Infectious diseases have occurred in the past and continue to emerge. Infectious diseases are termed “emerging” if they newly appear in a population or have existed but are increasing rapidly in incidence or geographic range [1]. Examples of emerging infectious diseases include acquired immunodeficiency syndrome, Ebola, dengue hemorrhagic fever, Lassa fever, severe acute respiratory syndrome (SARS), H1N1 flu, Zika, etc [2]. Evidence shows that emerging infectious diseases are among the leading causes of death and disability globally [3]. For instance, a 1-year estimate of the 2009 H1N1 flu pandemic shows that 43-89 million people were infected [4], and 201,200 respiratory deaths and 83,300 cardiovascular deaths were linked to the disease [5] worldwide. In addition, 770,000 HIV deaths were recorded in 2018 alone, with approximately 37.9 million people already infected with the virus globally [6]. Ebola is another deadly infectious disease that has an average case-fatality rate of about 50%, with a range of 25%-90% case-fatality rates in past outbreaks [2,7].

In December 2019, COVID-19, caused by the novel coronavirus, emerged and soon became the latest deadly infectious disease [8,9] worldwide, with more than 9.4 million confirmed cases and over 482,800 deaths in 188 countries and regions as of June 25, 2020 [10]. Hence, it was declared a pandemic by the World Health Organization. The COVID-19 pandemic has strained the global health systems and caused socioeconomic challenges due to job losses and lockdowns (and other restrictive measures) imposed by governments and public health agencies to curtail the spread of the virus. Evidence has already shown that emerging infectious diseases impose significant burden on global economies and public health [3,11-13]. To understand public concern, personal experiences, and factors that hinder or facilitate the efforts to control the spread of the COVID-19 pandemic, social media data can produce rich and useful insights that were previously impossible in both scale and extent [14].

Over the years, social media has witnessed a surge in active users to more than 3.8 billion worldwide [15], making it a rich source of data for research in diverse domains. In the health domain, social media data (ie, user comments or posts on Twitter, Facebook, YouTube, Instagram, online forums, blogs, etc) have been used to investigate mental health issues [16,17], maternal health issues [18,19], diseases [20-24], substance use [25,26], and other health-related issues [27,28]. Other domains (eg, politics, commerce, marketing, or banking) have also witnessed widespread use of social media data to uncover new insights related to election results [29-32], election campaigns [33], customer behavior and engagement [34,35], etc. Regarding the COVID-19 crisis, social media data can reveal public perceptions toward how governments and health agencies worldwide are handling the pandemic and the social, economic, psychological, and health impacts of the disease on people regardless of their geographic locations in line with various factors that hinder or facilitate the efforts to control the spread of the COVID-19 pandemic globally.

In this paper, we apply natural language processing (NLP) to understand public opinions, experiences, and issues with respect to the COVID-19 pandemic using data from Twitter, Facebook, YouTube, and three online discussion forums (ie, Archinect [36,37], LiveScience [38], and PushSquare [39]). NLP is a well-established method that has been applied in many JMIR papers and other health informatics papers to understand various health-related issues. For example, Abdalla et al [40] studied the privacy implications of word embeddings trained on clinical data containing personal health information, while Bekhuis et al [41] applied NLP to extract clinical phrases and keywords from a corpus of messages posted to an internet mailing list. Specifically, we aim to answer the following research questions (RQs):

The methodological approach used in answering our RQs are as follows:

Social media has been a rich source of data for research in many domains, including health [42]. Research that uses social media in conjunction with NLP within the health domain continues to grow and cover broad application areas such as health surveillance (eg, mental health, substance use, diseases, and pharmacovigilance), health communication, sentiment analysis, and so on [43]. For example, Park and Conway [44] used the lexicon-based approach to track prevalence of keywords indicating public interest in four health issues— Ebola, e-cigarettes, marijuana, and influenza—based on social media data. Afterward, they generated topics that explain changes in discussion volume over time using the latent Dirichlet allocation (LDA) algorithm. Similarly, Jelodar et al [45] applied LDA to extract latent topics in COVID-19–related comments and used the long short-term memory recurrent neural network technique for sentiment classification. Furthermore, Nobles et al [46] used social media data to examine the needs (including seeking health information) of the reportable sexually transmitted diseases community. Their NLP approach involves extracting the top 50 unigrams from the posts based on frequency and then generating topics using the nonnegative matrix factorization technique instead of LDA. Paul et al [47] applied the Ailment Topic Aspect Model to generate latent topics from Twitter data with the aim of detecting mentions of specific ailments including allergies, obesity, and insomnia. They used a list of key phrases to automatically identify possible systems and treatments. McNeill et al [48] investigated how the dissemination of H1N1-related advice in the United Kingdom encourages or discourages vaccine and antiviral uptake using Twitter data. They conducted an automated content analysis using the KH Coder tool (Koichi Higuchi) to explore potential topics based on frequency of occurrence and then performed a more detailed or conversational analysis to understand skepticism over economic beneficiaries of vaccination and the risks and benefits of medication based on public opinion. On the other hand, Oyebode et al [49] performed sentiment analysis on user reviews of mental health apps using the machine learning approach. They compared five classifiers (based on five different machine learning algorithms) and used the best performing classifier to predict the sentiment polarity of reviews. However, none of the aforementioned approaches considers the context in which words appear in unstructured texts, which instinctively plays a substantial role in conveying meaning.

To investigate the significance of contextual text analysis, Dave and Varma [50] compared the noncontextual n-gram chunking approach and the contextual part-of-speech (POS) chunking approach in their experimental research in the field of advertising. Although the n-gram chunking method simply extracts words of varying lengths within a sentence boundary as candidate key phrases, the POS chunking method infers the context of words using POS patterns such as one or more noun tags (NN, NNP, NNS, and NNPS) along with adjective tags (JJ) and optional cardinal tags (CD) and determiners (DT). They focused on key phrases up to a length of 6 for their experiments. Their initial assessment showed that the majority of the key phrases generated using the n-gram chunking method are not meaningful within the advertising context, hence not useful. Furthermore, they observed the impact of key phrases from both methods on the performance of classification systems based on naive Bayes, logistic regression, and bagging machine learning algorithms. Their findings revealed that systems using the POS chunking method outperformed those using the n-gram chunking method for feature extraction. We leveraged Dave and Varma’s [50] contextual method in this study and extended it to capture additional POS patterns, NLP preprocessing techniques, and sentiment scoring using a lexicon-based technique.

Finally, to uncover insights about the type of information shared on Twitter during the peak of the H1N1 (swine flu) pandemic in 2009, Ahmed et al [51] generated 8 broad themes using a coding method involving expert reviewers. Similarly, Bekhuis et al [41] involved two dentists to manually and iteratively classify clinical phrases into categories and subcategories. We also used this method in the key phrase categorization stage of our study to group related key phrases into categories or broad themes.

The main goal of this paper is to understand and reflect on people’s personal experiences and opinions with respect to the COVID-19 pandemic using social media data. To achieve this, we applied various standard and well-known computational techniques that are highlighted in the following section and summarized in Figure 1.

We used various automated techniques to collect 47,410,795 COVID-19-related or coronavirus-related comments from six social media platforms: Twitter, YouTube, Facebook, Archinect, LiveScience, and PushSquare. The following describes the techniques and the breakdown of the data collected from each platform:

Next, we applied the following NLP techniques to clean and prepare data for analysis using Python:

After the preprocessing tasks were completed, non-English and duplicated comments were removed, thereby reducing the total number of comments to 8,021,341.

Next, we randomly selected 1,051,616 comments (representing approximately 13% of the entire data set) and then extracted meaningful key phrases that conveyed the topical content of the comments. We refer to the data set containing the comments as corpus and each comment as document in the remaining parts of this paper. We focused on key phrases that are opinionated (ie, express or imply positive or negative sentiment [58]) since our goal was to determine public opinions and impact with respect to the COVID-19 pandemic. We extracted candidate key phrases from our corpus using a seven-stage NLP approach, shown in Figure 2. We implemented our approach using the Python programming language.

To derive meaningful key phrases, we defined the following regular grammar: <DT>? <JJ.*>* <NN.*>* <VB.*>? (<IN>? <DT>? <JJ.*>* <NN.*>*)? which specifies a meaningful POS pattern that the syntactic parser uses to deconstruct each sentence in the documents into their constituents [59]. Table 1 shows the various parts of speech (or syntactic categories) captured in the grammar. These syntactic categories are based on well-established POS tagging guidelines for English [60]. In the aforementioned regular grammar, the “?” and “*” characters represent “optional” and “zero or more occurrences,” respectively. Our regular grammar is aimed at generating key phrases that capture both context and sentiment of a conversation using nouns, adjectives, and verbs. Research has shown that nouns are most useful in knowing the context of a conversation (ie, what is being discussed) [61], while verbs and adjectives are important for sentiment detection [62]. Determiners and prepositions are also captured by the grammar since they usually co-occur with noun or adjective phrases (eg, a meal for six people or a hospital on the hilltop).

Next, each document is split into sentences, and then each sentence is split into tokens or words. The sentence breaking task is achieved using an unsupervised algorithm that considers abbreviations, collocations, capitalizations, and punctuations to detect sentence boundaries [63]. The tagging module associates each token with its POS. The POS tags are based on the Penn Treebank tagset [60,64], some of which are shown in Table 1. Each token is reduced to its root form, depending on its POS. This activity is called lemmatization. For example, worse and better, which are both adjectives, will become bad and good, respectively. Prior to lemmatization, each token is converted to lowercase. Although Witten et al [65] applied stemming for its tokens, we chose lemmatization over stemming since lemmatization returns root words that are always meaningful and exist in the English dictionary. Stemming, on the other hand, may return root words that have no meaning at all since it merely removes prefixes or suffixes based on a rule-based method [66].

Furthermore, the syntactic parsing module deconstructs each sentence into a parse tree and then creates chunks or phrases based on the regular grammar or POS pattern defined in the first step. In other words, the parser’s chunking process involves matching phrases composed of an optional determiner, zero or more of any time of adjective, zero or more of any type of noun, any type of verb (but optional), and an optional component. This component consists of an optional preposition, an optional determiner, zero or more of any type of adjective, and zero or more of any type of noun. The output of this stage is the candidate key phrases.

In the transformation and filtering stage, key phrases that are stop words (ie, words that are commonly used, such as the, a, an, with, in, and that) are removed from candidate key phrases using a predefined list L_stopwords compiled from multiple sources (eg, [67]). We excluded negation words, which are necessary for sentiment detection, such as not, from the list of stop words. In addition, a subset of L_stopwords were removed from the start and end of (and from within) the remaining key phrases in the candidate key phrases such that the meaning of the key phrases is preserved. Afterward, duplicates were removed from the candidate key phrases. Although previous research excluded key phrases above length 6 [50], we included key phrases up to length 10 in our analysis to avoid losing important key phrases that would have enriched insights from this paper. Hence, key phrases containing more than 10 words were removed from the candidate key phrases. Since our focus is on opinionated key phrases (ie, positive and negative key phrases), we applied a filtering technique that involves computing sentiment score for each key phrase and discarding nonopinionated key phrases.

Finally, to identify negative and positive key phrases in the candidate key phrases, the scoring module computes a sentiment score, S_score, ranging from –1 to 1 for each key phrase using the Valence Aware Dictionary for Sentiment Reasoning lexicon-based algorithm [68]. Afterward, each key phrase is assigned a polarity (negative or positive) based on the S_score using the criteria recommended by Hutto and Gilbert [68]. Specifically, a key phrase is negative if S_score<–0.05, while a key phrase is positive if S_score>0.05. A neutral key phrase (with S_score between –0.05 and 0.05) was removed from the candidate key phrases since it is not opinionated.

To answer our RQs, we categorized the final candidate key phrases into categories or broad themes using a thematic analysis approach used by Bekhuis et al [41] to classify clinical phrases into categories. In this approach, expert reviewers manually examine the key phrases and then assign them to appropriate categories. We recruited four reviewers to perform our key phrase categorization task. Specifically, we assigned the negative key phrases to a group of two reviewers (G1) and the positive key phrases to a second group of two reviewers (G2). Each reviewer independently examined the key phrases iteratively and continued to categorize related key phrases until a saturation level was reached (ie, no new categories were emerging from the key phrases). Reviewers used coding sheets in which they indicated the category each key phrase belonged to after examining it. Category names were decided by each reviewer such that a new category was created if none of the existing categories matched the key phrase being reviewed. Since key phrases are more specific than comments, the reviewers assigned each key phrase to only one category. In other words, reviewers assign a key phrase to the most appropriate category or to a new category if none of the existing categories was suitable. After categorizing the key phrases, the reviewers in each team validated each other’s work and agreed or disagreed with the category assigned to each key phrase, and offered suggestions to address every disagreement. The reviewers came together after completing their validations to apply the suggestions and ensure all category names were distinct while harmonizing names that are similar. We measured interrater reliability using the percentage agreement metric [69]. The percentage agreement score for G1 was 98.0%, while the score for G2 was 99.3%. We refer to the categories as themes and the various key phrases under each category as subthemes in the remaining part of this paper.

In this section, we discuss the results of our experiments and key phrase categorization. From the large corpus used for the experiment, 427,875 negative and 520,685 positive key phrases were automatically generated. However, the majority of these key phrases were similar; hence, the reviewers reached a saturation point (during key phrase categorization) where no new categories were emerging. In total, 18,694 negative and 19,841 positive key phrases were categorized.

Overall, 34 negative and 20 positive themes emerged after the key phrase categorization phase discussed in the Methods section. Out of the 34 negative themes, 15 were health-related, psychosocial, and social issues (which were the main focus of this paper and are shown in Tables 2-4). Table 5 shows the 15 negative themes and the corresponding number of key phrases under each theme, while Table 6 shows the negative themes and the total number of comments for each theme. Frustration due to life disruptions emerged as the top negative theme with the highest number of comments, followed by increased mortality, comparison with other diseases or incidents, nature of the disease, and harassment. On the other hand, Table 7 shows the 20 positive themes, description, and sample comments. Table 8 shows the corresponding number of key phrases under each positive theme, while Table 9 shows the total number of comments for each theme. Public awareness emerged as the top positive theme based on the number of comments, followed by spiritual support, encouragement, and charity. By identifying negative and positive themes from COVID-19–related comments, we have answered RQ1 and RQ2, respectively.

In this paper, we analyzed social media comments to uncover insights regarding people’s opinions and perceptions toward the COVID-19 pandemic using an NLP approach. Our empirical findings revealed negative and positive themes (see Tables 2-4 and Table 7) representing negative and positive impacts of the COVID-19 pandemic and coping mechanisms on the world population. To answer RQ3, we first discussed each of the negative issues (supported by research evidence) in this section and then suggest interventions to address the issues in a later section.

In this section, we suggest interventions that can help address the negative issues while drawing insights from the positive themes and relevant research evidence. This answers our RQ3.

As lockdown and physical distancing persists, people with health concerns should be able to receive medical attention without visiting a hospital. Considering the proliferation of smartphones and the current wave of global digitization, digital interventions using mobile, artificial intelligence (AI), internet of things (IoT), and virtual reality technologies have been shown to be effective for delivering remote health care (or telehealth) to patients [114-119]. This is based on our findings under the innovative research positive theme (see Table 7), which revealed global research efforts to create digital interventions using emerging technologies to address the health crisis caused by COVID-19. For example, mobile apps that detect mental health issues (eg, depression and anxiety) based on phone sensors (or wearable sensors) data and self-reports using machine learning and deep learning models, and then guide users through therapeutic procedures or treatments will be useful tools during and after the pandemic. In addition, these apps should allow users to book appointments with doctors, clinicians, or therapists and access remote medical advice, diagnosis, and treatments when necessary.

In addition, data-driven surveillance systems based on AI that predict the location of the next COVID-19 outbreak can enhance the effectiveness of containment efforts, thereby slowing the spread of the disease and reducing the case-fatality rate. Furthermore, the development of curative solutions or treatments (see Table 7) can be accelerated by leveraging machine learning and deep learning algorithms. For example, deep learning models can be used to predict chemical compounds that can halt viral replication and to suggest drugs that can be effective against the virus.

To address fitness issues during lockdown, physical activity (which is one of the positive themes in our results) programs or sessions with personalized feedback delivered through mobile apps would be helpful. Research has shown that smartphone-based health programs yield significant weight loss and increase physical activity [120]. There is also an urgent need to strengthen the global health care systems to cope with current and future pandemics through public and private investments in the health sector on an ongoing basis, such as provision of public health infrastructure that is robust and adequate for the target population and easily accessible and the provision of health insurance for everyone irrespective of financial status.

Public awareness (which emerged as the top positive theme in our findings) is also crucial for addressing negative issues arising from COVID-19 by providing timely and accurate information to people, which can be lifesaving. To reach a wider audience in an efficient manner and with less cost, public awareness can be delivered through mobile technologies, such as mobile-driven and voice-enabled conversational AI agents (or chatbots) with access to evidence-based and clinically validated resources (eg, precautionary or safety measures approved by public health agencies and organizations as well as government-approved policies or guidelines), can deliver accurate information regarding COVID-19 to people in their own native language (and in an interactive fashion) through their smartphones. These chatbots can also be made to route difficult questions to health experts for real-time feedback within the same chat session. This will help to improve people’s understanding of the disease, including how it differs from other infectious diseases, and how to protect themselves and their families from getting infected with COVID-19. In addition, people will be empowered with information required to effectively respond to fear mongering, domestic violence, and harassment. Evidence already shows the deployment of multilingual chatbots for public health awareness on COVID-19 symptoms, diagnosis, and precautionary measures [121]. Furthermore, chatbots can also respond to emergencies by contacting appropriate security agencies and emergency response teams on behalf of the users. Moreover, chatbots can deliver evidence-based therapeutic interventions to people while coordinating with specialists behind the scenes where necessary.

For people with nonsmartphone devices, public health agencies can partner with telecommunication companies to deliver COVID-19–related information directly to their phones as text messages at regular intervals. Social media is another platform through which evidence-based information can be shared with the public but may be overshadowed by fake news or false information, which is mostly shared on social media [122]. Nevertheless, official COVID-19–related channels managed by (or in conjunction with) reputable international health organizations (eg, World Health Organization) or local health authorities within the social media platforms, many of which have already been deployed, provide accurate information or updates about COVID-19 cases, fatality rates, and safety measures and guidelines [123,124]. In addition, people receive location-based updates on these channels, including emergency alerts, in a timely and effective manner.

Finally, based on our findings (see Table 7), connection with family and friends, encouragement, spiritual support, and charity can help to ease people’s frustrations, anxiety, and trauma (due to life disruptions caused by the pandemic) by addressing their emotional, physical, and spiritual needs. Evidence shows that psychological first aid and spiritual care can promote a sense of safety, calmness, self- and collective–efficacy, connectedness, and hope, as well as help people confront and overcome fear [125]. Therefore, people should endeavor to frequently communicate and follow up with loved ones (through direct voice or video calls or by using social media), encourage others in distress to stay calm and remain positive, identify people’s immediate needs and offer necessary assistance, help people find hope and meaning, and ensure the safety and comfort of vulnerable populations.

Mobile technology can play a key role in facilitating easy access to relief packages. For instance, mobile apps can be deployed with geolocation and multilingual features to help people locate the nearest food bank and charity organizations offering assistance in their geographical area. In addition, charity organizations can effectively mobilize and deliver relief items to more people, including individuals that are indisposed, based on data collected through these apps. In addition, older adults, the sick, and those in self-isolation can indicate their condition while requesting for relief so that their items can be delivered to their doorstep instead of picking it up. These apps can further integrate with other local and international charity organizations to widen the coverage of relief efforts. Recruitment of volunteers can also take place through these apps. The use data collected can be further analyzed in real time and used to predict the communities that are in dire need of assistance using machine learning or deep learning techniques.

In this study, we analyzed data from Twitter, Facebook, YouTube, and three discussion forums. However, people may have used other social media platforms such as Instagram and other discussion forums not covered in this study to disseminate information related to the COVID-19 pandemic. Therefore, our findings may not fully reflect the entire public’s opinion on social media with respect to the pandemic. Nevertheless, to have a reasonably broad understanding of public opinions, we analyzed over 1 million social media comments compared to only a few thousand commonly analyzed in many related studies. In addition, the thematic analysis used for theme categorization may be more robust; however, the large number of key phrases rendered this process time-consuming despite filtering out many irrelevant key phrases during experimentation. Accordingly, the saturation level and subsequent review and confirmation of the theme categories from a second reviewer and coder were introduced as an acceptable compromise.

In this paper, we explored the impact of the COVID-19 pandemic on people worldwide using social media data. We analyzed over 1 million comments obtained from six social media platforms using a seven-stage NLP approach to extract candidate key phrases, which we further categorized into broad themes using thematic analysis. Our results revealed 34 negative themes, out of which 15 were health-related issues, psychosocial issues, and social issues related to the COVID-19 pandemic from the public perspective. The top health-related issues were increased mortality, comparison with other diseases or incidents, nature of disease, and health concerns, while the top psychosocial issues were frustrations due to life disruptions, panic shopping, and expression of fear. The top social issues were harassment and domestic violence. Besides the negative themes, 20 positive themes emerged from our results. Some of the positive themes were public awareness, encouragement, gratitude, cleaner environment, online learning, charity, spiritual support, and innovative research. We reflected on our findings and recommend interventions that can help address the health, psychosocial, and social issues based on the positive themes and other research evidence.

Digital interventions using emerging technologies such as mobile apps, AI, IoT, and virtual reality will play a major role in delivering remote health care (ie, telemedicine or telehealth) to people in the comfort of their homes, including empowering them to self-manage their health and wellness. This will help to curb the spread of COVID-19 and future infectious diseases since many people will stay away from hospitals (or clinics) to book appointments or see doctors (or other health care professionals) unless it is absolutely necessary to visit, thereby keeping health workers and patients safe. These technologies are also useful in providing timely and accurate information about COVID-19 symptoms, diagnosis, treatment, precautionary and safety measures and guidelines, and other relevant information to target audience worldwide. Finally, digital interventions and other interventions discussed in this paper can help address the emotional, physical, and spiritual needs of people who are traumatized or frustrated by the disruptions caused by the pandemic. They also inform governments, health professionals and agencies, and institutions on how to react to the current COVID-19 pandemic and future pandemics.