We included English-language social networks and forums worldwide, such as Facebook public pages, Reddit, 4Chan, and the comments sections of news sites such as ABC News [23]. We defined forums as thread-based message boards and topic-specific pages [24]. We chose these sources to provide diversity because they have been studied less than Twitter in this area [25,26]. Additionally, the user base profile of our sources appeared to be more representative of the demographic profile of the broader US population than Twitter. While both Twitter and Reddit were popular among US adults aged ≤30 years, those who lived in urban areas, and were male [27], Facebook appeared to be more popular among female users and US adults older than 30 years [27].

Furthermore, even though there is an overlap between the affordance among our sources and Twitter, Reddit users have more anonymity as they do not need to register an account to access the majority of the content, thus allowing for greater participation [25]. Lastly, forums such as Reddit allow lengthy submissions and are usually topic-specific, which grant opportunities to cover the sensitive topics of our study (eg, mental health disorders and symptoms), which may not typically be discussed on social media [26,28]. The greater the length of the comments (eg, 40 words per comment for Reddit vs <15 words in tweets, on average) and less frequent use of hashtags associated with forums, which also makes it possible to apply more complex natural language processing (NLP) algorithms more accurately to classify sample posts [26].

We partnered with Signals Analytics, an advanced analytics company, to obtain access to target data sources from a third-party data vendor (NetBase) and to conduct the analysis [29,30]. In order to geotag posts, NetBase used a combination of geotagged social media messages, author profiles, and each country’s unique website domain suffix (eg, “.ca” for Canada). All the acquired data were then deidentified by NetBase and transferred to Signals Analytics for analysis.

We also gathered data on COVID-19 cases from the COVID-19 dashboard developed by the Center for System Science and Engineering at Johns Hopkins University, which provides the most comprehensive and up-to-date information on COVID-19 trends [1]. Using the RapidAPI application programming interface (API) [31], we updated the COVID-19 statistics (daily new cases or incidence) on a daily basis.

In this study, all personal identifying information such as usernames, emails, and IP addresses were removed before analysis. The study was exempt from institutional review board review at Yale University as it used publicly available, anonymized data.

For the analysis of data on symptoms and medical conditions being discussed on social media platforms between June 14 (when many countries began to lift major COVID-19 restrictions) and December 13, 2020 (when the first shipment of the COVID-19 vaccine arrived in the United States), we began by applying NLP algorithms to process social media posts collected from data sources during the study period, and then classified these posts in accordance with symptoms and medical conditions being mentioned.

To accomplish this, NetBase ran a daily scheduled data extraction query that we designed for the study on over 300 million web-based data sources (Multimedia Appendix 1). Additionally, we performed the following filtering steps to include posts relevant to our research questions. First, NLP algorithms were run, and advertisements and posts on sites for pornography were removed (Multimedia Appendix 1). Next, we applied a taxonomy of over 3000 health-related topics to identify key words, phrases, and statements mentioning symptoms and medical conditions (Multimedia Appendix 1). Social media posts that did not contain any of the taxonomy terms or symptoms and medical conditions as keywords were then deleted. Lastly, we removed redundant posts, blog posts, and news articles to ensure that the analysis was based on unique posts from social networks, forums, and comments only.

To evaluate the performance of the NLP algorithms and taxonomy classifications of symptoms and medical conditions, we applied the taxonomy to 4 sets of independent 100-post samples and calculated the positive predictive value and sensitivity of the classification (Multimedia Appendix 1). The algorithms used to identify symptoms and medical conditions topics in our study have been previously validated using real-world data to assess the public’s behaviors and perceptions toward COVID-19 [32]. Our study methodology has also been used to provide insights into the characterization and prediction of e-cigarette or vaping product use–associated lung injury outbreaks, known as the EVALI study [33].

Our taxonomy was organized into three levels: categories, subcategories, and topics. Symptoms and medical conditions were the 2 main categories in the taxonomy (Table 1 and Multimedia Appendix 1). The symptoms category included 98 non–COVID-19 topics (symptoms), which were grouped into 7 subcategories based on the affected organ or systems (eg, cardiovascular or respiratory systems). A list of 22 COVID-19–related topics (symptoms) was included as a separate symptom subcategory. The list of COVID-19–related symptoms was defined as outlined by the Centers for Disease Control and Prevention (CDC) on December 22, 2020 [34]. Because our algorithms captured all posts that mentioned any of the listed COVID-19 symptoms in the COVID-19–related symptom subcategory, the included posts may not necessarily represent discussions of symptoms experienced by patients with COVID-19. The medical condition category included 2200 topics (medical diagnoses), which were grouped into 10 subcategories. Categories, subcategories, and topics in the taxonomy were not mutually exclusive; each post could be assigned to multiple categories, subcategories, or topics.

We also created content filters to retain posts mentioning COVID-19 for further analysis. We applied 2 filters, COVID-19 disease status and COVID-19 diagnostic methods, to identify discussions on COVID-19 disease status (tested positive or negative, symptomatic or asymptomatic, recovered, and exposed to a confirmed patient) and diagnostic methods (COVID-19 testing, self-diagnosis, and remote diagnosis). These more restrictive searches were conducted by activating the 2 additional filters using the NLP algorithm, and the resulting posts from that search may not indicate the author’s COVID-19 status.

To explore how the discussion of symptoms and medical conditions on social media changed from June 14 to December 13, 2020, we determined the number of posts that included a discussion of each symptom and medical condition over time using NLP classification (Multimedia Appendix 1). To assess whether the frequency of symptom posts was associated with daily COVID-19 statistics, we performed Pearson correlation analysis to determine correlations among the top 5 most discussed symptoms and daily COVID-19 statistics (new cases, new deaths, new active cases [total cases minus recovered and those who have died], and new recovered cases). Additionally, we calculated Pearson correlation coefficients between frequency changes in each of the 5 symptoms and daily fluctuation in any COVID-19 statistic. A 2-tailed P value of <.05 was used to indicate statistical significance. Both posts and COVID-19 statistics used in these analyses were restricted to the United States.

Additionally, we compared the trends of the 5 most frequently mentioned symptoms and medical conditions from June 14 to August 31, 2021 (when the United States crossed the 6 million COVID-19 cases mark), to the trends observed from September 1 to December 13, 2020, by measuring the percent change between the 2 time periods in the number of posts including a discussion of each topic. We compared the 2 time periods to reveal changes in health-related conversations on social media at different stages of the pandemic, as prior literature focused primarily on the early stage of the pandemic (before June 2020). Our approach was also designed to contribute to a better understanding of the impact of COVID-19 on the public’s perceptions and attitudes toward different symptoms, medical conditions, and health care–seeking behaviors.

In this study, we collected and analyzed web-based posts from forums and comments on news sites between June 14 and December 13, 2020. We found that a wide variety of symptoms and medical conditions topics were discussed on non-Twitter social media. While the vast majority of discussions were about COVID-19 infection and COVID-19–related symptoms (as defined by the CDC), neuropsychological symptoms (eg, anxiety) and other medical conditions (eg, infectious diseases and psychiatric disorders) were also frequently mentioned. Additionally, we noticed that changes in posts frequency of anxiety, generalized pain, and weight loss were significant but negatively correlated with daily new COVID-19 cases in the United States, and that the frequency of posts on anxiety, generalized pain, weight loss, fatigue, and the changes in fatigue positively and significantly correlated with daily changes in both new deaths and new active cases in the United States. As COVID-19 cases continued to rise globally, the cumulative volume of posts mentioning anxiety, generalized pain, fatigue, influenza, unspecified CNS disorders, and depression increased from September 1 to December 13, 2020 (compared to June 13 to August 31, 2020).

Our findings expand on previous observations regarding the mental health effects of the COVID-19 pandemic among social media users by presenting a more complete picture of health-related topics discussed on social media [18]. Our results not only confirm the findings from previous studies that showed high levels of anxiety and depression mentioned by social media users during the pandemic [35,36] but also revealed that the frequency of anxiety and other general health symptom–related posts, including generalized pain, weight loss, and fatigue, was significantly correlated with daily COVID-19 statistics. These data support the idea that social media represents a potential powerful source of information for health care professionals to draw real-time estimations about population health status [18,21]. Understanding health symptom posts commonly associated with COVID-19 statistics may inform public health researchers, clinicians, and policymakers to take timely and appropriate public health and clinical measures accordingly.

Further, as access to the internet becomes more widely available and with the anonymity of social media, people who face barriers to accessing health care and those who have mental health symptoms may use social media to speak openly about their health experiences and seek help [21,37]. Collectively, these results further justify our approach to monitoring symptoms and medical condition posts on social media during the pandemic, and call for further investigation of the possibility of using social media analytics to gain insights into the population’s symptoms, including mental health symptoms, which are difficult to monitor outside of the health system, health threats, and to enhance public health preparedness.

As the pandemic progresses, obtaining information on the symptom profile of COVID-19 could help to better diagnose and treat the disease. There has been increasing recognition of the importance of extracting social media information to explore symptom experience and disease progression among patients with COVID-19 [38]. Although we did not restrict our analysis to only social media posts mentioning COVID-19 and could not verify the authors’ disease status, the most discussed COVID-19–related symptoms we found (eg, fatigue, cough, fever, headache, and difficulty breathing) were among the most common symptoms reported by patients with COVID-19 in other studies [39-41]. Based on information extracted by applying COVID-19 disease status and diagnostic methods filters, we found that nearly 40% of non-Twitter social media users who discussed the top 5 most commonly mentioned symptom topics, such as fatigue and cough, also talked about the topic of having tested positive for COVID-19.

We also noticed that approximately 15% of these discussions were related to asymptomatic COVID-19. While an in-depth exploration of these posts using qualitative analysis or sentiment analysis is necessary to help verify the users’ COVID-19 disease status, our preliminary data indicate the potential for extracting information from social media to understand the full spectrum of symptoms experienced by patients with COVID-19. Interestingly, we noticed an increase of over 60% in the volume of posts mentioning less common COVID-19 symptoms such as changes in the senses of taste and smell during the second stage of our study period (September 1 to December 13, 2020). This surge may be partly due to improvements in knowledge and awareness of COVID-19 symptoms in the general population as the 2 symptoms were recently added to the COVID-19 symptom lists of the CDC and the World Health Organization (late April 2020 and early May 2020, respectively).

While there have been fluctuations in the volume of social media posts on a day-to-day basis, there appeared to be seasonal variation in the volume of discussion of symptoms and medical conditions. We noticed that the volume of most health-related discussions increased more from September 1 to December 13, 2020, than from June 14 to August 31, 2020. These changes may have been due to a combination of colder weather in the northern hemisphere and social distancing and limitations on daily life during the pandemic as well as the second wave of COVID-19, resulting in more social media users and more people being restricted indoors [42]. Additionally, there were several inflection points in the volume of discussion of symptoms and medical conditions in the last 6 months. These changes appeared to have coincided with major news stories and national events, echoing findings from other studies that showed the potential impact of media coverage on web-based discussions [6,18]. For example, the volume of all 5 commonly mentioned symptoms (anxiety, generalized pain, weight loss, fatigue, and cough) and 2 medical conditions (unspecified CNS and depression) peaked on October 10, 2020, the day on which hurricane Delta struck Louisiana and nearby states and left 730,000 homes and businesses without power [43]. However, our study did not find evidence of an association between changes in the volume of symptom discussion over time and the trend of daily new confirmed cases of COVID-19 in the United States.

Our study has several limitations. First, information on geolocation, demographics, and COVID-19 disease status was not available for all social media users in the study, owing to various legal limitations (such as General Data Protection Regulation of the European Union). This might have introduced a sampling bias if there were significant differences between social media users’ characteristics in our project and the real world. However, by collaborating with social media analytics companies, we have maximized our ability to access thousands of social media data sources worldwide, thus minimizing the possibility of sampling bias. Additionally, the majority of social media users in our study were from the United States. The findings, therefore, may not be generalizable in their application to users located in other countries. Further, we did not conduct formal statistical analyses beyond comparing the trends differences in frequency of health-related posts and new COVID-19 cases; hence, further testing is needed to confirm the associations between patterns of changes in symptom/medical condition posts and the fluctuations of COVID-19 statistics over time. Finally, we did not perform sentiment analysis or qualitative analysis in the study and did not verify whether authors who discussed COVID-19–related topics had COVID-19 themselves. We hope to accomplish and report this analysis in a future study. We also hope that other studies on social media’s role in public health will replicate and validate our exploratory findings in non-Twitter social media platforms.

In this study, we classified web-based posts collected from June 14 to December 13, 2020, in accordance with discussions of symptoms and medical conditions. Neuropsychological symptoms such as anxiety were the most frequently mentioned symptom subcategory. Furthermore, COVID-19 infection was the most commonly mentioned medical condition. Our analysis also showed that frequency of anxiety and other general health symptoms posts, including generalized pain, weight loss, and fatigue, was significantly correlated with daily COVID-19 statistics in the United States. Additionally, health-related discussions were greater from September 1 to December 13, 2020, than from June 14 to August 31, 2020, aligning with the increase in COVID-19 cases in the United States during the winter months. These preliminary findings show promise for real-time monitoring of social media posts to measure the mental health status of a population during a global public health crisis and to assess the public’s main health needs that have not been captured or met by the existing health system. Future research may incorporate information from social media into predictive models for the detection of emerging infectious diseases.