Understanding the Engagement and Interaction of Superusers and Regular Users in UK Respiratory Online Health Communities: Deep Learning-Based Sentiment Analysis

Abstract

Background: Online health communities (OHCs) enable people with long-term conditions (LTCs) to exchange peer self-management experiential information, advice, and support. Engagement of "superusers," that is, highly active users, plays a key role in holding together the community and ensuring an effective exchange of support and information. Further studies are needed to explore regular users' interactions with superusers, their sentiments during interactions, and their ultimate impact on the self-management of LTCs.

Objective: This study aims to gain a better understanding of sentiment distribution and the dynamic of sentiment of posts from 2 respiratory OHCs, focusing on regular users' interaction with superusers.

Methods: We conducted sentiment analysis on anonymized data from 2 UK respiratory OHCs hosted by Asthma UK (AUK), and the British Lung Foundation (BLF) charities between 2006-2016 and 2012-2016, respectively, using the Bio-Bidirectional Encoder Representation from Transformers (BioBERT), a pretrained language representation model. Given the scarcity of health-related labeled datasets, BioBERT was fine-tuned on the COVID-19 Twitter Dataset. Positive, neutral, and negative sentiments were categorized as 1, 0, and -1, respectively. The average sentiment of aggregated posts by regular users and superusers was then calculated. Superusers were identified based on a definition already used in our previous work (ie, "the 1% users with the largest number of posts over the observation period") and VoteRank, (ie, users with the best spreading ability). Sentiment analyses of posts by superusers defined with both approaches were conducted for correlation.

Results: The fine-tuned BioBERT model achieved an accuracy of 0.96. The sentiment of posts was predominantly positive (60% and 65% of overall posts in AUK and BLF, respectively), remaining stable over the years. Furthermore, there was a tendency for sentiment to become more positive over time. Overall, superusers tended to write shorter posts characterized by positive sentiment (63% and 67% of all posts in AUK and BLF, respectively). Superusers defined by posting activity or VoteRank largely overlapped (61% in AUK and 79% in BLF), showing that users who posted the most were also spreaders. Threads initiated by superusers typically encouraged regular users to reply with positive sentiments. Superusers tended to write positive replies in threads started by regular users whatever the type of sentiment of the starting post (ie, positive, neutral, or negative), compared to the replies by other regular users (62%, 51%, 61% versus 55%, 45%, 50% in AUK; 71%, 62%, 64% versus 65%, 56%, 57% in BLF, respectively; P<.001, except for neutral sentiment in AUK, where P=.36).

Conclusions: Network and sentiment analyses provide insight into the key sustaining role of superusers in respiratory OHCs, showing they tend to write and trigger regular users' posts characterized by positive sentiment.

Keywords: asthma; bio-bidirectional encoder representations from transformers; chronic obstructive pulmonary disease; online health communities; sentiment analysis; social media; social network analysis.

Figure 1

Sentiment analysis workflow. BioBert: Bio-Bidirectional Encoder Representation from Transformers.

Figure 2

Distribution of different actions in distinct communities. (A) Asthma UK (AUK); (B) British Lung Foundation (BLF).

Figure 3

Comparisons of the number and length of posts written by superusers and regular users. (A,C) Asthma UK (AUK); (B,D) British Lung Foundation (BLF).

Figure 4

Number of posts and percentage of posts with different sentiments over time. (A,C) Asthma UK (AUK); (B,D) British Lung Foundation (BLF).

Figure 5

Average sentiment score (AVS) in different actions. (A) Asthma UK (AUK); (B) British Lung Foundation (BLF).

Figure 6

Trend of sentiment. All posts are sorted based on their publication time and regrouped into 15 bins with equal volume. Average sentiment of all posts written by regular users and superusers are calculated separately in each bin. (A) Asthma UK (AUK); (B) British Lung Foundation (BLF).

Figure 7

Sentiment trend in interactions between superusers and regular users. Average sentiment of regular users when replying to threads started by other regular users and superusers are calculated separately in each bin. (A) Regular users' sentiments Asthma UK (AUK); (B) Regular users' sentiments British Lung Foundation (BLF).

Figure 8

Sentiment trend in replies of superusers to regular users. Percentages of replies with corresponding sentiments are computed and shown in each bin. (A,C,E) Asthma UK (AUK); (B,D,F) British Lung Foundation (BLF).