key: cord-0056062-hjywcd10
authors: Hussain, Amir; Sheikh, Aziz
title: Opportunities for Artificial Intelligence–Enabled Social Media Analysis of Public Attitudes Toward Covid-19 Vaccines
date: 2021-02-05
journal: NEJM Catal Innov Care Deliv
DOI: 10.1056/cat.20.0649
sha: 48604cd4cbe172b7b55c4b0fc6f02b4e60740aca
doc_id: 56062
cord_uid: hjywcd10

Artificial intelligence can enable real-time analysis of public attitudes, and their demographic determinants, from social media and linked Web platforms. This analysis offers the opportunity to track changing public sentiments and develop proactive two-way communication strategies. In the context of Covid-19, an iterative learning cycle can help maximize vaccine uptake across demographic communities by identifying and addressing unforeseen areas of public concern.

Social media usage has significantly increased during the pandemic (for example, by up to 37% for Facebook6), and plays an important role in providing social support, including for those suffering from long-term Covid-19 consequences.7 Social media data is largely unstructured, but natural language processing (NLP) and machine learning (ML)8 can analyze it on a large scale and identify trends in public opinion. There is currently growing interest in utilizing established NLP techniques such as sentiment analysis and topic modeling,9 in conjunction with ML, to mine social media data for health care applications.10 Sentiment analysis uses computational methods to identify opinions in text, audio, and/ or video9 , 11 to determine the author's attitudes toward topic(s) of discussion. These can be expressed as polarities (positive, negative, neutral), or feelings (interested versus uninterested). A complementary approach is stance detection12 where a stance label (favorable, against, or none) is assigned to a post regarding a specific target, whether or not that target is explicitly addressed in the post. For example, if someone asserts that a disease can be devastating, it can be inferred that the author has a favorable stance toward a vaccine for that disease. Topic modeling9 enables discovery of related text, such as topics and themes.

These approaches are currently underutilized in health care. Social media analysis holds significant untapped potential to inform public policy research.

Social media platforms can enable real-time assessment of public confidence in Covid-19 vaccination. Changing levels of public interest and sentiment can be tracked for a range of concerns relating to vaccine safety and effectiveness, as well as trust in science, pharmaceutical companies, and governments. 13 The impact of policies and messaging can be monitored to assess public perceptions around decisions on prioritization and equitability.

Suitable search strategies can be developed to capture social media conversations and attitudes toward Covid-19 vaccinations. Anonymized data can be sourced from a wide range of social media and Web forums, such as Twitter, Facebook, news, moderated blogs, forums, and reviews. Some platforms such as Twitter give geographical locations for posts at a regional (e.g., city) level, while others such as Facebook Crowdtangle14 do so at national levels.

Social media "dashboards" for policy makers15 could analyze and visualize "mentions" of specific topics or keywords with the highest levels of engagement (that is, the number of times a topic or keyword is mentioned). These analyses can assess trends in the volume and sentiment of mentions for each social media platform. NLP algorithms, such as topic modelling and sentiment word clouds (visual representations of how frequently words are mentioned, using font size to indicate relative frequency), can be utilized for time periods of interest. Insights into the content of discussions can be gleaned, particularly around inflection points on a trend (for example, correlating changes with specific events or the circulation of particular pieces of misinformation). Policy experts, as part of "Expert-in-the-loop" NLP frameworks, can identify common positive and negative themes surrounding vaccination issues, through their manual reading of random subsets of social media and related news posts. Views of underrepresented online groups can also be identified and analyzed. Relevant dashboard data sets and outputs can be anonymized and presented as statistical aggregates to ensure compliance with national privacy laws.

Open challenges include the urgent need to develop NLP techniques to detect the sources of viral spread of fake news and misinformation, including deceptive and sarcastic language sometimes used by vaccine skeptics. This effort will require new multilingual vaccination lexicons (or dictionaries mapping commonly used terms to positive, negative, and neutral sentiments), to capture cultural differences in language use across communities. These lexicons will also need to keep up with the evolution of pandemic-related terminology (for example, the term "Rona" sometimes used as a nickname for Covid-19, or terms like "vaccine nationalism" or "viral anxiety"). State-of-the-art deep learning-based NLP models, such as Deep Bidirectional Transformers,16 can be trained through transfer learning and fine-tuning for enhanced language understanding. The use of social-network analysis techniques9 , 17 can help identify sources of misinformation and their patterns of spreading, which can shed light on the anatomy of key strategies used by those who intend to mislead and misinform, such as use of emotion and anecdotes to impact rational decision-making.2

Users with varied demographic characteristics are likely to engage differently with social media platforms on vaccination-related topics. Established AI approaches11 , 17 , 18 can enable policy makers to monitor how fake and genuine messages resonate with specific audiences. These insights can be used to help craft messages more likely to resonate with diverse audiences. In contrast with health care, such targeted messaging strategies aimed at different social groups are already extensively and successfully used in commercial marketing and advertising.18

Recent research18 has demonstrated the potential of AI-enabled demographic analysis to estimate age, gender, ethnicity, and geographic origin of social media profiles, using a combination of selfreported location data, names, social media photographs, and inferential statistics of national census data. This can help identify and categorize demographically distinct groups, such as the elderly, ethnic minorities, and incarcerated people,19 who are likely to experience disproportionate effects of Covid-19. Such analysis can inform development of real-time monitoring metrics for demographic-level engagement and more effective communication strategies to promote diversity and inclusion in vaccination campaigns.

Other challenges include linking social media with external trustworthy sources of data such as polling, surveys, news, and clinical sources. This linkage is important because social media users may not be representative of the population at large, particularly in the context of Covid-19 vaccinations that are likely to be preferentially targeted at high-risk groups (including communities who have historically lower rates of vaccination uptake13). In such cases, linked analysis with multiple social media and Web platforms, and use of network metadata such as likes and retweets, can extend the study population.

The linked analysis is also important to help reduce bias and increase trust20 in AI model predictions, because different demographics (ages, political affiliations, socioeconomic status) tend to use different platforms. Corroborated findings from independent sources can inform linkage decisions. For example, YouGov polls have shown that almost 1 in 5 Americans and 1 in 6 UK respondents state that they would be unlikely to get vaccinated against Covid-19.21 This reluctance can be attributed in part to concerns about politicization and widespread misinformation. Linked social media analysis can corroborate these findings and shed light on common determinants of vaccine hesitancy. These insights can inform more nuanced questionnaire designs around vaccine confidence, as part of interdisciplinary mixed-methods studies,22 integrating quantitative AI with qualitative surveys, interviews, and ethnographic approaches.

Social media and Web analysis at scale can help address the need to develop a benchmark metric of confidence, and a potential baseline for comparison, regarding variations in public knowledge and sentiment based on location or over time. These metrics can enable development of a realtime early warning system to prompt appropriate interventions to counteract misinformation and help mitigate the risks of decline in vaccine confidence and acceptance. Such a system can underpin more effective vaccine policy making, because policy makers can assess public mood and attitudes, develop interventions as needed, and measure their impact (Figure 1 ). The evolving iterative policies will be based on more informed and transparent community engagement, and will help transform current one-way communication strategies to two-way debates. These debates will promote urgently needed open discussions around complex Covid-19 vaccine deployment issues, in the absence of complete information about immune response and duration of immunity, repeated vaccination, multiple vaccines, transmission dynamics, and microbiological and clinical characteristics of the vaccine.2

Global efforts to develop Covid-19 vaccines have advanced rapidly. Harnessing the opportunities offered by AI-enabled high-value decision-making and linked analysis can address the range of behavioral factors underpinning vaccine confidence and trust, and further our understanding of their evolving impact on uptake. By analyzing social media and Web forum responses to ongoing vaccine trials and vaccine availability in real time, we can identify and prioritize demographic groups and regions where more targeted confidence and trust building is needed. Adopting continuously informed and proactive policy making and communication strategies will promote participatory dialogues to build support for ethical principles and maximize the deployment and uptake of life-saving vaccines.

Professor, School of Computing, Edinburgh Napier University, Edinburgh, United Kingdom Aziz Sheikh, MD Professor, Usher Institute, University of Edinburgh, Edinburgh, United Kingdom

Protect the NHS" in the United Kingdom

Covid-19 vaccine deployment: behaviour, ethics, misinformation and policy strategies

When Will We Have a Vaccine?" -Understanding questions and answers about Covid-19 vaccination

Evaluating the predictability of medical conditions from social media posts

Comparing Social media and Google to detect and predict severe epidemics

WhatsApp has seen a 40% increase in usage due to Covid-19 pandemic

Long-term consequences of COVID-19: research needs

Machine learning and the pursuit of high-value health care

A structural topic modeling-based bibliometric study of sentiment analysis literature

Mining social media data for biomedical signals and healthrelated behavior

Semi-supervised learning for big social data analysis

A novel approach to stance detection in social media tweets by fusing ranked lists and sentiments

Mapping global trends in vaccine confidence and investigating barriers to vaccine uptake: a large-scale retrospective temporal modelling study

CrowdTangle platform and API

COVID-19 AI-powered dashboard for sentiment and opinion mining in social media platforms

Pre-training of deep bidirectional transformers for language understanding

Social network-based distancing strategies to flatten the COVID-19 curve in a post-lockdown world

Identifying social media user demographics and topic diversity with computational social science: a case study of a major international policy forum

Covid-19 vaccine trials and incarcerated People -the ethics of inclusion

AI-enabled clinical decision support software: a trust and value checklist for clinicians

Why vaccine rumours stick-and getting them unstuck

Acceptability of artificial intelligence (AI)-led chatbot services in healthcare: A mixed-methods study