regarding medical treatments. Although first-aid information on burns or information on dental implants was limited in quantity and quality, the prevalence of misinformation for these topics was low. Surgical treatment misinformation was the least prevalent. This was due to the fact that the content related to surgical treatments mainly came from official accounts, which made the online information complete and reliable.

Regarding the methods used in the different studies, there were some differences between the diverse social media platforms. We classified the studies based on the methods applied into the following five categories: social network analysis (19/69), evaluating content (18/69), evaluating quality (16/69), content/text analysis (12/69), and sentiment analysis (4/69). Figure 3 shows the different methods applied in the studies classified by the type of social media platform and ordered by the percentage of misinformation posts. Among platforms, such as YouTube and Instagram, methods focused on the evaluation of health information quality and content were common, representing 22% (15/69) and 12% (8/69), respectively. On microblogging platforms, such as Twitter and Tumblr, social network analysis was the method most used by 19% (13/69) of the studies. Finally, on social media platforms, such as Facebook, VK, and WhatsApp, studies whose methods were related to social network analysis represented 3% (2/69) of the included studies and those focused on the evaluation of content represented 4% (3/69) of the included studies.

This work represents, to our knowledge, the first effort aimed at finding objective and comparable measures to quantify the extent of health misinformation in the social media ecosystem. Our study offers an initial characterization of dominant health misinformation topics and specifically assesses their prevalence on different social platforms. Therefore, our systematic review provides new insights on the following unanswered question that has been recurrently highlighted in studies of health misinformation on social media: How prevalent is health misinformation for different topics on different social platform types (ie, microblogging, media sharing, and social networks)?

We found that health misinformation on social media is generally linked to the following six topical domains: (1) vaccines, (2) diets and eating disorders, (3) drugs and new tobacco products, (4) pandemics and communicable diseases, (5) noncommunicable diseases, and (6) medical treatments and health interventions.

With regard to vaccines, we found some interesting results throughout the different studies. Although antivaccine ideas have been traditionally linked to emotional discourse against the rationality of the scientific and expert community, we curiously observed that in certain online discussions, antivaccine groups tend to incorporate scientific language in their own discourse with logically structured statements and/or with less usage of emotional expressions [53]. Thus, the assimilation of the scientific presentation and its combination with anecdotal evidence can rapidly spread along these platforms through a progressive increment of visits and “likes” that can make antivaccine arguments particularly convincing for health information seekers [53,55]. Furthermore, we found that the complex and heterogeneous community structure of these online groups must be taken into account. For instance, those more exposed to antivaccine information tend to spread more negative concerns about vaccines (ie, misinformation or opinions related to vaccine hesitancy) than users exposed to positive or neutral opinions [49]. Therefore, negative/positive opinions are reinforced through the network structure of particular social media platforms. Moreover, fake profiles tend to amplify the debate and discussion, thereby undermining the possible public consensus on the effectiveness and safety of vaccines, especially in the case of HPV; measles, mumps, and rubella (MMR); and influenza [23].

As observed in our review, health topics were omnipresent over all social media platforms included in our study; however, the health misinformation prevalence for each topic varied depending on platform characteristics. Therefore, the potential effect on population health was ambivalent, that is, we found both positive and negative effects depending on the topic and on the group of health information seekers. For instance, content related to eating disorders was frequently hidden or not so evident to the general public, since pro–eating disorder communities use their own codes to reach specific audiences (eg, younger groups) [98]. To provide a simple example, it is worth mentioning the usage of nicknames, such as proana for proanorexia and promia for probulimia, as a way to reach people with these health conditions and make it easier for people to talk openly about their eating disorders. More positively, these tools have been useful in prevention campaigns during health crises. For example, during the H1N1, Ebola, and Zika pandemics, and, even more recently, with the ongoing COVID-19 pandemic, platforms, such as Twitter, have been valuable instruments for spreading evidence-based health knowledge, expert recommendations, and educative content aimed at avoiding the propagation of rumors, risk behaviors, and diseases [31,89].

Throughout our review, we found different types of misinformation claims depending on the topic. Concerning vaccines, misinformation was often framed with a scientific appearance against scientific evidence [53]. Drug-related misinformation promoted the consumption and abuse of these substances [66]. However, these statements lacked scientific evidence to support them [32]. As with vaccines, false accounts that influenced the online conversation did so with a scientific appearance in favor of e-cigarettes [82]. In this sense, most accounts tended to promote the use and abuse of these items. With beauty as the final goal, misinformation about eating disorders promoted changes in the eating habits of social media users [91]. Furthermore, we found that social media facilitated the development of pro–eating disorder online communities [35]. In general, the results indicated that this type of content promoted unhealthy practices while normalizing eating disorders. In contrast, epidemic/pandemic-related misinformation was not directly malicious. Misinformation on this topic involved rumors, misunderstandings, and doubts arising from a lack of scientific knowledge [31]. The statements were within the framework of the health emergency arising from the pandemic. In line with these findings, we noted findings related to noncommunicable diseases. Messages that focused on this topic promoted cures for chronic diseases or for conditions with no cure through fallacies or urban legends [85].

In this study, we focused on analysis of the results obtained and the conclusions of the authors. Some of our findings are in line with those obtained in recent works [16]. The reviewed studies indicate, on one hand, the difficulty in characterizing and evaluating the quality of health information on social media [1] and, on the other, the conceptual fuzziness that can result from the convergence of multiple disciplines trying to apprehend the multidisciplinary and complex phenomenon of health misinformation on social media. This research field is being studied by health and social scientists [70,73], as well as by researchers from the fields of computer science, mathematics, sociophysics, etc [99,100]. Therefore, we must understand that the inherent multidisciplinary and methodological diversity of studies and the highly dynamic world of social media are a perfect match for making it more difficult to identify comprehensive and transversal solutions to the problem of health misinformation. In fact, as we have found, misinformation on vaccines, drugs, and new smoking products is more prevalent on media-sharing platforms (eg, YouTube) and microblogging applications (eg, Twitter), while misinformation on noncommunicable diseases is particularly prevalent on media sharing platforms where users can widely describe disease symptoms, medical treatments, and therapies [76,85]. Platforms, such as YouTube, owing to their characteristics, allow more space for users to share this type of information, while the natural dynamism of Twitter makes it an ideal medium for discussion among online communities with different political or ideological orientations (eg, pro/antivaccination communities).

Finally, we should mention that the current results are limited to the availability and quality of social media data. Although the digitalization of social life offers researchers an unprecedented amount of health and social information that can be used to understand human behaviors and health outcomes, accessing this online data is becoming increasingly difficult, and some measures have to be taken to mitigate bias [40,43,67,79]. Over the last few years, new concerns around privacy have emerged and led governments to tighten regulations around data access and storage [101,102]. Consequently, in response to these new directives, as well as scandals involving data sharing and data breaches such as the Cambridge Analytica case, social media companies are developing new controls and barriers to data in their platforms. This is why free access to application programming interfaces (APIs) is becoming increasingly difficult and the range of social data accessible via APIs is gradually decreasing. These difficulties in accessing data are also determining which platforms are most frequently used by researchers, which are not used, and which will be used in the near future.

The present study has some limitations. First, the conceptual definition of health misinformation is one limitation. In any case, taking into account that we were facing a new field of study, we considered a broad definition in order to be more inclusive and operative in the selection of studies. Therefore, we included as many papers as possible for the review in order to perform an analysis of the largest number of possible topics. Second, from a methodological perspective, our findings are limited to research published in English language journals and do not cover all the social media platforms that exist. Besides, we discovered some technical limitations when conducting this systematic review. Owing to the newness of this research topic, our study revealed difficulties in comparing different research studies characterized by specific theoretical approaches, working definitions, methodologies, data collection processes, and analytical techniques. Some studies selected involved observational designs (using survey methods and textual analysis), whereas others were based on the application of automatic or semiautomatic computational procedures with the aim of classifying and analyzing health misinformation on social media. Finally, taking into account the particular features of each type of social media (ie, microblogging service, video sharing service, or social network) and the progressive barriers in accessing social media data, we need to consider the information and selection bias when studying health misinformation on these platforms. According to these biases, we should ponder which users are behind these tools and how we can extrapolate specific findings (ie, applied to certain groups and social media platforms) to a broader social context.

Despite the limitations described above, it is necessary to mention the strengths of our work. First, we believe that this study represents one of the first steps in advancing research involving health misinformation on social media. Unlike previous work, we offer some measures that can serve as guidance and a comparative baseline for subsequent studies. In addition, our study highlights the need to redirect future research toward social media platforms, which, perhaps due to the difficulties of automatic data collection, are currently being neglected by researchers. Our study also highlights the need for both researchers and health professionals to explore the possibility of using these digital tools for health promotion and the need for them to progressively colonize the social media ecosystem with the ultimate goal of combating the waves of health misinformation that recurrently flood our societies.

Health misinformation was most common on Twitter and on issues related to smoking products and drugs. Although we should be aware of the difficulties inherent in the dynamic magnitude of online opinion flows, our systematic review offers a comprehensive comparative framework that identifies subsequent action areas in the study of health misinformation on social media. Despite the abovementioned limitations, our research presents some advances when compared with previous studies. Our study provides (1) an overview of the prevalence of health misinformation identified on different social media platforms; (2) a methodological characterization of studies focused on health misinformation; and (3) a comprehensive description of the current research lines and knowledge gaps in this research field.

According to the studies reviewed, the greatest challenge lies in the difficulty of characterizing and evaluating the quality of the information on social media. Knowing the prevalence of health misinformation and the methods used for its study, as well as the present knowledge gaps in this field will help us to guide future studies and, specifically, to develop evidence-based digital policy action plans aimed at combating this public health problem through different social media platforms.