This is a manual mixed methods descriptive study to characterize public tweets mentioning “monkeypox.” This study was deemed exempt from detailed review by the Emory University institutional review board (category 4; publicly available data). Informed consent was not obtained for our study because our data were publicly available at the time of the study. Despite the data being public, we took an additional step to protect the identity of the subscribers whose posts were included in the analysis by removing their subscriber handles before passing the data on to the annotators.

We collected data via the Twitter academic application programming interface (API) using monkeypox-related keywords and hashtags (eg, “monkeypox,” “#monkeypox,” “#mpoxvirus,” and “monkey pox”). We used the relevant API query filter to only collect posts in the English language. For multiword expressions, individual words were first used for streaming data collection and then regular expression-based matching was performed to filter out posts that did not contain the exact multiword expression. We collected data from August to October 2022—a 3-month period. All tweets were preprocessed via natural language processing methods, and duplicates were removed. We randomly sampled the remaining unique posts for manual analysis. Multiple posts from the same Twitter subscriber were not included. For anonymity purposes, while manually analyzing the data, the researchers used the integer IDs for the tweet extracts to hide the Twitter handles of the original posters from the annotators.

The 3 authors of this manuscript first manually reviewed a sample of posts to prepare a guideline for further annotation or coding. Based on the guidelines, the authors or annotators coded the random sample of posts into topics. Following an initial round of coding of a subsample, the authors resolved disagreements via discussion and identified overarching themes. The more fine-grained coded topics were mapped onto the overarching themes. The rest of the sampled tweets were then manually categorized into the chosen topics and then into themes. In total, we identified 8 themes. Each tweet was assigned only 1 theme based on the most appropriate fit, as determined by the annotators. The categorizations were finalized via discussion after the entire sample was coded.

In total, we collected 128,615 posts within the specified time period. After preprocessing and removing duplicates, we were left with 101,433 posts. A total of 200 posts were sampled and chosen for manual categorization. The initial discussion included 50 posts. In the end, all 200 posts were revisited to ensure accurate categorization.

Table 1 presents the relative distribution of the 8 themes. The description of each theme along with examples is provided in Table 2. Table 1 shows that the most common themes were monkeypox doubts and media (44/200, 22%); the former included posts that questioned the validity and seriousness of the disease, and the latter included news articles shared by Twitter subscribers; 19% (38/200) of the posts discussed how the virus was transmitted, including speculations that may or may not have been accurate, and 13% (25/200) of the posts represented knowledge shared about monkeypox, such as symptoms. Politics (18/200, 9%), monkeypox vaccine (16/200, 8%), effect of monkeypox (12/200, 6%), and general comment (3/200, 1%) were the rest of the identified themes, occurring less than 10% in the sample. Note that some posts included information that could be put into multiple themes; in our final discussion, we adjudicated the most appropriate theme label for each post in such cases.

Our study demonstrated that there was a substantial amount of monkeypox-related information on Twitter during the outbreak. This finding is unsurprising since much of the infodemiology research during and prior to COVID-19 presented similar findings. Our work adds further support to a growing body of research focusing on the intersection of social media and infectious diseases; it also supports the possibility of leveraging knowledge from social media to obtain early insights into infectious diseases and outbreaks.

In terms of the contents of the posts and their themes, we found a mixture of relevant and irrelevant or inaccurate information. Many posts expressed doubts or promoted conspiracy theories about the existence and origin of monkeypox. Posts also expressed political alignment, demonstrating how regional politics can shape people’s perceptions of the disease. This is perhaps the new normal in a post–COVID-19 world, where the spread of misinformation is inevitable. The presence of misinformation makes it even more important to conduct social media–based surveillance so that it can be detected early and public health agencies across the world can take the necessary steps to combat the issue.

Our review also found important information, such as symptoms, knowledge, and emerging news on monkeypox. Such information can be of particularly high utility at the early stages of infectious disease outbreaks when little is known about the topic. This utility was observed during the early stages of COVID-19 when social media chatter revealed symptoms associated with infection [9]. Studies, including ours, on COVID-19 and monkeypox have consistently found that social media subscribers share their personal experiences once infected, and there is the potential to conduct syndromic surveillance using such data. Many posts also quelled misinformation (eg, about how monkeypox spreads) posted by others.

About 40,000 monkeypox-related tweets were posted on the day monkeypox was declared a global public health emergency by the World Health Organization [8]. An early sentiment analysis of the monkeypox tweets revealed a larger number of neutral tweets, followed by negative and then positive tweets [8]. Most of the data revolved around Twitter subscribers’ discussions related to monkeypox. In another study, “expressed fear” or negative sentiments were the most frequent observations, followed by neutral or positive sentiments [10]. Fear and sadness were the most common emotions in negative sentiment Tweets. In addition to sharing their concerns and opinions, Twitter subscribers reportedly spread a lot of mis- and disinformation about monkeypox [11]. In a separate study also involving Twitter, the authors applied unsupervised learning, namely topic modeling, to identify topics related to monkeypox from a large number of posts [12]. Our work, in contrast, focused on a thorough manual categorization, so that topics and themes that are most relevant for subject matter experts were identified and coded. Besides Twitter, other social networks have also been leveraged to study monkeypox. In a study [13], 100 videos were analyzed; 55% of them were from health care professionals, 38% were from non–health care professionals, and 7% were from news organizations. The study found that while health care–related content was more accurate and of higher quality, viewers engaged more with the non–physician-created content [13].

Our study has several limitations. First, only a small sample of data was manually analyzed. Although we could have applied automated strategies, such as topic modeling, for analyzing the data, we intentionally chose to conduct a manual thematic analysis since our past research [14-16] has shown that such analyses provide more fine-grained and detailed insights. Second, we used a limited set of keywords for data collection and did not include lexical variants, potentially reducing the total number of posts. The general limitations of social media also apply to our work—social media subscribers are younger, on average, compared to the overall population, and they also tend to be more tech-savvy. People without access to the internet are automatically excluded from social media studies. Finally, our data collection period was short, due to the formative nature of this study.

Our study revealed the major categories of discussion regarding monkeypox on Twitter. The findings shed light on the possible distribution of social media chatter associated with the most important manually identified themes. Unlike prior work that leveraged automated analyses, our work focused on identifying pertinent themes manually. This is a particular strength of this study since automated methods may not incorporate prior knowledge about the subject matter when identifying topics or themes that are the most relevant. Future work may use manual characterization to train supervised learning models for large-scale, automated analyses. Although our use case was monkeypox, our study suggests that social media data analysis can reveal important information about public perceptions in the early stages of infectious disease outbreaks. Social media data may also enable the early detection of misinformation regarding disease outbreaks, allowing public health agencies and policy makers to act accordingly.