key: cord-0916937-aoy5bs9k
authors: Patel, Vishal R.; Gereta, Sofia; Blanton, Christopher J.; Chu, Alexander L.; Patel, Akash P.; Mackert, Michael; Zientek, David; Nortjé, Nico; Khurshid, Anjum; Moriates, Christopher; Wallingford, Gregory
title: Perceptions of Life Support and Advance Care Planning During the COVID-19 Pandemic: A Global Study of Twitter Users
date: 2022-01-22
journal: Chest
DOI: 10.1016/j.chest.2022.01.023
sha: a0d05514df6122efe11e159ef5afb1ac0ad2172e
doc_id: 916937
cord_uid: aoy5bs9k

Background The COVID-19 pandemic has presented new challenges surrounding end-of-life planning and has been associated with increased online discussion about life support. Research Question How has online communication about advance care planning (ACP) and specific life-sustaining interventions (LSIs) changed during the pandemic? Study Design and Methods Conversations on Twitter containing references to LSIs (e.g., ‘ECMO’) or ACP (e.g., ‘DNR/DNI’) were collected from January 2019 to May 2021. User account metadata was used to predict user demographic information and classify users as either organizations, individuals, clinicians, or influencers. We compared the number of impressions across these user categories and analyzed the content of tweets using natural language processing models to identify topics of discussion and associated emotional sentiment. Results There were 202,585 unique tweets about LSIs and 67,162 unique tweets about ACP. Users who were younger, men, or influencers were more likely to discuss LSIs online. Tweets about LSIs were associated with more positive emotional sentiment scores than tweets about ACP (LSIs: 0.3, ACP: -0.2; P < 0.001). Among tweets about ACP, most contained personal experiences related to the death of loved ones (27%) or discussed discrimination through DNR orders directed at the elderly and disabled (19%). Personal experiences had the greatest retweet-to-tweet-ratio (4.7), indicating high levels of user engagement. Tweets about discrimination contained the most negative net sentiment score (-0.5). Interpretation The observed increase in tweets regarding LSIs and ACP suggests that Twitter was consistently utilized to discuss treatment modalities and preferences related to intensive care during the pandemic. Future interventions to increase online engagement with ACP may consider leveraging influencers and personal stories. Finally, we identified DNR-related discrimination as a commonly held public fear, which should be further explored as a barrier to ACP completion and can be proactively addressed by clinicians during bedside goals-of-care discussions.

Advance care planning (ACP) discussions clarify a person's values and preferences about medical treatments in the event they are unable to communicate or if their clinical status rapidly deteriorates. Media coverage of life-sustaining interventions (LSIs) has greatly increased during the COVID-19 pandemic, which presents an important opportunity to integrate ACP into discussions about LSIs and further engage patients and health care professionals in this practice. [1] [2] [3] [4] It is important to distinguish between routine discussions about LSIs, which focus on specific medical interventions, and ACP discussions, which focus on aligning patient's values and goals with the specific medical interventions they recieve. 3, [5] [6] [7] Participation in ACP is important since lack of planning often leads to goal-discordant care in which patients receive LSIs that they may not have wanted had they established their preferences in advance. However, only about one-in-three US adults have completed an advance directive, owing to multiple barriers for both patients and physicians, including, but not limited to: 1) lack of understanding about ACP, 2) insufficient time during physician appointments to discuss ACP, 3) a belief that ACP does not apply to young or healthy patients, 4) reliance on family or physician-led decision making over autonomous decision making, and 5) discomfort with talking about death and dying. [8] [9] [10] [11] [12] [13] [14] Online discussions about ACP may have a significant impact on people's perceptions and behaviors related to ACP. Although the online completion of advance directives has increased nearly 5-fold during the COVID-19 pandemic, no prior studies have examined online discussions about ACP during the pandemic in relation to discussions about LSIs. 15 To characterize how online communication about ACP and LSIs may have changed during the pandemic, we studied Twitter discussions that referenced LSIs -including cardiopulmonary resuscitation (CPR), ventilators, and extracorporeal membrane oxygenation (ECMO) -and those that included more deliberate conversations related to ACP. Twitter provides a real-time source of data from over 300 million active users worldwide contributing over 500 million tweets daily and has been extensively used to identify ongoing healthcare issues. [16] [17] [18] [19] [20] [21] [22] [23] [24] Analysis of social media content may inform on public J o u r n a l P r e -p r o o f viewpoints during critical situations and accordingly could be used to tailor the messaging of public health recommendations.

We collected English-language tweets related to LSIs and ACP between January 2019, and May 2021. Twitter provides an application programming interface (API) that enables collection of historical tweet content and metadata such as retweets and geolocation. With Twitter's API, we searched for tweets containing variations of the words 'DNR' (do-not-resuscitate), 'DNI' (do-notintubate), 'advance directives', 'ECMO', 'CPR', 'high flow oxygen', 'dialysis', and 'ventilation'.

Variations for the search term 'advance directives' included 'living will', 'medical directive', and 'medical power of attorney'. Tweets containing 'DNR' or 'DNI' were grouped together as 'DNR/DNI' since most tweets referenced both terms concurrently. Tweets returned from the search terms 'DNR/DNI' and 'advance directives' were categorized as tweets about ACP. While LSIs refer to a broad set of interventions intended to prolong life following organ failure, we chose to selectively examine discussions about cardiopulmonary interventions related to the treatment of COVID-19.

Thus, only tweets referencing 'ECMO', 'CPR', and 'ventilation' were included in the LSIs category.

Twitter data reflects information users choose to share publicly, thus IRB approval was not required.

Twitter users were classified as clinicians, individuals, influencers, or organizations. A user was marked as an organization and further classified as a news, health, or government organization if either its display name or profile biography contained relevant keywords (e.g., 'department', 'organization', 'news', 'government', 'hospital', etc.). Keywords representing organizations were chosen based on review of Twitter profiles from users in the data set and keywords used in previous related studies. 17 Users were labeled as influencers if they did not meet the criteria for an organization but had at least 100,000 followers or a verified Twitter account. 17 Verified status on J o u r n a l P r e -p r o o f Twitter distinguishes authentic high-profile users from impersonators. 25 Of the remaining users,   clinicians were identified and further classified as ICU physicians, non-ICU physicians, or nurses if   profile biographies contained relevant keywords (e.g., 'MD', 'DO', 'RN', 'physician', 'surgeon', 'nurse',   ' intensivist', 'critical care', etc.). Clinicians with over 100,000 followers or verified status were classified as influencers. All remaining users were classified as individuals. The performance of our classification algorithm was compared to a manually verified random sample of 400 unique users.

The estimated overall accuracy of the algorithm was 93% (e- Figure 1 ). Finally, we applied a multimodal deep neural architecture model to predict the age and gender of non-organization accounts using profile pictures and user biographies. 26

Text parsing was used to remove components from tweets without semantic value (e.g., special characters, punctuations, URLs, etc.). 27 Each message was divided into individual words and lemmatized to the base dictionary form (e.g., 'better' reduced to 'good'). Words appearing in less than 5 tweets, auxiliary verbs, and conjunctions were excluded. Latent Dirichlet Allocation (LDA), an unsupervised machine learning approach, was then performed to analyze the content of tweets. 28 The LDA algorithm uses the frequency and cooccurrence of linguistic units to generate a probabilistic model for assigning individual tweets to distinct topics. 29 To validate findings from the LDA analysis and create representative names for each topic, we manually reviewed the identified topic areas using a commonly utilized iterative sixstep thematic analysis. 30 To evaluate the sentiment of tweets, we used VADER (Valence Aware Dictionary and sEntiment Reasoner), an established lexicon-based model for computing sentiment. 31 This tool provides a compound score for each tweet that accounts for the intensity of expressed emotion, ranging from +1 (extreme positive) to -1 (extreme negative).

All analyses were performed using MATLAB R2021a (MathWorks, Inc.). Chi-square tests were used to assess differences between types of users and the identified topics. To identify cells J o u r n a l P r e -p r o o f contributing to the significance of the chi-square results, adjusted residuals were calculated for each cell and a Bonferroni-adjusted P < 0.001 was used to indicate statistical significance. 32 Mann-Whitney U tests were used to assess differences in non-normally distributed user characteristics between groups. followed by references to 'ECMO' (14.5%) and CPR (13.5%). Tweets referencing CPR showed the smallest increase during the first peak in COVID-19 deaths, shifting from a maximum of 93 tweets/day before March 2020 to a maximum of 138 tweets/day after.

Despite a considerable increase from baseline during this period, tweets about ACP were less frequently shared than tweets about LSIs. ACP tweets reached a maximum of 472 tweets on April 8 th , 2020. After May 2020, the number of daily tweets for all search terms gradually decreased yet remained higher than the baseline prior to COVID-19.

When comparing the geolocation of created tweets across groups, LSIs were more frequently mentioned than ACP in parts of Africa and the Indian subcontinent (e- Figure 3 ). This discrepancy suggests that the terms 'advance directives' and 'DNR/DNI' may have been introduced to users in these English-speaking developing countries but are less utilized by these populations.

Twitter users posting about LSIs and ACP were classified as either organizations, influencers, individuals, or clinicians ( Table 1) . Across all types of users, more total users J o u r n a l P r e -p r o o f participated in conversations about LSIs than ACP (Figure 2) . Organizations (N=26,182) were further classified into health (34%), news (33%), and government (13%) organizations. Clinicians (N=36,158) were further classified as non-ICU physicians (20%), nurses (10%), and ICU physicians (4%). Among all users, ICU physicians had the greatest percentage of tweets referencing LSIs (85.7%), while nurses had the least (69.2%) (e- Figure 4) .

While users in both groups were mostly male, more males tweeted about LSIs than about ACP (LSIs: 68%, ACP: 56%; P < 0.0001). Most users were older than 40 years of age in both groups, however, users tweeting about LSIs tended to be younger than those tweeting about ACP.

The impact of each tweet was calculated using impressions, which quantify the number of followers that may be exposed to a user's tweets and is a commonly used metric for potential exposure. 17, 33 Impressions for each user were computed by multiplying the number of created tweets by the number of their followers. Users tweeting about LSIs had more total followers, unique tweets, and impressions, likely as this group had more overall users (e- Table 1 ). Further, influencers created more median impressions related to LSIs than about ACP (LSIs: 33,917, ACP: 24,048; P < 0.001) (Figure 2 ; e- Table 1) . Organizations, individuals, and clinicians had more impressions about ACP content than about LSIs, however these impressions were an order of magnitude lower than the impressions of influencers.

Tweets referencing LSIs had a greater median sentiment than those referencing ACP (LSIs: 0.3, ACP: -0.2; P < 0.001), shown in Figure 3 . A topic analysis was performed to further understand the content domains of conversations within the ACP group. Table 2 shows the keywords, representative tweets, and mean sentiment for identified topics. There was a significant association between types of users and identified topics within ACP tweets ( 2 = 4810, P < 0.0001) and retweets ( 2 = 37508, P < 0.0001) ( Table 3) . Organizations were more likely to share messages about NHDM, research, and calls to establish ACP, but less likely to share personal experiences, legal advice, or content related to discrimination. In contrast, individuals were more likely to tweet about personal experiences, legal advice, discrimination, and COVID-19

precautions, but less likely to mention NHDM, research, or calls to establish ACP. Similarly, clinicians were more likely to reference personal experiences and less likely to reference NHDM or calls for action. However, clinicians were more likely to retweet posts about all topics besides personal experiences, suggesting a willingness to share pre-created content about ACP over original messaging. Influencers were more likely to tweet about research and less likely to tweet about discrimination.

J o u r n a l P r e -p r o o f Tweets sharing personal experiences had the highest retweet-to-tweet ratio (4.7) across all users, suggesting that these tweets had the most support. Content related to legal advice (0.8) or public appeals to establish ACP (0.9) had the lowest retweet-to-tweet ratios.

This study tracked discussions about ACP and LSIs on Twitter and employed machine learning techniques to characterize participating users and public perceptions surrounding these topics during the COVID-19 pandemic. Our findings suggest that while the COVID-19 pandemic has increased public discussion about previously uncommonly discussed LSIs, conversations about ACP received far less viewership and were associated with more negative emotional sentiment. We also found that systemic discrimination through advance directives is a commonly cited perspective in online discussions.

The excessive demand for mechanical ventilation and ECMO services, in addition to complexities involving allocation of equipment, personnel, and resources, likely drove the observed surge in Twitter conversations related to LSIs. 34 Younger age and male gender users were more likely to contribute to these discussions. This finding may represent that younger users perceive ACP discussions as less relevant given their lower overall risk than older users. 35 All user types were less likely to tweet about ACP, which was also associated with more negative emotional sentiment. This is consistent with the perception that ACP is associated with death or dying, and that many users feel it is less appropriate to share negative emotions online. 35 While clinicians tweeting about ACP left more impressions on a per user basis, fewer clinicians talked about ACP than about LSIs.

This result suggests clinicians have the potential to promote ACP online but more often choose to participate in conversations about LSIs. These findings could be explained by previous research that suggests that, although clinicians are educated about ACP, they are often afraid of talking about death and may forget to initiate ACP conversations with patients in the rush of clinical practice. [36] [37] [38] [39] Approximately 19% of ACP tweets recognized DNR/DNI orders as a form of discrimination during the pandemic and were associated with the most negative sentiment. These findings suggest J o u r n a l P r e -p r o o f that public reservations about DNR/DNI may be rooted in fear of biased treatment of groups such as the elderly or people with disabilities. Such reservations are not unexpected, as one in five patients over the age of 50 experiences discrimination in healthcare settings, often because of new or worsening disability. 40 Additionally, older patients tend to receive shorter, less intensive care and have higher rates of DNR/DNI orders, independent of clinical status. [41] [42] [43] Unfortunately, public discourse during COVID-19 has reintroduced the potential for discriminatory behaviors, such as counting physical and mental disability as part of resource allocation criteria, blanket DNR orders for people in care homes, and Twitter hashtags attacking "baby boomers". [44] [45] [46] Clinicians may more effectively approach ACP conversations at the bedside by proactively acknowledging concerns related to systemic DNR/DNI-related discrimination, knowing that it is brought up in nearly one out of five Twitter conversations about ACP.

One potential opportunity to increase ACP-related content may be to leverage awareness campaigns like NHDM to promote messages that reach online audiences more effectively. NHDM has been hosted by the Institute of Healthcare Improvement each April since 2008 to educate both the public and providers about the importance of ACP. [47] [48] [49] Tweets containing appeals to complete advance directives peaked during every April of the study period, demonstrating existing social media awareness that might be further leveraged to change the narrative towards more positive messaging around ACP. However, the frequency of NHDM-related tweets did not increase from 2019 to 2021 despite elevated mortality during the pandemic, reflecting the challenge of connecting this campaign to real-time events. Encouraging collaboration between institutions participating in NHDM may allow for more coordinated situational responses to future crises that may result in more goal-concordant care.

Notably, tweets containing personal experiences with ACP had considerably more public support than tweets simply promoting NHDM. While these personal stories were primarily shared by individuals or clinicians, organizations participating in NHDM might produce more meaningful and effective campaigns by collaborating with these users or modifying their shared content to focus more on patient stories. There is considerable research to suggest that storytelling can improve J o u r n a l P r e -p r o o f learning, which is vital to increase awareness and improve public sentiment relating to ACP. 50 Similarly, organizations may consider collaborating with influencers to share content, since these users had the greatest impact on a per user basis. In addition, users who select to follow ACPfocused organizations are more likely to already be familiar with these resources. For example, tweets sharing research articles related to ACP had extensive public exposure, but organizations created and shared these posts more than any other user. This suggests that among followers of organizations, most users are not engaging with the tweets and only a small percentage share the original messaging. To increase user engagement and reduce unidirectional messaging, NHDM organizers may even consider directing tweets about ACP to specific users using the "@" symbol.

There are several limitations to the current analysis. First, we may have under-identified the number of tweets about LSIs or ACP. Tweets relevant to either topic may have been missed because they were not covered by our search terms, which did not account for misspellings or non-English-language tweets. Our findings may also not be generalizable to the entire public, as Twitter users are a self-selected group that may not adequately represent certain demographics, including elderly people. For international tweets from non-English speaking countries, English-speaking users could represent a biased sample of highly educated or affluent users rather than the general public.

Additionally, the use of keywords to classify organizations and clinicians may have resulted in misclassification of users that did not provide adequate description in their Twitter biography. The aforementioned limitations would have resulted in non-differential misclassification of the user types and bias our findings toward the null. Finally, there may have been tweets in the LSIs group that were focused on setting preferences about life support (e.g., "I don't want ECMO if I get COVID-19").

As we could not discriminate these tweets, our result only applies to the official use of ACP terminology. Further attempts to understand the content of discussions related to LSIs is warranted.

During the pandemic, discussions about LSIs gained more impressions and were perceived with more positive emotional sentiment than discussions about ACP, suggesting that routine J o u r n a l P r e -p r o o f discussions about life support therapies were more popular than value-based discussions surrounding life support preferences. Our findings suggest that younger individuals, men, and users with influencer status were more likely to focus on LSIs than ACP, and thus are potential targets for future interventions aimed at increasing engagement in ACP. Further, personal experiences with death were commonly shared by individuals and clinicians tweeting about ACP and had the most user engagement. Incorporating these empowering personal stories into media campaigns and influencer content may increase public exposure and re-sharing of ACP content. Finally, we identified DNR-related discrimination of vulnerable groups as a commonly held fear during the pandemic. Proactively addressing patient concerns about discrimination may be helpful for clinicians when discussing ACP at the bedside.

Further studies are warranted to better characterize the online prevalence of specific end-oflife priorities, the socioeconomic and cultural influences governing these priorities, and the evolving role of social media in promoting ACP. Given the remarkable ability of social media to disseminate information, our findings provide some guidance for the medical community to better leverage this resource to engage with the public during emergencies. 

Calls for action NHDM Discrimination

During COVID-19, Outpatient Advance Care Planning Is Imperative: We Need All Hands on Deck

Shaping End-of-Life Care: Behavioral Economics and Advance Directives

Value of Advance Care Directives for Patients With Serious Illness in the Era of COVID Pandemic: A Review of Challenges and Solutions

The importance of addressing advance care planning and decisions about do-not-resuscitate orders during novel coronavirus 2019 (COVID-19)

AGS Position Statement: Resource Allocation Strategies and Age-Related Considerations in the COVID-19 Era and Beyond

The impact of advance care planning on end of life care in elderly patients: randomised controlled trial

Advance Directives and Outcomes of Surrogate Decision Making before Death

Advance Care Planning: Ensuring Your Wishes Are Known and Honored If You Are Unable to Speak for Yourself

Future Care Planning for patients approaching endof-life with advanced heart disease: an interview study with patients, carers and healthcare professionals exploring the content, rationale and design of a randomised clinical trial

Clinician barriers and facilitators to heart failure advance care plans: a systematic literature review and qualitative evidence synthesis

Integrating Advance Care Planning Into Practice

Approximately One In Three US Adults Completes Any Type Of Advance Directive For End-Of-Life Care

Barriers to advance care planning: a qualitative study of seriously ill Chinese patients and their families

Exploring patient-reported barriers to advance care planning in family practice

Completion of Advance Directives and Documented Care Preferences During the Coronavirus Disease 2019 (COVID-19) Pandemic

Using Twitter for breast cancer prevention: an analysis of breast cancer awareness month

Twitter as a Potential Data Source for Cardiovascular Disease Research

Gender Differences in Twitter Use and Influence Among Health Policy and Health Services Researchers

Twitter to engage, educate, and advocate for global antibiotic stewardship and antimicrobial resistance

Tweaking and Tweeting: Exploring Twitter for Nonmedical Use of a Psychostimulant Drug (Adderall) Among College Students

The #PalliativeCare Conversation on Twitter: An Analysis of Trends, Content, and Caregiver Perspectives

An analysis of COVID-19 vaccine sentiments and opinions on Twitter

Cancer Communication in the Social Media Age

Twitter verification requirements -how to get the blue check

Demographic Inference and Representative Population Estimates from Multilingual Social Media Data | The World Wide Web Conference

Comparing Methods for Single Paragraph Similarity Analysis

Latent Dirichlet Allocation

Finding scientific topics

Using thematic analysis in psychology

VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text

Multiple Regression Approach to Analyzing Contingency Tables: Post Hoc and Planned Comparison Procedures

Setting minimum standards for measuring public relations effectiveness

Norms of online expressions of emotion: Comparing Facebook

Barriers to Advance Care Planning in Cancer, Heart Failure and Dementia Patients: A Focus Group Study on General Practitioners' Views and Experiences

Communication in heart failure: perspectives from older people and primary care professionals

Implementing advance care planning: a qualitative study of community nurses' views and experiences

Advance Care Planning Conversations in the Oncology Setting: Tips from the Experts

Discrimination in healthcare settings is associated with disability in older adults: health and retirement study

End of Life Care and Do Not Resuscitate Orders: How Much Does Age Influence Decision Making? A Systematic Review and Meta-Analysis

Age as a factor in do not attempt cardiopulmonary resuscitation decisions: a multicentre blinded simulation-based study

Age discrimination in out-of-hospital cardiac arrest care: a case-control study

Ageism and COVID-19: what does our society's response say about us?

Disability Discrimination, Medical Rationing and COVID-19

Boomer? Intensification of Ageism and Intergenerational Tensions on Social Media Amid COVID-19

National Healthcare Decisions Day -April 16 [Internet]. The Conversation Project

National Healthcare Decisions Day

National Healthcare Decisions Day

Storytelling to Enhance Teaching and Learning

Characteristics ACP, N (%) LSIs, N (%)