%0 Journal Article
%@ 1438-8871
%I JMIR Publications
%V 24
%N 12
%P e41517
%T Sentiment Analysis of Insomnia-Related Tweets via a Combination of Transformers Using Dempster-Shafer Theory: Pre– and Peri–COVID-19 Pandemic Retrospective Study
%A Maghsoudi,Arash
%A Nowakowski,Sara
%A Agrawal,Ritwick
%A Sharafkhaneh,Amir
%A Kunik,Mark E
%A Naik,Aanand D
%A Xu,Hua
%A Razjouyan,Javad
%+ Department of Medicine, Baylor College of Medicine, One Baylor Plaza, Houston, TX, 77030, United States, 1 713 798 4951, javad.razjouyan@bcm.edu
%K COVID-19
%K coronavirus
%K sleep
%K Twitter
%K natural language processing
%K sentiment analysis
%K transformers
%K Dempster-Shafer theory
%K sleeping
%K social media
%K pandemic
%K effect
%K viral infection
%D 2022
%7 27.12.2022
%9 Original Paper
%J J Med Internet Res
%G English
%X Background: The COVID-19 pandemic has imposed additional stress on population health that may result in a change of sleeping behavior. Objective: In this study, we hypothesized that using natural language processing to explore social media would help with assessing the mental health conditions of people experiencing insomnia after the outbreak of COVID-19. Methods: We designed a retrospective study that used public social media content from Twitter. We categorized insomnia-related tweets based on time, using the following two intervals: the prepandemic (January 1, 2019, to January 1, 2020) and peripandemic (January 1, 2020, to January 1, 2021) intervals. We performed a sentiment analysis by using pretrained transformers in conjunction with Dempster-Shafer theory (DST) to classify the polarity of emotions as positive, negative, and neutral. We validated the proposed pipeline on 300 annotated tweets. Additionally, we performed a temporal analysis to examine the effect of time on Twitter users’ insomnia experiences, using logistic regression. Results: We extracted 305,321 tweets containing the word insomnia (prepandemic tweets: n=139,561; peripandemic tweets: n=165,760). The best combination of pretrained transformers (combined via DST) yielded 84% accuracy. By using this pipeline, we found that the odds of posting negative tweets (odds ratio [OR] 1.39, 95% CI 1.37-1.41; P<.001) were higher in the peripandemic interval compared to those in the prepandemic interval. The likelihood of posting negative tweets after midnight was 21% higher than that before midnight (OR 1.21, 95% CI 1.19-1.23; P<.001). In the prepandemic interval, while the odds of posting negative tweets were 2% higher after midnight compared to those before midnight (OR 1.02, 95% CI 1.00-1.07; P=.008), they were 43% higher (OR 1.43, 95% CI 1.40-1.46; P<.001) in the peripandemic interval. Conclusions: The proposed novel sentiment analysis pipeline, which combines pretrained transformers via DST, is capable of classifying the emotions and sentiments of insomnia-related tweets. Twitter users shared more negative tweets about insomnia in the peripandemic interval than in the prepandemic interval. Future studies using a natural language processing framework could assess tweets about other types of psychological distress, habit changes, weight gain resulting from inactivity, and the effect of viral infection on sleep. 
%M 36417585
%R 10.2196/41517
%U https://www.jmir.org/2022/12/e41517
%U https://doi.org/10.2196/41517
%U http://www.ncbi.nlm.nih.gov/pubmed/36417585