Twitter Sentiment Analysis of Long COVID Syndrome

Abstract

Background Long COVID syndrome originated as a patient phrased terminology which was initially used to describe a group of vague symptoms that persisted after recovering from COVID-19. However, it has moved from a patient lingo to a recognized pathological entity which refers to a group of symptoms that lasts weeks or months after the COVID-19 illness. The novelty of this condition, the inadequacy of research on long COVID syndrome, and its origin as a patient-coined terminology necessitated exploring the disease's sentiments and conversations by analyzing publicly available tweets. Method Tweets were extracted using the Twarc2 tool for tweets in the English language with the keywords (long COVID syndrome, long COVID, post-COVID syndrome, post-acute sequelae of SARS-CoV-2, long-term COVID, long haulers, and chronic COVID syndrome) between March 25, 2022, and April 1, 2022. The analyses included frequency of the keywords, sentiment analysis, and topic modeling to identify and explore discussion topics over time. A natural language approach and the latent Dirichlet allocation algorithm were used to determine the most shared tweet topics, categorize clusters, and identify themes based on keyword analysis. Results The search yielded 62,232 tweets. The tweets were reduced to 10,670 tweets after removing the duplicates. The vast majority of the tweets originated from the United States of America (38%), United Kingdom (30%), and Canada (13%), with the most common hashtags being #longcovid (36%) and #covid (6.36%), and the most frequently used word being people (1.05%). The top three emotions detected by our analysis were trust (11.68%), fear (11.26%), and sadness (9.76%). The sentiment analysis results showed that people have comparable levels of positivity (19.90%) and negativity (18.39%) towards long COVID. Conclusions Our analysis revealed comparable sentiments about long COVID syndrome, albeit slightly positive. Most tweets connoted trust (positive), fear (negative), and sadness (negative). These emotions were linked with concerns about the infection, pandemic, chronic disability, and governmental policies. We believe this study would be important in guiding information dissemination and governmental policy implementation necessary in tackling long COVID syndrome.

Keywords: health informatics; long covid syndrome; long haulers; sentiment analysis; twitter.

Figure 1. Countries of origin of tweets analyzed in this study

The Y-axis contains the top 25 countries with the most 'long COVID syndrome' tweets, while the X-axis is the percent of each country relative to the whole tweet.

Figure 2. Top ten hashtags used in association with long COVID tweets

The Y-axis is the percent of each hashtag relative to all analyzed tweets, while the X-axis contains the top ten hashtags used in association with long COVID tweets.

Figure 3. Top ten words used in long COVID tweets

The Y-axis is the percent of each word relative to all analyzed tweets, while the X-axis contains the top ten words used in long COVID tweets.

Figure 4. Word cloud of long COVID words with a frequency of occurrence greater than twenty-five

Figure 5. Sentiment analysis of long COVID tweets

The Y-axis is the percent of sentiments of all analyzed tweets, while the X-axis refers to the sentiments (negative or positive) in long COVID tweets.

Figure 6. Word cloud of the top 100 negative long COVID words

Figure 7. Word cloud of the top 100 positive long COVID words

Figure 8. Emotion quotient analysis of long COVID tweets

The Y-axis is the percent of sentiments of all analyzed tweets, while the X-axis refers to the emotion in long COVID tweets.

Figure 9. Comparative word cloud of the long COVID words and portrayed emotions

Figure 10. A network analysis of common co-occurring long COVID words

The nodes represent words while the edges (lines) are the connection's weight between the nodes. The thicker the lines, the stronger the connection.

Figure 11. Modeled topics and their associated frequent terms among the analyzed long COVID syndrome tweets analyzed

Figure 12. Result of LDA tuning using "Griffiths2004" and "Arun2010" metrics

The Y-axis contains the metric unit while the X-axis contains the candidate number of topics. We identified topic numbers 28-30 (Griffith) and 20-30 (Arun) to be the optimal range. Based on both results, we set our K (topic number) at 28. LDA - latent Dirichlet allocation