JMIR Publications

ABSTRACT

Background:

The coronavirus (COVID-19) pandemic is a globally, rapidly spreading event with close to 2.5 million cases as of mid-April, representing an outbreak of historical scope and one with an accelerating trajectory. However, there are ongoing concerns about the accuracy of COVID-19 case counts due to issues such as lack of access to testing and difficulty in measuring recoveries.

Objective:

The aims of this study were to detect and characterize user-generated conversations of COVID-19-related symptoms, experiences with access to testing, and mentions of recovery using an unsupervised machine learning approach.

Methods:

Tweets were collected from the Twitter public API from March 3-20 filtered for general COVID-19-related keywords and then further filtered for terms related to COVID-19 symptoms. After data cleaning and processing data, tweets were analyzed using an unsupervised machine learning approach called the biterm topic model (BTM), where groups of tweets containing the same word-related themes were separated into topic clusters related to COVID-19 symptoms, testing, and recovery conversations. Tweets in these clusters were then extracted and manually annotated for content analysis and then analyzed for statistical and geographic characteristics.

Results:

A total of 4,492,954 tweets were collected that contained COVID-19-related symptom terms. After using BTM to identify relevant COVID-19 clusters and removing duplicate tweets, we identified a total of 3,465 (<1%) tweets that included user generated conversations about experiences perceived to be related to COVID-19. These tweets were grouped into five main categories including first and second-hand reports of COVID-19-related symptoms, symptom reporting concurrent with lack of access to testing, discussion of recovery, confirmation of negative COVID-19 diagnosis after receiving testing, and users recalling past symptoms and questioning whether they had been previously infected with COVID-19. Co-occurrence of themes was statistically significant for users reporting symptoms with lack of testing and with discussion of recovery. Sixty-three percent (n=1112) of tweets with geospatial coordinates were from the U.S.

Conclusions:

In this study, we analyzed Twitter for the purposes of characterizing conversations regarding self-reporting of COVID-19-related symptoms, access to testing, and experiences with purported recovery for the purposes of digital contact tracing. It appears that many users reported COVID-19-related symptoms, but never got tested due to lack of access. However, it is unclear how many of these users were actual cases and in the absence of further testing, accurate case estimations may never be known. Future studies should continue to explore the utility of social media and other forms of electronic data to estimate COVID-19 disease severity.