Emotional Analysis of Tweets About Clinically Extremely Vulnerable COVID-19 Groups

Abstract

Background Clinically extremely vulnerable (CEV) individuals have a significantly higher risk of morbidity and mortality from coronavirus disease 2019 (COVID-19). This high risk is due to predispositions such as chronic obstructive pulmonary disease (COPD), diabetes mellitus, hypertension, smoking, or extreme age (≥75). The initial COVID-19 preventive measures (use of face masks, social distancing, social bubbles) and vaccine allocation prioritized this group of vulnerable individuals to ensure their continued protection. However, as countries start relaxing the lockdown measures to help prevent socio-economic collapse, the impact of this relaxation on CEVs is once again brought to light. In this study, we set out to understand the impact of policy changes on the lives of CEVs by analyzing Twitter data with the hashtag #highriskcovid used by many high-risk individuals to tweet about and express their opinions and feelings. Methodology Tweets were extracted from the Twitter API between March 01, 2022, and April 21, 2022, using the Twarc2 tool. Extracted tweets were in English and included the hashtag #highriskcovid. We evaluated the most frequently used words and hashtags by calculating term frequency-inverse document frequency, and the location of tweets using the tidygeocoder package (method = osm). We also evaluated the sentiments and emotions depicted by these tweets using the National Research Council sentiment lexicon of the Syuzhet package. Finally, we used the latent Dirichlet allocation algorithm to determine relevant high-risk COVID-19 themes. Results The vast majority of the tweets originated from the United States (64%), Canada (22%), and the United Kingdom (4%). The most common hashtags were #highriskcovid (25.5%), #covid (6.82%), #immunocompromised (4.93%), #covidisnotover (4.0%), and #Maskup (1.40%), and the most frequently used words were immunocompromised (1.64%), people (1.4%), disabled (0.97%), maskup (0.85%), and eugenics (0.85%). The tweets were more negative (19.27%) than positive, and the most expressed negative emotions were fear (13.62%) and sadness (12.47%). At the same time, trust was the most expressed positive emotion and was used in relation to belief in masks, policies, and health workers to help. Finally, we detected frequently co-tweeted words such asmass and disaster, deadly and disabling, high and risk, public and health, immunocompromised and people, mass and disaster, and deadly and disabling. Conclusions The study provides evidence regarding the concerns and fears of high-risk COVID-19 groups as expressed via social media. It is imperative that further policies be implemented to specifically protect the health and mental wellness of high-risk individuals (for example, incorporating sentiment analyses of high-risk COVID-19 individuals such as this paper to inform the evaluation of already implemented preventive measures and policies). In addition, considerable work needs to be done to educate the public on high-risk individuals.

Keywords: cevs; covid-19; face masks; healthcare policy; highriskcovid; sentiment analysis; twitter.

Figure 1. The analytic workflow used in this paper.

Figure 2. Top 10 hashtags used in association with #highriskcovid tweets.

The Y axis is the percentage of each hashtag relative to all analyzed tweets, while the X axis contains the top 10 hashtags associated with #highriskcovid tweets.

Figure 3. Top 10 words used in #highriskcovid tweets.

The Y axis is the percentage of each word relative to all analyzed tweets, while the X axis contains the top 10 words used in #highriskcovid tweets.

Figure 4. Word cloud of #highriskcovid words with a frequency of occurrence greater than 25.

Figure 5. Countries of origin of tweets analyzed in this study.

The Y axis contains the top 25 countries with the most #highriskcovid tweets, while the X axis is the percentage of each country relative to the whole tweet.

Figure 6. Sentiment analysis of #highriskcovid tweets.

The Y axis is the percentage of sentiments of all analyzed tweets, while the X axis refers to the sentiments (negative or positive) in #highriskcovid tweets.

Figure 7. Word cloud of the top 100 negative #highriskcovid words.

Figure 8. Word cloud of the top 100 positive #highriskcovid words.

Figure 9. Emotional quotient analysis of #highriskcovid tweets.

The Y axis is the percentage of sentiments of all analyzed tweets, while the X axis refers to the emotion in #highriskcovid tweets.

Figure 10. Comparative word cloud of the #highriskcovid words and the emotions they portray.

Figure 11. Network analysis of common co-occurring #highriskcovid words.

The nodes represent words, while the edges (lines) are the weight of the connection between the nodes.

Figure 12. Modeled topics and their associated frequent terms among the analyzed #highriskcovid tweets.

Figure 13. Five-fold cross-validation of topic modeling.

The Y axis contains the perplexity score, while the X axis contains the candidate number of topics.  We identified a topic number between 10 to 17 to be the optimal range based on the perplexity score.

Figure 14. Result of LDAtuning using Griffiths2004 and Arun2010 metrics.

The Y axis contains the metric unit, while the X axis contains the candidate number of topics. We identified a topic number between 10 to 26 (Griffith) and 10-25 (Arun) to be the optimal range based on the perplexity score.