JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Original Paper Stroke Survivors on Twitter: Sentiment and Topic Analysis From a Gender Perspective Alejandro Garcia-Rudolph1,2,3, PhD; Sara Laxe1,2,3, PhD; Joan Saurí1,2,3, PhD; Montserrat Bernabeu Guitart1,2,3, MD 1 Institut Guttmann Hospital de Neurorehabilitacio, Badalona, Spain 2 Universitat Autònoma de Barcelona, Bellaterra (Cerdanyola del Vallès), Spain 3 Fundació Institut d’Investigació en Ciències de la Salut Germans Trias i Pujol, Badalona, Spain Corresponding Author: Alejandro Garcia-Rudolph, PhD Institut Guttmann Hospital de Neurorehabilitacio Camí de Can Ruti, s / n Badalona, 08916 Spain Phone: 34 934977700 Email: alejandropablogarcia@gmail.com Abstract Background: Stroke is the worldwide leading cause of long-term disabilities. Women experience more activity limitations, worse health-related quality of life, and more poststroke depression than men. Twitter is increasingly used by individuals to broadcast their day-to-day happenings, providing unobtrusive access to samples of spontaneously expressed opinions on all types of topics and emotions. Objective: This study aimed to consider the raw frequencies of words in the collection of tweets posted by a sample of stroke survivors and to compare the posts by gender of the survivor for 8 basic emotions (anger, fear, anticipation, surprise, joy, sadness, trust and disgust); determine the proportion of each emotion in the collection of tweets and statistically compare each of them by gender of the survivor; extract the main topics (represented as sets of words) that occur in the collection of tweets, relative to each gender; and assign happiness scores to tweets and topics (using a well-established tool) and compare them by gender of the survivor. Methods: We performed sentiment analysis based on a state-of-the-art lexicon (National Research Council) with syuzhet R package. The emotion scores for men and women were first subjected to an F-test and then to a Wilcoxon rank sum test. We extended the emotional analysis, assigning happiness scores with the hedonometer (a tool specifically designed considering Twitter inputs). We calculated daily happiness average scores for all tweets. We created a term map for an exploratory clustering analysis using VosViewer software. We performed structural topic modelling with stm R package, allowing us to identify main topics by gender. We assigned happiness scores to all the words defining the main identified topics and compared them by gender. Results: We analyzed 800,424 tweets posted from August 1, 2007 to December 1, 2018, by 479 stroke survivors: Women (n=244) posted 396,898 tweets, and men (n=235) posted 403,526 tweets. The stroke survivor condition and gender as well as membership in at least 3 stroke-specific Twitter lists of active users were manually verified for all 479 participants. Their total number of tweets since 2007 was 5,257,433; therefore, we analyzed the most recent 15.2% of all their tweets. Positive emotions (anticipation, trust, and joy) were significantly higher (P<.001) in women, while negative emotions (disgust, fear, and sadness) were significantly higher (P<.001) in men in the analysis of raw frequencies and proportion of emotions. Happiness mean scores throughout the considered period show higher levels of happiness in women. We calculated the top 20 topics (with percentages and CIs) more likely addressed by gender and found that women’s topics show higher levels of happiness scores. Conclusions: We applied two different approaches—the Plutchik model and hedonometer tool—to a sample of stroke survivors’ tweets. We conclude that women express positive emotions and happiness much more than men. (J Med Internet Res 2019;21(8):e14077) doi: 10.2196/14077 KEYWORDS stroke; emotions; Twitter; infodemiology; infoveillance; sentiment analysis; topic models; gender http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 1 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al #Stroke Introduction The expansion of social media has changed the way in which General Background patients, physicians, and other health care stakeholders interact Tweets can contain information about the mood of their authors. [10]. Twitter has led to the development of disease-specific Even when users are not specifically posting about their personal communities that can categorize and aggregate their interactions emotive status, the message can reflect their mood. As such, using “hashtags.” These Twitter communities serve as readily tweets are regarded as microscopic instantiations of emotions. accessible, no-cost platforms that provide significant educational Twitter has been extensively analyzed for health-related and professional benefits. conditions. Nevertheless, to the best of our knowledge, no study Within stroke medicine, social media, specifically, Twitter has has been conducted in chronic stroke, with a focus on the been recently highlighted for its potential to benefit patients, emotional aspects and topics addressed by stroke survivors. stroke organizations, and medical education [11]. Stroke in Young Adults The stroke-related Twitter network has been recently studied Stroke is the third leading cause of long-term disability and one [12], through 621,653 tweets containing the #Stroke hashtag of the leading causes of depression worldwide [1]. Evidence from March 20, 2012, to January 31, 2018, in relation to tweet suggests that stroke incidence in young adults is increasing in content, activity metrics, engagement, and user characteristics. high-income countries [2]. It has been recently reported that The most commonly discussed topics were prevention, diabetes, ischemic stroke is no longer a disease affecting just elderly atrial fibrillation, aphasia, dementia, thrombectomy devices, people, and an estimated 3.6 million young people (age<55 thrombolysis, and tobacco. Specifically, the content of years) are affected each year [3]. The burden of stroke in young discussions included recognition of the signs of a stroke, people may be increasing further, since multiple recent studies associated risk factors (eg, atrial fibrillation, heart disease, and have reported increasing incidence of ischemic strokes, diabetes), and findings of peer-reviewed journals regarding particularly at younger ages, while the incidence at older ages stroke treatment. Tweets were mainly composed by has been declining during the same period [4]. advocacy/support organizations (21.5%), physicians (8.4%), individuals not known to be directly working in the health care Globally, almost half of the entire stroke burden is on young industry (14.0%), other health care professionals (5.5%), individuals, as they have a greater likelihood to survive strokes, organizations related to research/academia (2.3%), and with long life spans ahead, and because strokes occur at younger academics (2.2%), while stroke patients contributed to 6.7% of ages in low- and middle-income countries [5]. Moreover, the tweets (n=41,822). There was a similar proportion of total tweets overall population burden of cerebrovascular disease in young with the #Stroke hashtag generated by physicians (8.4%) and people may be underestimated, since clinically silent infarcts patients (6.7%) during the study period and apparent minimal and white-matter changes are prevalent even in young stroke network communication between physicians and patients, as patients [6]. reported in the study conclusions [12]. About one-fourth of ischemic strokes occur in working-aged Emotional Distress in the Adjustment Process for individuals in high-income countries, with the incidence Stroke Survivors increasing worldwide in this age group from the 1980s to present [3]. Brennan emphasizes the importance of assumptions in adapting to the world around us. According to Brennan’s model, we each Gender Differences in Stroke Outcomes have a cognitive map or representation of the world, resulting After experiencing a stroke, women experience more activity from our social and cultural context and the accumulation of limitations, worse health-related quality of life (HRQoL), and our life experience. This highly complex “assumptive world” more poststroke depression than men, as recently reported in is biologically adaptive in that it allows us to anticipate and plan an updated systematic review of sex differences [7]. for the future [13]. Recent research published in January 2019 in the European In the case of a typical stroke patient, their assumptive world Journal of Neurology reported that women are twice as likely will almost always be challenged or disconfirmed by the to suffer from severe depression following a stroke than men. experience of stroke and its immediate repercussions [14]. As Ayis and colleagues [8] followed the progression of symptoms Brennan states, “adjusting core assumptions involves huge over 5 years after stroke onset in 2313 people (1275 men and amounts of cognitive processing and emotional distress, this 1038 women) from the South London Stroke Register and found often leads to acute emotional difficulties, such as feelings of that 20% of women suffered from severe depression compared confusion, loss, sadness and anger.” to 10% of men [8]. Moreover, the experience of stroke and disability may also The higher prevalence of depression among women may reflect confirm previously held negative beliefs for some individuals a higher prevalence in the general population, where depression (eg, “I am worthless” or “Others see me as weak”) and may was identified as the leading cause of disease burden in women lead to emotional distress in this manner [14]. worldwide [9]. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 2 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Spontaneous, Emotional Language, and Everyday are coded for Plutchik’s basic human emotions [23], and has Topic Discussions on Twitter been applied in several sentiment analysis studies [24]. Over the last few years, Twitter has become a notable data Plutchik’s categories also have the advantage of providing a source in sociolinguistics, as it captures opinions and sentiments balanced list of positive (trust, joy, anger, and anticipation) and on a wide range of topics. Although Twitter users are a negative (disgust, sadness, fear, and surprise) emotions, which, self-selected group, it has been argued that analyses of Twitter to the best of our knowledge, have not been applied in chronic data produce results congruent with those obtained using conditions, in general, or stroke, in particular. standard research methods and data sources [15]. Hedonometer Considering the frequent use of emotional language in tweets After performing emotional analysis based on Putchik’s model, that relate to everyday experiences [16], for a large proportion we propose another point of view, by assigning happiness scores of the population, Twitter provides unobtrusive access to to tweets with the hedonometer tool. The hedonometer [25] was time-sensitive and ecologically valid samples of spontaneously developed from Twitter, Google Books, music lyrics, and the expressed emotions [17]. New York Times for measuring expressed happiness—positive Sentiment analysis in the health care setting is not a new and negative sentiment—in large-scale text corpora. Since its phenomenon, for example, in previous research, greater positive development, the hedonometer has been applied to studies on sentiment within discharge summaries was associated with a predictive markers of depression on Instagram [26] or the significantly decreased risk of readmission [18]. climate change sentiment on Twitter [27]. The hedonometer calculates a happiness score based on the happiness of the This Study individual words used in the text. A total of 10,222 of the most In the following subsections, we describe the specific frequently used English words in four disparate corpora were characteristics and objectives of our study. given happiness ratings using Amazon’s Mechanical Turk online marketplace. Twitter Lists Previous studies have shown that topical experts are often the Adding Covariate Information With Structural Topic primary drivers of interesting discussions on Twitter [19]. In Models contrast to random sampling for gathering Twitter data, Although Latent Dirichlet Allocation (LDA) is, perhaps, the alternative sampling methods have been put forward; one of most common form of topic modeling, a number of associated them proposed to retrieve content only from topical experts, techniques now exist, including dynamic topic models, that is, Twitter users whose followers consider them to be correlated topic models, and hierarchical topic models. One of knowledgeable on some topic, to reduce the number of unwanted the most increasingly popular techniques to emerge in recent tweets in the sampled data while still gathering useful tweets years, however, is structural topic modeling (STM). STM related to a specific topic. The key challenge, however, lies in provides a flexible way to incorporate “metadata” associated identifying a good set of experts [20]. with the text, such as when the text was written, where (eg, Twitter users can organize the accounts that they follow into which country) it was written, who wrote it, and characteristics Twitter user lists. These lists are used in a variety of ways. In of the author, into the analysis using document-level covariates. some cases, they may correspond to personal lists of a given In turn, it allows analysis of relationships between metadata user’s friends and families, but frequently, lists are employed and topics in the text corpus. to group together Twitter accounts based on a common topic Study Objectives or theme. In this way, every Twitter user can effectively become As Brennan states [13], the adjusting process involves huge a community curator. Therefore, previous research has proposed amounts of emotional distress. This often leads to acute that we consider a Twitter user a “topical expert” if the user emotional difficulties such as feelings of confusion, loss, belongs to several lists on a particular topic [20]. sadness, and anger. Considering that women experience more In our study, we propose to take advantage of user lists in the activity limitations, worse health-related quality of life, and field of stroke. To the best of our knowledge, lists have not been more poststroke depression, in this study, we propose to take used in studies related to chronic health conditions. advantage of unobtrusive access to samples of spontaneously expressed emotions and opinions provided by Twitter and to Plutchik’s Human Emotions analyze them from a gender perspective using two different, Currently, there is no single accepted psychological theory of well-established approaches (Plutchik model and the basic human emotions; nevertheless, there is an agreement that hedonometer tool), with the following specific aims: a simple positive-negative dichotomy is not enough to capture • To compare tweets by gender of stroke survivor for the 8 the full range of emotions [21]. basic emotions (anger, fear, anticipation, surprise, joy, In this work, we use the Plutchik [22] approach, which sadness, trust, and disgust) while considering the raw postulates the following eight basic human emotions: joy, frequencies of words in the collection of tweets posted by sadness, anger, fear, trust, disgust, anticipation, and surprise. a stroke survivors’ sample. There have been extensive applications of this approach, for • To determine the proportion of each emotion in the example, the National Research Council (NRC) Word-Emotion collection of tweets and statistically compare each of them Association Lexicon, which contains 10,170 lexical items that http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 3 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al by gender. This measurement thus allows us to track the Data Cleaning proportion of each emotion for each individual tweet and The final sample was prepared for analysis by using the is less affected by single outliers. quanteda R package. This included the process of basic • To extract the main topics (represented as sets of words) normalization (eg, remove punctuation and lowercase all text), that occur in the collection of tweets, related to each gender. stop word removal (eg, the words “a” and “the”), normalization • To assign happiness scores to tweets and topics (using the of Twitter user mentions (eg, “@janedoe” is converted to hedonometer) and compare them by gender. “@user”), lemmatization (eg, “dog,” “dogs,” and “dog’s” are all converted to “dog”), and nonprintable character removal (eg, Methods emojis). All analyses relied on public, anonymized data; adhered to the terms and conditions, terms of use, and privacy policies Data Collection of Twitter; and were performed under Institutional Review We considered the network analysis from previous research Board approval from the authors’ institution. [12] (see #Stroke in the Introduction) as the starting point. Node size is related to user influence, which is directly correlated to We do not report any specific tweets that could be used to the amount a user is mentioned. The top identified nodes and identify the original Twitter user who posted the content online, their corresponding number of followers are as follows: as this is an important concern that has been discussed @TheStrokeAssoc (102 million), @signagnststroke (68 million), extensively in recent literature on the ethics of using Twitter @StrokeHope (93.8 million), @PeterCoghlan1 (7.2 million), data for research [29]. @strokefdn (11.5 million), @StrokeAssocNW (5.5 million), Sentiment Analysis @StrokeAHA_ASA (10.6 million), and @HeartandStroke (45.3 We calculated the overall frequencies of emotion words for million), @HeartandStroke (45.3 million). each Plutchik category for each user (and therefore gender) by Twitter data collection was performed using the rtweet R using the syuzhet R package [30]. The NRC Word-Emotion package [28] via Twitter's REST (representational state transfer) Association Lexicon is available via open access and has been and stream application program interfaces (APIs). We initially implemented in the get_nrc_sentiment() function of the syuzhet applied the lists_users function to obtain all lists that the top R package. Finally, the data were subjected to statistical nodes subscribe to, including their own. Subsequently, we used analyses: For each tweet, given an emotion X, an emotion the lists_members function to obtain Twitter list members (users proportion score was calculated as: on a given list). To retain a list member, we imposed the proportionX = frequency of words with emotion X in condition that it should appear in at least 3 different lists. a tweet / (frequency of negative words in a tweet + For each identified user, we retrieved tweets with the frequency of positive words in a tweet) (equation 1) get_timelines() function (it retrieves the most recent 3200 tweets The emotion proportion scores for men and women were then for each Twitter user, without any time restriction). We collected subjected to a Wilcoxon rank sum test in R, since the F-test had 1,300,845 tweets from a thousand users and further classified revealed that the two distributions did not meet the criterion of them in tweets from particulars (woman or man) and from variance homogeneity [31]. organizations (institutions and associations); the last step was to collect only tweets from particulars where the gender could Structural Topic Models be clearly determined, as explained in the next section. Considering the final sample of tweets from the data cleaning Participant Selection Process phase presented above as the starting point, we proceeded as follows: We modeled our data collection methods on prior studies that have used the Twitter platform for generating a convenience 1. Convert cleaned tweets to tm corpus and create a term sample of users with publicly available accounts, who document matrix (TDM) using the tm R Package [32]. self-identify as stroke survivors in their profile or tweets. 2. Calculate the term frequency inverse document frequency (TF-IDF) for all the words in TDM. We then confirmed the self-reported stroke diagnosis by having 3. Exclude all the words with TF-IDF≤0.1 to remove all the one researcher generate this initial list of Twitter users and a words that are less frequent. second researcher check the details for each Twitter user on the 4. Calculate the optimal number of topics (K) in the corpus list to ensure correct identification of stroke survivors users. using the log-likelihood method for the calculated TDM We then employed a stepwise process for coding each Twitter using Gibbs sampling and exploring different metrics: user’s gender as male, female, or unknown/insufficient data. “Griffiths2004,” “CaoJuan2009,” “Arun2010,” and Two researchers independently used these codes, beginning “Deveaud2014” using the FindTopicsNumber function from with each Twitter user’s username, followed by profile name, the ldatuning R package [33]. profile description, profile photo, and tweets. Both researchers 5. Apply the spectral method using the stm package to discover then reviewed their final gender codes for each Twitter user to topics. ensure consistency and resolve disagreements. 6. Topic validation (semantic coherence and exclusivity). 7. Visualization and interpretation of results from the calculated model. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 4 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al A unique feature of STM, implemented by the stm R package [34], is that it can model how the document level covariates Results affect the topical prevalence parameter μ with a generalized Sample Description linear model. As mentioned in the Sentiment Analysis section above, our covariate is the gender factor with two levels After the selection process, a final sample of 479 Twitter users (“Woman” and “Man”). who posted 800,424 tweets between August 1, 2007, and December 1, 2018, were selected. Women (n=244) posted a Besides the inclusion of the gender covariate, the stm R package total of 3,788,069 tweets; from them, we collected 396,898 supports the explicit estimation of correlations among topics. tweets (the most recent ones, up to December 2018), and the This feature provides further information on the corpus structure. mean number of tweets posted by our selected sample was 1620. Correlations are estimated by replacing the Dirichlet distribution In addition, 54% of the selected sample posted more than 1000 in the standard LDA framework with a logistic normal tweets and 71% posted more than 500 tweets. The total number distribution as in the Correlated Topic Model [35]. of followers of the selected sample was 182,807. This allows us to identify when two topics are likely to cooccur Men (n=235) posted a total of 1,469,364 tweets, from which within tweets (here, we focus on both positive and negative we collected 403,526 tweets (the most recent ones, up to correlations, which are also useful to identify gender December 2018), and the mean number of tweets posted by our differences). selected sample was 1717. In addition, 59% of the selected Hedonometer sample posted more than 1000 tweets and 73% posted more than 500 tweets. The total number of followers of the selected We applied the hedonometer tool to all tweets and to the main sample was 255,053. identified topics as follows: For each word in each tweet, we obtained a happiness score, calculated the mean happiness score Figure 1 shows the date of the first and last posted tweets for for each day, and plotted it by date grouping by gender; STM each selected participant included in our sample (women in red, allows us to identify the main topics and label the topics as men in blue; same code colors throughout the analysis). Each “More likely Women” and “More likely Men.” As each topic vertical line in the plot represents a participant whose first tweet is defined by a set of words, we obtained the happiness score was posted at the top of the vertical line and last tweet was of each word using the hedonometer, and therefore, we are able posted at the bottom of it. We ordered participants in the plot to compare topics according to their happiness score. This also from left to right, where the earliest date of the first tweet is allows us to select, for example, the top 25 words with the shown leftmost for each participant. For example, the leftmost highest levels of happiness and identify if such words belong participant is a man whose first tweet was posted in 2007 and to female or male topics. last tweet was posted in 2014. Figure 1. Topics and gender covariate obtained with spectral structural topic modeling. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 5 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al In Multimedia Appendix 1, we present the number of tweets emotions [23] and implemented in the syuzhet R package, posted by year; a larger number of tweets was posted in 2018 associates an emotion (or more than one emotion) to each of (about 300,000) and the other 500,000 posts were distributed the 10,170 lexical items. Given a word and emotion X, the NRC with growing tendency since 2007, as presented in previous Word-Emotion Association Lexicon associates a score (range: research (described in the #Stroke section). 0 to 1) with it. A score of 1 indicates that the word conveys the highest amount of emotion X. A score of 0 indicates that the For each of the 479 participants, we reviewed their profiles to word conveys the lowest amount of emotion X. verify their geographic locations, obtained by means of the rtweet library. We were able to identify the geographic locations We then identified (via the get_nrc_sentiment() function) the of 378 of the 479 users (78.91%). number of words that, according to the NRC, express positive or negative sentiment as well as one (or more than one) of In Multimedia Appendix 1, we present the total number of users Plutchik’s eight basic emotions. (N=378) by country, showing that most of the users are from four countries: 95% are from Australia, Canada, the United Table 1 summarizes the raw number of words (and their Kingdom, or the United States. percentages) obtained with the get_nrc_sentiment() function of the syuzhet R package. United States had the most users (206/378; 55%). The United Kingdom had 113 users (29.89%). As such, both countries Among both men and women, the most frequent emotions were together accounted for more than 85% of the participants. trust, anticipation, and joy (top 3), as shown in Figure 2. In Multimedia Appendix 1, we present wordclouds of the top Women used considerably more words from all positive 500 words in all participants’ profile description. Most words categories (except anger), and men used more words in all are repeated in both wordclouds, but some distinctive negative categories (except surprise), as shown in Table 1. characteristics can be observed (women clearly refer to Music, When considering negative or positive words, women used 12% Live, and Time, while men do not). of negative words, while men used 13.6% of negative words. Sentiment Analysis In contrast, women used 21.8% of positive words, while men The NRC Word-Emotion Association Lexicon, which contains used 20.5% of positive words. Positive and negative labels for 10,170 lexical items that are coded for Plutchik’s basic human words are also obtained from the NRC lexicon using the get_nrc_sentiment() function of the syuzhet R package. Table 1. Raw frequencies of words identified for each emotion. Emotion Men, n (%) Women, n (%) Anger 76,650 (5.7) 74,858 (5.4) Anticipation 155,608 (11.6) 166,150 (12.0) Disgust 55,512 (4.1) 54,785 (4.0) Fear 117,221 (8.7) 104,826 (7.6) Joy 131,243 (9.8) 161,933 (11.7) Sadness 101,475 (7.6) 89,868 (6.5) Surprise 77,109 (5.7) 83,663 (6.1) Trust 170,0176 (12.7) 178,718 (12.9) Negative 182,288 (13.6) 166,000 (12.0) Positive 276,124 (20.5) 300,751 (21.8) a Not applicable because . b N/A: not applicable. Figure 2. Ranking of emotions (in percentage of the total words. Men: left; women: right). Each bar represents the percentage of total words presented in Table 1. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 6 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al We then calculated the emotion proportion score for each negative words than women; in addition, each individual positive emotion X, as shown in equation 1 in the Methods section. emotion (joy, anticipation, and trust, except surprise) was favorable to women and each individual negative emotion (fear, Table 2 reports statistical comparisons; for example, for the sadness, and disgust) was preferred by men. global positive emotion, women (median=100%, mean=65.57%) used considerably more positive words than men Global negative-positive proportion comparisons are presented (median=66.67%, mean=60.73%). Since the F-test indicated in Figure 3. Women used considerably fewer negative words that the two distributions have a significantly different variance and more positive words than men (shown at the top and bottom (F237040,242190=1.0468, P<.001), they were subjected to a of Figure 3, respectively) Wilcoxon rank sum test. This test showed that the difference Plutchik’s eight emotions are subdivided into four between men and women is highly statistically significant complementary pairs, namely, joy–sadness, (W=2.6817e+10, P<.001). Similar results are shown in Table anticipation–surprise, trust–disgust, and anger–fear [23]. 2 for global negative emotion: Men used considerably more Table 2. Statistical comparison of words identified for each emotion. Emotion, participants Median Mean F (df) P value W P value Joy 0.9030 (237040,242190) <.001 2.63e+10 <.001 Men 0 0.2972 — — — — Women 0 0.3611 — — — — Negative 1.0468 (237040,242190) <.001 3.05e+10 <.001 Men 0.3333 0.3927 — — — — Women 0 0.3443 — — — — Fear 1.1183 (237040,242190) <.001 3.01e+10 <.001 Men 0 0.2481 — — — — Women 0 0.2121 — — — — Positive 1.0468 (237040,242190) <.001 2.68e+10 <.001 Men 0.6667 0.6073 — — — — Women 1.0000 0.6557 — — — — Sadness 1.0261 (237040,242190) <.001 2.45e+10 <.001 Men 0 0.2204 — — — — Women 0 0.1868 — — — — Anger 1.0399 (237040,242190) <.001 2.90e+10 <.001 Men 0 0.1573 — — — — Women 0 0.1495 — — — — Anticipation 0.9837 (237040,242190) <.001 2.79e+10 <.001 Men 0 0.3299 — — — — Women 0 0.3488 — — — — Surprise 1 (237040,242190) >.99 N/Ab N/A Men 0 0.1672 — — — — Women 0 0.1779 — — — — Trust 1.0134 (237040,242190) <.001 2.81e+10 <.001 Men 0.1667 0.3628 — — — — Women 0.2500 0.3755 — — — — Disgust 1.0559 (237040,242190) <.001 2.88e+10 <.001 Men 0 0.1153 — — — — Women 0 0.1109 — — — — http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 7 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 3. Topics and gender covariate obtained with spectral structural topic modeling. Figure 4 plots such scores for each emotion summarized monthly scores for men in almost every emotions throughout the along all the time periods in the study of emotion words for considered period. each pair of emotions, obtained with syuzhet R package and As shown in Figure 4, joy and global positive words clearly plotted with the ggplot2 R package. It clearly shows higher present higher values for women throughout the considered scores for women in positive emotions along time and lower period. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 8 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 4. Monthly mean scores for NRC emotions in the 2008-2018 period. As shown in Figure 5, happiness ratings obtained by Assigning Happiness Scores With the Hedonometer hedonometer summarized on a daily basis for each user are also We then calculated the happiness score using the hedonometer higher for women than for men, almost throughout the for each word in each tweet, summarized the mean happiness considered time period, with remarkable differences in favor score for each day during the whole period under study, and of women, for example, in the 2013-2014 period, 2016, and plotted it by date, grouping by gender. 2018. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 9 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 5. Topics and gender covariate obtained with spectral structural topic modeling. and we applied the hedonometer to each of them; the happiness Structured Topic Modelling scores are shown in brackets for each word. Before the application of STM, we performed an exploratory cluster analysis using VosViewer [36]. As defined by We summarized the happiness scores and obtained mean VosViewer, a term map is a two-dimensional representation, happiness scores of 5.31 (SD 1.31) for all the words present in which strongly related terms are located close to each other only in men clusters and 6.25 (SD 1.07) for those present only and less strongly related terms are located further away from in women clusters. We then considered each of the largest each other. Each point in a term map has a color that depends clusters separately and obtained the following for men: mean on the density of items at that point. It is argued that the VOS happiness score in cluster 1=5.52 (SD 0.99) and mean happiness mapping technique yields more satisfactory term maps than score in cluster 2=5.06 (SD 1.25). For women, the mean popular multidimensional scaling–based approaches to happiness score was 6.30 (SD 1.12) in cluster 1, 5.70 (SD 1.13) bibliometric mapping. Maps constructed using these in cluster 2, and 6.68 (SD 0.81) in cluster 3. Again, happiness multidimensional scaling–based approaches are shown to suffer scores of women were higher than those of men when from certain artifacts. Maps constructed using the VOS mapping considering the scores at the cluster level. technique do not have this problem, as reported by Waltman et Before STM, we also performed LDA analysis for seven topics; al [36]. Details are presented in Multimedia Appendix 1 the number of topics was determined as shown in Multimedia (VosViewer Cluster Analysis). Appendix 1 by using different metrics and the We tested different parameter configurations to increase FindTopicsNumber function from the ldatuning R package. intercluster distances and reduce intracluster distances. The obtained topics are presented in Multimedia Appendix 1, VosViewer allowed us to identify seven clusters for men and but as with cluster analysis, we could not identify topics clearly five clusters for women (Multimedia Appendix 1). In the related to men or women. obtained clusters for the most relevant 250 words for men and women, we highlighted words that are common to clusters Therefore, we applied STM to associate covariates (Gender) to obtained by men and women. Unfortunately, this is the case for the identified topics and plot results as presented in Figure 6. most of the words; therefore, it did not allow us to visually As with most topic models, the objective function maximized identify gender differences. by STM is multimodal. Therefore, the way we choose the starting values for the variational EM algorithm can affect our Nevertheless, in Multimedia Appendix 1, we present the clusters final solution. for the words that are not common to both men and women, http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 10 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 6. Topics and gender covariate obtained with spectral structural topic modeling. We applied LDA initialization (the default option), which uses We then assigned happiness scores to topics with the several passes of collapsed Gibbs sampling to initialize the hedonometer tool (we proceeded as was done with VosViewer algorithm. clusters described in Multimedia Appendix 1). As shown in Figure 6, STM allows us to label the topics as “More likely Table 3 shows the top three topics selected for men and women, Women” and “More likely Men.” with 3 different types of word profiles: highest probability, frequency-exclusivity (FREX), and score values. Detailed As presented in Table 3, each topic is defined as a list of 20 descriptions of all identified topics are presented in Multimedia words. To assign happiness scores, we selected the words with Appendix 1. the highest probability (first row of each topic in Table 3). Therefore, we applied the hedonometer tool to assign a FREX measures exclusivity of the words to the topic in a way happiness score to each word with the highest probability of that balances it with word frequency. The score metric divides each topic. the log frequency of the word in the topic by the log frequency of the word in other topics, and highest probability considers In the Multimedia Appendix 1, we present the complete list of words within each topic with the highest probability, inferred all words defining each topic. For each word, we present the directly from topic-word distribution. happiness score and its corresponding topic (and therefore the associated gender to the topic). We selected a subset of these In Multimedia Appendix 1, we present the evaluation of the words (Tables 4 and 5). We show the 25 words with the highest obtained topics. Semantic coherence is a criterion developed happiness scores and those with the lowest happiness scores by Mimno and colleagues [37]; it is maximized when the most along with the corresponding gender. probable words in a given topic frequently cooccur together, and it has been shown that the semantic coherence metric Table 4 shows that 19 of the 25 words with the highest happiness correlates well with human judgment of topic quality [37]. scores correspond to women’s topics and only 6 correspond to men’s topics. Table 5 shows the 25 words with the lowest Topics 19, 17, and 10 show all semantic coherence values to happiness scores, and only 7 corresponded to women, while 18 the right side of the plot, with topic 19 to the rightmost possible corresponded to men. position and close to topic 10. Topics 1, 6, and 3 are in the central positions, while upper right side positions are six optimal selected topics showing acceptable values. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 11 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Table 3. Top three identified topics and percentages for women (topics 1, 3,6) and men (topics 10,17,19). Topic (%) Highest probability FREXa Score 1 (5.23) year, happy, tomorrow, open, birthday, take, merri, birthday, appl, ang, eve, awak, con, happi, year, birthday, tomorrow, open, come, busy, christmas, baby, sleep, friday, clay, decemb, angel, happi, est, syracus, christma, sleep, friday, come, busi, babi, sunday, monday, list, bed, smile, market, ako, relax, closet, lang, store, carousel sunday, take, bed, store, monday, holiday, treat, guess list, date, market 3 (7.21) good, great, video, hope, night, morn, lol, playlist, chicken, grill, peter, egg, chees, good, video, hope, night, morn, great, lol, tonight, long, done, head, weekend, fun, movi, delici, potato, cooki, bbq, soup, recip, weekend, playlist, movi, tonight, don, luck, readi, celebrate, citi, movie, luck, earli, for- video, kitti, cup, chili, luck, pan, belli fun, long, sweet, forget, gonna, saturday, get dinner 6 (5.73) time, life, thing, world, god, famili, twitter, god, lord, pray, faith, amen, bless, prayer, time, god, thing, life, famili, twitter, world, hear, power, hate, pass, speak, human, step, psalm, soul, negat, holi, heal, thank, charl, lord, bless, hear, power, step, super, hate, posit, bless, super, continu, messag, creat nchousingbuild, merci, evil, accomplish, pray, prayer, congrat, pop, faith, posit yea, compass 10 (3.82) end, heart, walk, news, stop, run, hand, pay, mile, rate, bpm, attitud, anthem, bioness, heart, walk, end, news, mile, stop, run, rate, mile, rate, worth, success, dead, offer, singl, hawk, failur, shoulder, flaw, casual, com- bpm, hand, pay, dead, success, attitud, hero, reach, staff, fail, snow, hero plic, tattoo, zombi, hero, pinterest, hand, snow, worth, bbc, offer, reach virus, vancouv 17 (5.25) back, play, game, team, job, place, boy, yard, hole, playoff, player, joe, nfl, eagl, game, team, play, back, boy, job, footbal, man, point, won, lost, black, park, perfect, cowboy, bronx, kiss, dalla, theater, doodl, player, park, perfect, black, act, place, north, act, lose, john, footbal, film, player lewi, cunt, throw, golden, barn, korea, brave beat, test, film, lose, tour, kick 19 (6.25) stroke, support, find, survivor, learn, lot, aware, raise, foundat, risk, research, patient, stroke, survivor, learn, lot, find, support, brain, care, health, patient, help, aware, region, medic, lot, recoveri, donat, increas, brain, patient, awar, care, rais, health, re- money, raise, children, research, import, factor, studi, resourc, rehab, treatment, search, risk, foundat, injuri, studi, region, risk, experience, hospital cancer, rehabilit, learn recoveri, disease a FREX: frequency-exclusivity. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 12 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Table 4. Top 25 words with highest happiness scores, topics, and gender of participants. Word Participant Score Topic Love Women 8.42 T16 Happy Women 8.3 T1 Win Women 8.12 T15 Smile Women 8.1 T1 Won Men 8.1 T17 Music Women 8.02 T2 Weekend Women 8.0 T3 Celebrate Women 7.98 T3 Christmas Women 7.96 T1 Fun Women 7.96 T3 Free Men 7.96 T11 Great Women 7.88 T3 Success Men 7.86 T10 Award Women 7.86 T15 Positive Women 7.8 T6 Hero Men 7.8 T10 Sun Men 7.8 T11 Birthday Women 7.78 T1 Winner Women 7.78 T15 Beauty Men 7.76 T5 Family Women 7.72 T6 Gift Women 7.72 T15 Brilliant Women 7.68 T2 Super Women 7.68 T6 Amazing Women 7.66 T16 http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 13 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Table 5. Top 25 words with the lowest happiness scores, topics, and gender of participants. Word Participant Score Topic Death Men 1.54 T18 Kill Women 1.56 T12 Die Men 1.74 T18 Fail Men 1.96 T10 Dead Men 2.0 T10 Pain Men 2.1 T4 Hell Men 2.22 T9 Poor Men 2.32 T9 Hate Women 2.34 T6 Sad Women 2.38 T12 Attack Men 2.42 T8 Shot Women 2.5 T2 Shit Men 2.5 T18 Aphasia Men 2.58 T11 Stroke Men 2.58 T19 Lie Men 2.6 T13 Bad Women 2.64 T16 Fight Women 2.7 T16 Lost Men 2.76 T17 Lose Men 2.76 T17 Disabled Men 2.82 T18 Problem Men 2.98 T4 Wrong Men 3.14 T18 Forget Women 3.22 T3 Cut Men 3.42 T9 We then calculated the boxplots of the happiness scores for each in a tweet. In Figure 9, we plot both positive and negative topic (Figure 7), ordered from “More Likely Men” to “More correlations for all identified topics. Likely Women”; the means and regression line are shown in Topic 1 shows the highest positive correlation with Topic 3. red circles and a red line, respectively (P<.001). This can be further confirmed in Table 3, as both topics address Figure 7 shows higher happiness scores from Topic 4 to the actual positive everyday life situations like celebrations right (ie, women’s topics) with the exception of Topic 16, which (birthday, Christmas, holiday, merry), and Topic 1 was strongly contains several words with low happiness scores (eg, “bad” or negatively correlated with Topic 19, which refers to research, “lone”; Multimedia Appendix 1). The regression line shows a studies, risks, factors, hospital, disease, stroke, and care. positive slope in the direction of women’s topics (P<.001). Topic 3, therefore, is also strongly negatively correlated with We then compared happiness scores by pairs from the leftmost Topic 19 and Topic 10. and rightmost topics in Figure 6 to the center (Topic 19-Topic Topic 10 refers to running, beats per minute, heart rate, attitude, 1, Topic 17-Topic 6, Topic 10-Topic 3, etc). We found stop, walk, and reach, while Topic 3 refers to fun, celebrate, significant differences in favor of women in 4 of the 10 pairs movie, Saturday, dinner, barbeque, chicken, grill, egg, cheese, of topics (and none in favor of men) when comparing the delicious, potato, and cooking. Topic 6 addresses religion—god, happiness scores by pairs of topics (Figure 8; men blue, women lord, pray, faith, amen, bless, prayer, psalm, soul—while Topic red). The complete list of comparisons is presented in 17 addresses sports—playoff, nfl, game, yard, football, Multimedia Appendix 1. player—showing clear differences in topics of interest addressed STM also permits correlations between topics. Positive by men and women. correlations indicate that both topics are likely to be discussed http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 14 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 7. Topics and gender covariate obtained with spectral structural topic modeling. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 15 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Figure 8. Topics and gender covariate obtained with spectral structural topic modeling. Figure 9. Topics and gender covariate obtained with spectral structural topic modeling. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 16 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al positive social interaction and positive feedback from other Discussion SNS users (according to the results of an ethnographic diary Principal Findings study on Facebook use from Sas et al [39]). In this work, we proposed the application of Twitter lists to a Second, prior research indicates that the positivity of chronic health condition in a GNU framework (R-3.5.1). We self-presentation on SNSs has an influence on both the quantity applied a set of publicly available R libraries for collecting and and quality of reactions from other SNS contacts. For example, organizing Twitter data via Twitter’s REST and stream API Utz [40] found that SNS users were least likely to receive (rtweet), sentiment analysis (syuzhet), text mining (tm, reactions from their online friends when they expressed sadness quanteda), and structural topic model (stm). We also applied in their postings. Similarly, Forest and Wood [41] demonstrated the hedonometer tool to assign happiness scores to topics by that more positive status updates on Facebook received more gender. positive and favorable feedback from friends than negative status updates. According to our findings, men use significantly more words while expressing negative emotions in their tweets than women, A third explanation to our findings can be related to the while women use significantly more words when expressing existence of “the positivity bias in SNS communication,” which positive emotions. states that “while the SNS environment generally enables authentic self-presentation, it favors positive forms of The results also show that the top three most frequent emotions authenticity over the presentation of negative aspects of the true evoked by both men and women are trust, anticipation, and joy. self” [42]. Besides, the statistical analysis of the basic emotions detects significant preferences for each gender: While words from the Therefore, according to Reinecke et al, due to the positivity bias emotional fields of trust, anticipation, and joy appear in SNS communication, individuals with higher levels of significantly more often in women’s tweets, men’s tweets psychological well-being have a higher chance of experiencing significantly exhibit a preference for evoking disgust, anger, authenticity through the use of SNSs than SNS users with low fear, and sadness. psychological well-being. We also applied another tool that was specifically designed for The fourth explanation is related to a recent Facebook analysis considering Twitter inputs—the hedonometer. Happiness ratings involving 15,000 users [43]. The authors concluded that obtained by the hedonometer, summarized on a daily basis for “language used more by self-identified females was each user, are also higher for women than for men almost all interpersonally warmer, more compassionate, polite, along the considered time period. and—contrary to previous findings—slightly more assertive in their language use, whereas language used more by Finally, we applied structural topic modelling (to the best of self-identified males was colder, more hostile, and impersonal.” our knowledge, for the first time to a chronic health condition) In fact, the following text from their publication, can also be to identify main topics addressed by gender and determined applied to our own findings: positive and negative correlations between topics by gender. The most strongly female-linked topics included words Topics in this context are defined as sets of words; therefore, describing positive emotions (e.g., “excited”, we assigned happiness scores to the words with highest “happy”, “<3”, “love”,), social relationships (e.g., probabilities in the identified topics and found that the topics “friends”, “family”, “sister”), and intensive adverbs women talk about show higher happiness scores than the topics (e.g., “sooo”, “sooooo”, “ridiculously”). Strongly addressed by men. male-linked topics included words related to politics A common stereotype in both Western and Eastern cultures (e.g., “government”, “tax”, “political”), sports and suggests that women are more emotional than men, particularly competition (e.g.,“football”, “season”, “win”, when responding to negative emotions [38]. As remarked in the “battle”), and specific interests or activities, such as Introduction section, after stroke, women experience more shooting guns, playing musical instruments, or activity limitations, worse health-related quality of life, and playing video games. more poststroke depression than men [7] and are twice as likely Therefore, according to this fourth explanation, our findings in to suffer from severe depression following a stroke than men. another SNS such as Twitter are similar to those involving users We identify several explanations for our findings, listed below. not necessarily identified as stroke survivors on Facebook. First, according to Ayis et al [8], women draw larger Limitations components of their sense of self and self-worth from The collected sample was not intended to be representative or interpersonal relationships and networks, and they are more a comprehensive set of all tweets posted by stroke survivors sensitive to adversities of these. Therefore, female stroke during the period under study. Although the collected data also survivors may experience (to a larger extent in comparison to included tweets directed at other users (ie, conversational men) the interpersonal and intrapersonal benefits of sharing tweets), the results cannot be considered to reflect all topics of positive events and emotions on social network sites (SNSs). conversation appearing in Twitter for stroke survivors. The intrapersonal benefits of sharing positive events and emotions on SNSs consist of re-experiencing and prolonging Data collection relied on Twitter’s streaming API, which these positive events; the interpersonal benefits comprise prevents collection of tweets from private Twitter accounts. As http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 17 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al a result, findings may not represent individuals with private the situational context in which those traits were expressed is accounts. unclear. Possible users affected by severe depression may not be active on Twitter; this could be a source for another Furthermore, recent analysis [44] shows that 62% of all Twitter significant bias in the data sample. users are less than 49 years old; our participants are skewed toward such an age range, and most of them from the United Comparison with Prior Work States. One of the scarce previous research about tweet topics or Nevertheless, as discussed in the Introduction section, multiple sentiment analysis on chronic health conditions was recently recent studies have reported a sustained increasing incidence conducted by Brunner and colleagues [48]. Tweets tagged with of stroke at younger ages and the included participants were traumatic brain injury (TBI)-related hashtags were harvested randomly selected after checking their membership to several over a one-month period in 2016 and analyzed qualitatively and Twitter stroke–related lists and manually double checked in quantitatively. A total of 29,199 tweets included tweets sent by relation to gender and stroke survivor condition. 893 users, 219 of whom had a brain injury. Twitter was used to discuss health issues, raise awareness of TBI, talk about life We analyzed women (n=244) who posted a total of 3,788,069 after TBI, talk about sport and concussion, and communicate tweets. From them, we included 396,898 tweets in our analysis inspirational messages. (the most recent ones, up to December 2018); therefore, we analyzed 10.5% of all posted tweets by women participating in In relation to depression, Lachmar and colleagues [49] captured this study. 3225 original tweets for the hashtag #MyDepressionLooksLike that circulated in May 2016. Cleaning resulted in a total of 1978 We analyzed men (n=235) who posted a total of 1,469,364 tweets. Using qualitative content analysis revealed seven themes: tweets. From them, we included 403,526 tweets in our analysis dysfunctional thoughts, lifestyle challenges, social struggles, (the most recent ones, up to December 2018). Therefore, we hiding behind a mask, apathy and sadness, suicidal thoughts analyzed 27.4% of all posted tweets by men participating in and behaviors, and seeking relief. Contrary to Lachmar and this study. colleagues' [49] analysis or the #Stroke analysis (the one The total number of tweets posted by women from whom we presented in the Introduction section), our analysis is not linked extracted our sample is clearly larger than tweets posted by to a specific hashtag. men. This seems to be coincidental with general Twitter use It is important to remark the need for further research from a statistics: Women are usually more active, and each month, 40 gender perspective, as promoted by initiatives such as the million more women than men visit Twitter [45]. Women’s Brain Project [50]. Other relevant factors to be mentioned as limitations to our Conclusions study are related to geographic location, spatial trajectory, or the time of the day a tweet has been posted. As remarked by This study explored emotional expressivity for eight specific Padilla et al [46] and Gore et al [47], such factors may affect types of emotion and identified 20 main topics of interest tweets’ sentiments. We observed that 85% of our participants through Twitter posts in stroke survivors from a gender profiles are from the United Kingdom and United States, but perspective. Numerous studies have shown that, compared with spatiotemporal aspects are not controlled in our study. men, women usually experience more frequent and stronger negative emotions. Nevertheless, our results show that men Finally, the individual psychological differences that stroke present more frequent and stronger negative emotions in their survivors may experience must also be mentioned. Certain tweets, when considering both globally positive-negative or individuals might have personality traits that make them more individual tweets and analyzing them using two different predisposed to positive or negative sentiments. The degree to well-established approaches: the Plutchik model and the which sentiment reflects variance in psychological traits versus hedonometer tool. Acknowledgments This research was partially funded by EU H2020 PRECISE4Q - Personalized Medicine by Predictive Modeling in Stroke for better Quality of Life (Grant Agreement 777107 – Research and Innovation Action). Authors' Contributions AG-R and SL conceived the study. AG-R, JS, and SL collected, selected, and cleaned the data. AG-R and JS analyzed the data. AG-R drafted the initial manuscript. SL, JS, and MBG revised the manuscript critically for important intellectual content and approved the final manuscript. AG-R, SL, JS, and MBG received funding for the study. Conflicts of Interest None declared. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 18 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Multimedia Appendix 1 Demographics, wordclouds, VosViewer cluster analysis, latent Dirichlet allocation topics, correlation analysis, STM topics, hedonometer scores, and the Plutchik psychoevolutionary model. [PDF File (Adobe PDF File), 2 MB-Multimedia Appendix 1] References 1. Johnson W, Onuma O, Owolabi MS. Bulletin of the World Health Organization. 2016. Stroke: a global response is needed URL: https://www.who.int/bulletin/volumes/94/9/16-181636.pdf [accessed 2019-07-22] 2. Béjot Y, Daubail B, Jacquin A, Durier J, Osseby G, Rouaud O, et al. Trends in the incidence of ischaemic stroke in young adults between 1985 and 2011: the Dijon Stroke Registry. J Neurol Neurosurg Psychiatry 2014 May;85(5):509-513. [doi: 10.1136/jnnp-2013-306203] [Medline: 24249786] 3. Putaala J. Ischemic stroke in the young: Current perspectives on incidence, risk factors, and cardiovascular prognosis. Eur Stroke J 2016 Mar 08;1(1):28-40 [FREE Full text] [doi: 10.1177/2396987316629860] [Medline: 31008265] 4. Vangen-Lønne AM, Wilsgaard T, Johnsen SH, Carlsson M, Mathiesen EB. Time trends in incidence and case fatality of ischemic stroke: the tromsø study 1977-2010. Stroke 2015 May 07;46(5):1173-1179. [doi: 10.1161/STROKEAHA.114.008387] [Medline: 25851772] 5. Krishnamurthi RV, Moran AE, Feigin VL, Barker-Collo S, Norrving B, Mensah GA, et al. Stroke Prevalence, Mortality and Disability-Adjusted Life Years in Adults Aged 20-64 Years in 1990-2013: Data from the Global Burden of Disease 2013 Study. Neuroepidemiology 2015 Oct 28;45(3):190-202. [doi: 10.1159/000441098] 6. Putaala J, Kurkinen M, Tarvos V, Salonen O, Kaste M, Tatlisumak T. Silent brain infarcts and leukoaraiosis in young adults with first-ever ischemic stroke. Neurology 2009 May 26;72(21):1823-1829. [doi: 10.1212/WNL.0b013e3181a711df] 7. Gall S, Phan H, Madsen TE, Reeves M, Rist P, Jimenez M, et al. Focused Update of Sex Differences in Patient Reported Outcome Measures After Stroke. Stroke 2018 Mar;49(3):531-535. [doi: 10.1161/STROKEAHA.117.018417] 8. Ayis SA, Rudd A, Ayerbe L, Wolfe CDA. Sex differences in trajectories of depression symptoms and associations with 10-year mortality in stroke patients: The South London Stroke Register. London: The South London Stroke Register; 2019. 9. Mathers C. The global burden of disease update. In: The Global Burden of Disease: 2004 Update. Geneve: World Health Organization; 2008:2004. 10. Radu G, Solomon M, Gheorghe CM, Hostiuc M, Bulescu IA, Purcarea VL. The adaptation of health care marketing to the digital era. J Med Life 2017;10(1):44-46 [FREE Full text] [Medline: 28255375] 11. Cabrera-Maqueda JM, Minhas JS. New Horizons for Stroke Medicine: Understanding the Value of Social Media. Stroke 2018 Feb;49(2):25-27. [doi: 10.1161/STROKEAHA.117.020068] 12. Bundy JJ, Chick JFB, Hage AN, Srinivasa RN, Chaudhary N, Srinivasa RN, et al. #Stroke. J NeuroIntervent Surg 2018 Apr 20;10(12):e33-e33. [doi: 10.1136/neurintsurg-2018-013877] 13. Brennan J. Adjustment to cancer - coping or personal transition? Psychooncology 2001;10(1):1-18. [Medline: 11180573] 14. Taylor GH, Todman J, Broomfield NM. Post-stroke emotional adjustment: A modified social cognitive transition model. Neuropsychological Rehabilitation 2011 Dec;21(6):808-824. [doi: 10.1080/09602011.2011.598403] 15. Sylwester K, Purver M. Twitter Language Use Reflects Psychological Differences between Democrats and Republicans. PLoS One 2015 Sep;10(9):e0137422 [FREE Full text] [doi: 10.1371/journal.pone.0137422] [Medline: 26375581] 16. Kramer A. The spread of emotion via Facebook. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. USA: ACM; 2012 Presented at: Presented atIGCHI Conference on Human Factors in Computing Systems; 2012; Austin, Texas p. 05-10. 17. Kern ML, Park G, Eichstaedt JC, Schwartz HA, Sap M, Smith LK, et al. Gaining Insights From Social Media Language: Methodologies and Challenges. Psychol Methods 2016 Aug 8:507-525. [doi: 10.1037/met0000091] [Medline: 27505683] 18. Gohil S, Vuik S, Darzi A. Sentiment Analysis of Health Care Tweets: Review of the Methods Used. JMIR Public Health Surveill 2018 Apr 23;4(2):e43 [FREE Full text] [doi: 10.2196/publichealth.5789] [Medline: 29685871] 19. Ardon S. Spatio-Temporal Analysis of Topic Popularity in Twitter. arXiv 2011 Apr 05:2904-2908. 20. Ghosh S, Zafar M, Bhattacharya P, Sharma N, Ganguly N. On sampling the wisdom of crowds: random vs. expert sampling of the twitter stream. 2013. URL: http://ama.liglab.fr/wikiDeAMA/DOCS/randomvsexpert.pdf [accessed 2019-08-06] 21. Hoffmann T. “Too many Americans are trapped in fear, violence and poverty”: a psychology-informed sentiment analysis of campaign speeches from the 2016 US Presidential Election. Linguistics Vanguard 2018;4(1):1-9. [doi: 10.1515/lingvan-2017-0008] 22. Plutchik R. The Nature of Emotions: Human emotions have deep evolutionary roots, a fact that may explain their complexity and provide tools for clinical practice. American Scientist 2001:344-350. 23. Mohammad S, Turney P. Emotions evoked by common words and phrases: Using Mechanical Turk to create an emotion lexicon. 2010 Presented at: NAACL-HLT Workshop on computational approaches to analysis and generation of emotion in text; 2010; California, US. 24. Schweinberger M. A sociolinguistic analysis of emotives in Irish English. 2016 Presented at: Annual Meeting of the Society for Text & Discourse; 2016; Kassel. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 19 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al 25. Dodds PS, Harris KD, Kloumann IM, Bliss CA, Danforth CM. Temporal patterns of happiness and information in a global social network: hedonometrics and Twitter. PLoS One 2011 Dec;6(12):e26752 [FREE Full text] [doi: 10.1371/journal.pone.0026752] [Medline: 22163266] 26. Reece A, Danforth C. Instagram photos reveal predictive markers of depression. EPJ Data Sci 2017 Aug 8;6(1):15. [doi: 10.1140/epjds/s13688-017-0110-z] 27. Cody E, Reagan A, Mitchell L, Dodds P, Danforth CM. Climate Change Sentiment on Twitter: An Unsolicited Public Opinion Poll. PLoS One 2015;10(8):e0136092 [FREE Full text] [doi: 10.1371/journal.pone.0136092] [Medline: 26291877] 28. Kearney M. Package ‘rtweet’. URL: https://cran.r-project.org/web/packages/rtweet/rtweet.pdf [accessed 2019-03-19] 29. Webb H, Jirotka M, Stahl B, Housley W, Edwards A, Williams M, et al. The ethical challenges of publishing Twitter data for research dissemination. 2017 Presented at: ACM on Web Science Conference; 2017; Troy, NY p. 25-28. 30. Jockers M. Introduction to the Syuzhet package. URL: https://cran.r-project.org/web/packages/syuzhet/vignettes/ syuzhet-vignette.html [accessed 2019-08-06] 31. Gries S. Statistics for linguistics with R: A practical introduction. NY: Mouton de Gruyter; 2009. 32. Feinerer I. Introduction to the tm Package: Text Mining in R. URL: https://cran.r-project.org/web/packages/tm/vignettes/ tm.pdf [accessed 2019-03-19] 33. Murzintcev N. ldatuning: Tuning of the Latent Dirichlet Allocation Models Parameters. URL: https://cran.r-project.org/ web/packages/ldatuning/index.html [accessed 2019-03-19] 34. Roberts M. Journal of Statistical Software. stm: R Package for Structural Topic Models URL: https://cran.r-project.org/ web/packages/stm/vignettes/stmVignette.pdf [accessed 2019-03-19] 35. Blei DM, Lafferty JD. A correlated topic model of Science. Ann Appl Stat 2007 Jun;1(1):17-35. [doi: 10.1214/07-AOAS114] 36. Waltman L, van Eck NJ, Noyons EC. A unified approach to mapping and clustering of bibliometric networks. Journal of Informetrics 2010 Oct;4(4):629-635. [doi: 10.1016/j.joi.2010.07.002] 37. Mimno D, Wallach H, Talley E, Leenders M, McCallum A. Optimizing semantic coherence in topic models. In: In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ?11, pp. 262?272. Association for Computational Linguistics, Stroudsburg, PA, USA. 2011 Presented at: Conference on Empirical Methods in Natural Language Processing; 2011; Stroudsburg, PA, US. 38. Deng Y, Chang L, Yang M, Huo M, Zhou R. Gender Differences in Emotional Response: Inconsistency between Experience and Expressivity. PLoS ONE 2016 Jun 30;11(6):e0158666. [doi: 10.1371/journal.pone.0158666] 39. Sas C, Dix A, Hart J, Su R. Dramaturgical capitalization of positive emotions: The answer for Facebook success? 2009 Presented at: 23rd British HCI group annual conference on people and computers: Celebrating people and technology; 2009; Swinton, UK. 40. Utz S. Social network site use among Dutch students: Effects of time and platform. In: Networked Sociability and Individualism: Technology for Personal and Professional Relationships. Hershey, PA: IGI Global; 2011:103-125. 41. Forest AL, Wood JV. When Social Networking Is Not Working. Psychol Sci 2012 Feb 07;23(3):295-302. [doi: 10.1177/0956797611429709] 42. Reinecke L, Trepte S. Authenticity and well-being on social network sites: A two-wave longitudinal study on the effects of online authenticity and the positivity bias in SNS communication. Computers in Human Behavior 2014 Jan;30:95-102. [doi: 10.1016/j.chb.2013.07.030] 43. Park G, Yaden DB, Schwartz HA, Kern ML, Eichstaedt JC, Kosinski M, et al. Women are Warmer but No Less Assertive than Men: Gender and Language on Facebook. PLoS One 2016;11(5):e0155885 [FREE Full text] [doi: 10.1371/journal.pone.0155885] [Medline: 27223607] 44. Omnicore. Twitter by the Numbers: Stats, Demographics & Fun Facts URL: https://www.omnicoreagency.com/ twitter-statistics/ [accessed 2019-03-18] 45. Who is more active on Social Media?. URL: https://www.digitalinformationworld.com/2014/10/ who-is-more-active-on-social-media-men-or-women-infographic.html [accessed 2019-03-19] 46. Padilla J, Kavak H, Lynch C, Gore R, Diallo SY. Temporal and spatiotemporal investigation of tourist attraction visit sentiment on Twitter. PLoS One 2018;13(6):e0198857 [FREE Full text] [doi: 10.1371/journal.pone.0198857] [Medline: 29902270] 47. Gore R, Diallo S. You Are What You Tweet: Connecting the Geographic Variation in America?s Obesity Rate to Twitter Content. PLoS ONE 2015;You Are What You Tweet:1-16 [FREE Full text] 48. Brunner M, Hemsley B, Dann S, Togher L, Palmer S. Hashtag #TBI: A content and network data analysis of tweets about Traumatic Brain Injury. Brain Injury 2017 Dec 08;32(1):49-63. [doi: 10.1080/02699052.2017.1403047] 49. Lachmar EM, Wittenborn AK, Bogen KW, McCauley HL. #MyDepressionLooksLike: Examining Public Discourse About Depression on Twitter. JMIR Ment Health 2017 Oct 18;4(4):e43 [FREE Full text] [doi: 10.2196/mental.8141] [Medline: 29046270] 50. Womens Brain Project. URL: http://www.womensbrainproject.com/ [accessed 2019-06-10] http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 20 (page number not for citation purposes) XSL• FO RenderX JOURNAL OF MEDICAL INTERNET RESEARCH Garcia-Rudolph et al Abbreviations API: application program interface EM: expectation maximization FREX: frequency-exclusivity HRQoL: health-related quality of life LDA: latent Dirichlet allocation NRC: National Research Council REST: representational state transfer SNS: social network site STM: structural topic modeling TDM: term document matrix TF-IDF: term frequency – inverse document frequency Edited by G Eysenbach; submitted 23.03.19; peer-reviewed by R Gore, L Subirats, S Kiritchenko; comments to author 18.04.19; revised version received 11.06.19; accepted 16.06.19; published 26.08.19 Please cite as: Garcia-Rudolph A, Laxe S, Saurí J, Bernabeu Guitart M Stroke Survivors on Twitter: Sentiment and Topic Analysis From a Gender Perspective J Med Internet Res 2019;21(8):e14077 URL: http://www.jmir.org/2019/8/e14077/ doi: 10.2196/14077 PMID: 31452514 ©Alejandro Garcia-Rudolph, Sara Laxe, Joan Saurí, Montserrat Bernabeu Guitart. Originally published in the Journal of Medical Internet Research (http://www.jmir.org), 26.08.2019. This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included. http://www.jmir.org/2019/8/e14077/ J Med Internet Res 2019 | vol. 21 | iss. 8 | e14077 | p. 21 (page number not for citation purposes) XSL• FO RenderX