We collected coronavirus-related tweets between February 2, 2020, and March 15, 2020, using the Twitter standard search application programming interface (API) consisting of a set of predefined search terms (“corona,” “2019-nCov,” and “COVID-19”), which are the most widely used scientific and news media terms relating to the novel coronavirus. We extracted and stored the text and metadata of the tweets using the time stamp, number of likes and retweets, and user profile information including the number of followers. We stored the tweets in a database table, where the primary key of the table was tweet ID. As a result, the duplicates were not stored in our database. Only English language tweets were collected in the study. Since the metadata of tweets such as the number of likes and retweets might change over time, we recollected the updated metadata of the tweets at the end of the study period using the tweet IDs of the already collected tweets. Twitter standard search API allows the access of old tweets using tweet IDs. We used the Tweepy Python (Python Software Foundation) library for accessing the Twitter API and PostgreSQL (PostgreSQL Global Development Group) database for storing the collected tweets.

We identified non-English tweets using the language field in the tweets metadata and removed them from the analysis. We identified and removed retweets from the analysis. We also removed punctuation, stop words such as an and the, and nonprintable characters such as emojis from the tweets. We normalized Twitter user mentions by converting, for example, “@Alaa” to “@username.” Furthermore, various forms of the same word (eg, travels, traveling, and travel’s) were lemmatized by converting them to the main word (eg, travel) using the WordNetLemmatizer module of the Natural Language Toolkit Python library. The data preprocessing is depicted in Figure 1. Following the terms and conditions, terms of use, and privacy policies of Twitter, all data were anonymized and were not reported verbatim to any third party.

The processed tweets were analyzed using word frequencies of single words (unigram) and double-word (bigrams) combinations, and they were visualized through word clouds to identify the most common topics. In addition, we used the topic modeling technique [13] to identify the most common topics in the tweets. Topic modeling is an unsupervised machine learning technique that can find clusters in a collection of documents (tweets in this case). We used the latent Dirichlet allocation (LDA) algorithm from the Python sklearn package. LDA requires a fixed set of topics, where each topic is represented by a set of words. The objective of LDA is to map the given documents to the set of topics so that the words in each document are mostly captured by those topics. LDA is a widely used topic modeling algorithm. We used it to find natural clusters in the language of tweets. We applied topic modeling by specifying the number of topics required by the LDA to separate the set of tweets into various clusters. Based on our previous work, we selected 30 to be the number of topics for running the LDA [14].

We took the top representative words of each of the 30 topics produced by the LDA topic modelling algorithm (see LDA output in Multimedia Appendix 1) and the common words from the word cloud (see word cloud in Multimedia Appendix 2) and manually analyzed both sets of words. From this manual analysis, the authors reached a consensus on 12 topics and associated terms, unigram and bigram, for each topic (see associated terms for each topic in Multimedia Appendix 3). These terms were used to classify tweets, using a rule-based classification script, into different topics and compute the prevalence of each topic.

Next, we developed a rule-based classification script written in Python to check for the presence of any of the preidentified unigrams and bigrams in each tweet. The classification script used a simple string-matching technique to see if a given tweet contains the selected keywords of the topics. A tweet that contained a selected keyword related to a certain topic was classified as belonging to that topic.

We also performed other analyses such as sentiment analysis, which extracts the mean number of retweets, likes, and followers for each topic and then calculates the interaction rate for each topic. The sentiment analysis was performed on the tweet text using the Python textblob library. The sentiment score varied between –1.0 to 1.0, with –1.0 as the most negative text and 1.0 as the most positive text. We calculated the mean sentiment and the mean number of likes, retweets, and followers for each topic. We also calculated the interaction rate for each topic by summing the total number of retweets and likes per topic divided by the sum of the total number of followers per topic. These measures provided additional insight into the topics and users who posted in these topics.

As shown in Figure 2, a total of 2,787,247 tweets were obtained between February 2, 2020, and March 15, 2020. Of these tweets, 1,636,422 (58.71%) non-English tweets were removed. Of the 1,150,825 remaining English tweets, 735,182 ‬(63.88%) retweets were excluded. A further 248,570 (21.60%) tweets with no coronavirus-related terms in the text were also removed. These tweets were captured by Twitter API either because the name or the profile description of users matched the search terms. Accordingly, the study analyzed 167,073 unique tweets from 160,829 unique users.

As shown in Table 2, the mean of sentiment was positive in all topics except two: deaths caused by COVID-19 and increased racism. The highest mean of positive sentiments was for the eating meat topic, followed by the wearing masks topic. The highest mean of negative sentiments was for “deaths caused by COVID-19” topic.

The mean of followers for tweeters who posted the collected tweets ranged from 2878 (in increased racism) to 13,361 followers (in economic losses). The economic loss topic had the highest mean of likes. On the other hand, travel ban and warning-related topics had the lowest mean of likes. The mean of retweets for the collected tweets varied between 0.89 (for panic buying) and 7.11 (for eating meat). The lowest interaction rate was for panic buying–related tweets, and the highest interaction rate was for racism-related tweets followed by bioweapon-related tweets and eating meat–related tweets (Table 2).

User mentions were the most common in China-related tweets, but they were the least common in racism-related tweets (Table 2). Similarly, link sharing was the most common in China-related tweets, whereas they were the least common in racism-related tweets (Table 2). Multimedia Appendix 4 shows more descriptive statistics (ie, medians, variances, standard deviations, maximums, and minimums) for all previously mentioned measures.

Users on Twitter discussed 12 main topics across four main themes related to COVID-19 between February 2, 2020, and March 15, 2020. User mentions and link sharing were the most common in the analyzed tweets. These findings might demonstrate that users on Twitter are interested in notifying or warning their friends and followers about COVID-19. These interpersonal communications indicate that people bond around the topic of COVID-19 on Twitter.

Users on Twitter also focused on the impact of coronavirus on people and countries. Specifically, numerous tweets were posted on the number of deaths linked to the coronavirus. Furthermore, the emotional and psychological impact of the coronavirus was mentioned in many tweets. Users on Twitter may show their fear and stress about COVID-19 and the lack of vaccine treatment options to prevent it or specific antiviral treatments [15]. However, the sensationalistic use of Twitter can be a great challenge for public health and outbreak response efforts because of the wild spread of misinformation and conspiracy theories [16]. The infectious outbreak of “fake news” and “distorted evidence” in the digital world can create mass panic and cause damaging and devastating consequences in the real world, distorting evidence and impeding the response efforts and activities of health care workers and public health systems [17].

Additionally, the economic impact of COVID-19 on companies and countries were discussed in several tweets. Tweeters might talk about the economic impact of COVID-19 due to, for example, temporary closures of major fast-food chains and retailers (eg, McDonald’s, KFC, Apple, and Adidas) [18], decreases in auto sales, drops in oil demand, production delays such as with the iPhone, the canceling or postponing of sporting events such as the Formula One World Championship, or decreases in airline revenues due to flight cancellations [18,19]. It has been estimated that the spread of COVID-19 could cost the worldwide economy a total of US $2.7 trillion [20]. The last impact of COVID-19 discussed by Twitter users was travel. This topic might have been common because most countries have banned travel from and to countries that confirmed the presence of COVID-19 inside their borders.

Tweets also focused on two possible sources of the coronavirus: the eating of meat and a Chinese biological military laboratory. Tweeters mentioned two main methods used to decrease the spread of COVID-19: masks and quarantine. The first method (masks) was discussed frequently on Twitter mainly due to the face mask shortage reported in several countries (eg, China, the United Kingdom, and the United States). The quarantine was a common topic in tweets because it was the first step that countries applied to control the outbreak of COVID-19.

Several strengths and limitations can be attributed to this study analyzing tweets related to the recent COVID-19 outbreak. In this study, no geographical restrictions were applied on the tweets analyzed considering the worldwide spread of the disease. However, the study only analyzed tweets in the English language, which may limit the generalizability of the findings about this worldwide outbreak. In addition, given that the Twitter standard search API does not allow researchers to obtain tweets posted more than 1 week ago [25], we could not get COVID-19-related tweets posted before February 2, 2020. Thus, the findings may not be generalizable to that period. Moreover, this study could not collect tweets from accounts marked as private. Therefore, findings may not represent all the topics discussed by users on Twitter related to COVID-19. Only posts on Twitter were analyzed in this study, thereby, our findings may not be generalizable to other social media platforms. Furthermore, the findings reported in this study are limited to only those that have access to and use Twitter. Therefore, caution is advised before assuming the generalizability of the results, as Twitter is not used by everyone in the population.

The COVID-19 pandemic has been affecting many health care systems and nations, claiming the lives of many people. As a vibrant social media platform, Twitter projected this heavy toll through the interactions and posts people made related to COVID-19. It is clear that coordinating public health crisis response activities in the real world and online is paramount, and should be a top priority for all health care systems. We need to build more national and international detection and surveillance systems to detect the spread of infectious diseases and combat the fake news that is usually accompanied by these diseases.