Tracking Self-reported Symptoms and Medical Conditions on Social Media During the COVID-19 Pandemic: Infodemiological Study

Abstract

Background Harnessing health-related data posted on social media in real time can offer insights into how the pandemic impacts the mental health and general well-being of individuals and populations over time. Objective This study aimed to obtain information on symptoms and medical conditions self-reported by non-Twitter social media users during the COVID-19 pandemic, to determine how discussion of these symptoms and medical conditions changed over time, and to identify correlations between frequency of the top 5 commonly mentioned symptoms post and daily COVID-19 statistics (new cases, new deaths, new active cases, and new recovered cases) in the United States. Methods We used natural language processing (NLP) algorithms to identify symptom- and medical condition–related topics being discussed on social media between June 14 and December 13, 2020. The sample posts were geotagged by NetBase, a third-party data provider. We calculated the positive predictive value and sensitivity to validate the classification of posts. We also assessed the frequency of health-related discussions on social media over time during the study period, and used Pearson correlation coefficients to identify statistically significant correlations between the frequency of the 5 most commonly mentioned symptoms and fluctuation of daily US COVID-19 statistics. Results Within a total of 9,807,813 posts (nearly 70% were sourced from the United States), we identified a discussion of 120 symptom-related topics and 1542 medical condition–related topics. Our classification of the health-related posts had a positive predictive value of over 80% and an average classification rate of 92% sensitivity. The 5 most commonly mentioned symptoms on social media during the study period were anxiety (in 201,303 posts or 12.2% of the total posts mentioning symptoms), generalized pain (189,673, 11.5%), weight loss (95,793, 5.8%), fatigue (91,252, 5.5%), and coughing (86,235, 5.2%). The 5 most discussed medical conditions were COVID-19 (in 5,420,276 posts or 66.4% of the total posts mentioning medical conditions), unspecified infectious disease (469,356, 5.8%), influenza (270,166, 3.3%), unspecified disorders of the central nervous system (253,407, 3.1%), and depression (151,752, 1.9%). Changes in posts in the frequency of anxiety, generalized pain, and weight loss were significant but negatively correlated with daily new COVID-19 cases in the United States (r=-0.49, r=-0.46, and r=-0.39, respectively; P<.05). Posts on the frequency of 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 (r ranged=0.39-0.48; P<.05). Conclusions COVID-19 and symptoms of anxiety were the 2 most commonly discussed health-related topics on social media from June 14 to December 13, 2020. Real-time monitoring of social media posts on symptoms and medical conditions may help assess the population’s mental health status and enhance public health surveillance for infectious disease.

Innovations in public health surveillance: An overview of novel use of data and analytic methods

“…Some of the published innovations using AI for PHS still reside within academic collaborations. One such study from the Yale School of Medicine used NLP, which applies AI methods to the interpretation of human language, to provide real-time monitoring of population health by identifying symptoms mentioned on social media platforms (( 51 )).…”

Social media for early characterization of pandemic symptoms: A qualitative analysis of patient‐reported COVID‐19 experiences

“…15 Due to the anonymous nature of these platforms, users often feel comfortable with providing in-depth description of their symptoms, sometimes, Advanced analytical approaches such as natural language processing and machine learning techniques 16 have been used to process various social media platforms. [17][18][19] Although, these advanced techniques can effectively extract "themes" and "terms" from a large volume of data, the capture of in-depth descriptions of patient-reported symptoms is often lost with automated processing of the data. Further refinement of these analytical approaches to reflect the data is possible but may be difficult to implement.…”

Dynamics of social media behavior before and after SARS-CoV-2 infection

“…From a medical point of view, it has been highlighted that the number of self-reports of symptoms displayed a temporal correlation with the number of confirmed cases (13,14). The frequency of symptoms reported on Twitter was shown to be in good agreement with the prevalence of symptoms following confirmed infections (15) and allowed to identify patterns revealing the long-term criticality of the post-acute sequelae of COVID-19 (16).…”