Our study approach was to undertake content analysis of Twitter data by applying topic modeling. Content analysis is a widely used technique for qualitative research [24] that enables studying patient experience in depth by deriving topics of interest from text documents [14,25].

Twitter was used as the data source rather than other social media platforms, blog posts, or news articles because individuals use this platform for expressing and sharing their feelings and opinions on health-related topics by posting short messages that can be easily collected through application programming interfaces (APIs) or other open sources [14-17,26]. We used an open-source Twitter scraping tool called Twint [27] for collecting tweets related to LBP that were written in English. Twint enables the collecting of Twitter data without using Twitter's API through its publicly available library in the Python programming language [27,28]. We collected tweets posted between January 1, 2014, and December 31, 2018 (inclusive). The time frame of 5 years was selected to provide us with sufficient data to examine the patterns in emerging topics and the number of tweets over time. Since the number of active users on the social media platform increased in recent years and we needed a large volume of data for topic modeling, we did not consider tweets posted before 2014. We selected the search keywords based on 3 studies on back pain [15,29,30]. These are detailed in Table 1. Search keywords were verified by our domain experts (FC, a rheumatologist; DU, a physiotherapist) who have extensive research and clinical expertise in the area of LBP. Selecting search keywords and an appropriate time frame are important considerations in the data collection process. The Monash University Human Research Ethics Committee approved this study (project ID 19738).

Our data processing and analysis consisted of 4 steps (see Figure 1).

We removed duplicates, retweets, URLs, and tweets related to marketing and advertisements, which reduced the data set from 7,892,210 to 2,825,645. We filtered the data further by removing tweets that did not contain first person pronouns [15]. As a result, the remaining data set size was 2,010,295.

We replaced contractions with their expanded forms (eg, “didn’t” to “did not”). We converted the HTML characters to ASCII characters and removed hashtags, Unicode strings (eg, “\u2026”), numbers, and punctuation. We replaced abbreviations, elongated words (eg, “gooood” to “good”), and emoticons and emojis with their equivalent English expressions. We then performed spelling correction, lowercasing, tokenization, and lemmatization, created n-grams, removed stop words (eg, common terms such as “the” and “is”). We again removed the duplicates and the remaining data set was 1,249,576 tweets.

After completing the abovementioned steps, we excluded tweets with less than three words because in topic modeling, the document size is important to achieve high accuracy [31]. This reduced the data set to 896,867 tweets.

Topic modeling is a technique used to provide a summary of a large collection of documents by extracting “topics” that represent the dominant themes [32]. It allows the uncovering of common, hidden themes from a corpus of text documents such as tweets. We tested 5 well-established topic modeling algorithms for detecting topics in a text-based corpus, namely latent Dirichlet allocation (LDA) [33], Dirichlet multinomial mixture (DMM) [34], GPU-DMM [35], biterm topic model (BTM) [36], and nonnegative matrix factorization (NMF) [37].

LDA is a generative probabilistic model that assumes each document can be represented by distribution over topics and each topic by distribution over words [33,38]. DMM is also a generative model but it assumes that each document is associated with one single topic [34,39]. GPU-DMM is an extended method of DMM that considers semantic similarity between words to provide semantic understanding of text documents and improve topic inference [35,40]. BTM uncovers topics by modeling the word co-occurrence patterns (ie, biterms) rather than using the document-level word co-occurrences [36,41]. NMF is able to learn the latent features in data using a nonnegative representation and improve latent semantic topic identification [37,42,43].

To use these models (except for NMF), we used a Java-based open-source library for short text topic modeling algorithms called STTM (version 1.8) [44], whereas for NMF we used the sklearn [45] library. For each approach, we performed a series of experiments ranging from 5 topics to 200 topics. We applied the 5 algorithms to the 896,867 tweets to determine the best model and the optimal number of topics.

Choosing the right number of topics is a crucial step in topic modeling because it can affect the accuracy of results. The quantitative approach computes the coherence score and perplexity, which helps in determining the optimal number of topics [46]. The coherence score measures the sum of the pairwise word-similarity scores of the words in the topic, using the pointwise mutual information (PMI) score [47]. Best collocation pairs usually have a high PMI. On the other hand, the qualitative approach requires humans and domain experts to examine the topics. Human judgment is extremely important because topic modeling uses a form of unsupervised learning.

As a quantitative approach, we calculated the coherence score of each model on different numbers of topics ranging from 5 to 200, based on the PMI score [47,48]. The coherence score was used to evaluate the quality of the topic-word distribution. LDA outperformed other approaches (ie, DMM, GPU-DMM, BTM, and NMF).

Additionally, we used a qualitative approach to select the most representative topics. We manually examined the topics, their top 20 terms, and a random sample of tweets in each topic. We also created a word cloud for each topic and evaluated word clouds and their sample tweets. We identified the number of topics that provided us with distinct and meaningful topics; if we exceeded this number of topics, we started to notice an increase in duplicates and overlapping topics. We used both quantitative and qualitative approaches to select the optimal number of topics.

Topic labeling is a process of representing the meaning of a topic by assigning each topic a descriptive word or phrase [49]. Although automatic labeling approaches can reduce costs and time required, they are not able to achieve high semantic validity and accuracy [50,51]. In our study, we used the “eyeballing” method, which refers to reading and inspecting the top words in a topic and manually assigning a label [50]. We made sure that the results met the requirements of a “good” label: (1) semantically relevant, (2) meaningful, (3) representative, (4) adequate, and (5) understandable [34,49].

LDA assumes that each document (tweet) is a mix of topics with different proportions [33]. We were interested to examine tweets based on their dominant topic to gain a better understanding of the frequency of topics across all tweets. Therefore, we performed further analysis, and used the label of the dominant topic to represent each tweet, and then calculated the total number of tweets per topic.

To improve the results of thematic analysis, low-order topics can be grouped under broad, higher-order categories [52]. Higher-level categories can provide a better overview of the key topics discussed by individuals. To this end, after manual topic labeling, we performed topic categorization and assigned a category label to the topics that represented common themes. To identify the important and widely discussed categories, we then calculated the percentage of all tweets that corresponded to each individual category.

Two domain experts (FC, a rheumatologist; DU, a physiotherapist), actively working clinically and researchers in the area of LBP, independently examined the selected topics from the previous step where each topic included the top 20 words to determine face validity. As previously described, in topic modeling, the top words of each topic provide the description of that topic, thereby assisting the domain experts with inferring its meaning [49]. The experts then met to reconcile any differences and develop the final labels.

The total number of collected tweets about LBP was 7,892,210 from 2,420,258 unique users from 2014 to 2018. The average number of words in each tweet increased from 2017 onward (Multimedia Appendix 1), in line with Twitter doubling the character limit of tweets from 140 to 280 characters as of November 2017 [53].

After performing comprehensive data preprocessing, the final number of retained tweets was 896,867, which represents 11% (896,867/7,892,210) of the original raw data we collected, with a vocabulary size of 29,539. The minimum length of tweets was 4 words and the maximum length was 20 words.

After testing 5 topic modeling algorithms and the number of topics based on the coherence score and our manual examination, we selected the best model that included 60 topics, detected from 896,867 self-reported tweets about LBP. Multimedia Appendix 2 shows the coherence score of different models with a different number of topics ranging from 5 to 200. The best model was the LDA model with 60 topics, which had a coherence score of 0.562. Multimedia Appendix 3 shows the best model selected with 60 topics and their top 20 terms.

The 60 topics were examined and manually given a topic label. The common and duplicate labels were then grouped into higher-order categories. Word clouds for the two categories of “pain regions” and “sleep” after combining the related topics are provided in Multimedia Appendix 4. The prevalence of the 60 manually labeled topics is presented in Multimedia Appendix 5.

Independent examination of selected topics by two domain experts and reconciliation of any differences resulted in 19 contextual categories, with details presented in Multimedia Appendix 6. The total number of tweets within each of 19 contextual categories is presented in Figure 2, with more details in Multimedia Appendix 7. The “emotion and beliefs” category had the largest proportion of the total tweets, followed by “physical activity” and “daily life.” The lowest proportion of tweets belonged to the categories of “food and drink,” “weather,” and “not being understood.”

The proportion of tweets for each higher-level category over the years showed that all 19 categories had been discussed by individuals with relatively similar frequency every year (see Figure 3). However, the proportion of “emotion and beliefs” decreased from 2014 to 2018. The number of tweets about other categories, such as “aggravating factors” and “symptoms,” increased over that time period. An example of a tweet for each category is presented in Table 2 to illustrate the type of personal point of view related to each category.

In this study, we identified 60 specific topics from 896,867 tweets about LBP and grouped them into 19 categories that relate to contextual variables of LBP. The top category was “emotion and beliefs,” with 157,563/896,867 tweets (17.6%), followed by “physical activity” (124,251/896,867, 13.85%) and “daily life” (80,730/896,867, 9%), while “food and drink,” “weather,” and “not being understood” had the lowest proportions of tweets (11,551/896,867, 1.29%; 10,109/896,867, 1.13%; and 9180/896,867, 1.02%, respectively). There were 11 topics within the category of “emotion and beliefs”; of 157,563 tweets in this category, 113,562 (72%) expressed negative sentiment. Our results were consistent with the general findings from traditional study methods in the area of LBP but provided more in-depth detail on the context of LBP from the individual perspective.

Our study examined contextual variables to provide a novel insight into first-person perspectives of the LBP experience and confirmed the broad areas that have previously been identified using more traditional data collection methods from qualitative and quantitative studies. For example, psychosocial factors have an important role in LBP [54] and, from our analysis of tweets, “emotion and beliefs” was the most common topic we identified, with 157,563 of 896,867 tweets. This is consistent with LBP being widely recognized as a biopsychosocial condition, and growing evidence to show that psychological factors, such as beliefs and emotions, play an important role [55]. For instance, systematic reviews have highlighted that beliefs about back pain and negative consequences resulting from these beliefs are common across different countries and populations [56], and affect both treatment efficacy and prognosis [57]. Moreover, mass media campaigns that target negative beliefs have been implemented in an effort to influence how people manage their back pain on a population level [58]. Our study has also provided novel findings with respect to emotions. Although we found a range of emotions, from positive emotions (such as happy, love, or fun) to negative emotions (including hell, bad, or disgusting), the majority were found to be of negative affect. Although several studies have examined the role of specific emotions, such as anger [59,60], in LBP research, our understanding of the array of emotions experienced by individuals with back pain, specifically negative emotions, is limited.

Our study also highlighted areas related to the pain experience in individuals that have not been adequately explored in the literature but that play an important role in the effectiveness of LBP interventions and self-management behaviors, such as the “not being understood,” “religion,” and “food and drink” categories. We found that although the category of “not being understood” had the smallest proportion of tweets with a total of 9180 tweets, it had the top five words: “make,” “people,” ”stop,” “thing,” and “complain.” This is consistent with a previous systematic scoping review that examined what patients want from their medical care, which reported that patients felt misunderstood and wanted legitimation of their LBP [11]. Patients with LBP report negative social stereotyping from health care professionals, family and friends, and the community [61] and that they are dissatisfied with the inadequate advice they receive from medical practitioners and have identified an unmet need for care providers that show more understanding and empathy [11].

The category of “food and drink” is novel and interesting. The tweets included words relating to the type of food (eg, pizza, chocolate, cookies and cream), mealtimes (such as breakfast and lunch), and the process of bringing or making food. Although they reflect important daily habits of eating and drinking, they may also highlight issues around pain affecting an individual’s capacity to eat and drink and/or problems associated with weight and in particular obesity [62], which is a major public health issue [63].

There are well-described sex differences in the prevalence of back pain [64]. Analysis of tweets identified 3 topics under the “women” category including “motherhood,” “large breasts problem,” and “female health complaints.” LBP has been reported in more than two-thirds of pregnant women [65]. Improving psychological well-being, physical fitness, and general well-being may reduce LBP in women [65-67]. The topics identified in tweets may provide more direction in relation to the personal topics that warrant further examination (eg, the potential effect of “large breasts problem” and whether this is a cause of LBP or a potential confounding variable). Identifying possible mechanisms for the association with topics such as “motherhood” or “female health complaints” could also help with understanding whether these associations are due to psychosocial factors or biomechanical factors such as the lifting and carrying of children. Understanding the context of LBP could offer valuable insights into how people with LBP view and experience their condition; this could lead to the identification of new areas of research in exploring the causes of LBP, as well as the opportunity to identify areas of potential misinformation that need to be addressed.

There are some limitations to our study. Although the keywords were taken from existing studies about LBP and approved by domain experts, some keywords, such as “back hurt” and “back pain,” were very broad. Therefore, the data collected might not have been specific to LBP. Selection of the right keywords in Twitter data analysis is very important to avoid unrelated data that could reduce the accuracy of results. Filtering and cleaning of Twitter data is also crucial for achieving high accuracy of results. In our study, we performed vigorous data cleaning, but our manual examination showed that there was a group of tweets that contained a few lines from the lyrics of a famous hip-hop song (Bad and Boujee) by Migos. These lines included “…So my money makin' my back ache.” One of our search keywords was “back ache.” Although there are many tools and methods available to automatically perform data cleaning, it is always necessary to manually inspect the results.

Twitter users tend to be younger and might not represent the general population; therefore, the results must be carefully interpreted [68]. Similar to other social media studies in health care, we cannot verify that individuals who tweeted about LBP were actually real patients [15]. However, the filtering based on first-person pronouns (eg, I, my, or mine) that we performed is likely to have reduced this.

To determine the optimal number of topics, we used the coherence score, a widely used method, and then manually examined and compared the models. This process can be further improved by using other measures such as heuristic approaches [69] or perplexity measures [70].

We also recognize that manual labeling of topics can be subjective. Two domain experts with extensive knowledge were involved in the labeling and examination of selected topics but future work in this area could involve a greater number of and more diverse domain experts to further reduce this subjectivity.

Our findings provided useful insights into individuals’ beliefs and perspectives regarding their needs and concerns related to LBP that complement the information available in the literature. Considering the contextual factors identified in this study rather than simply focusing on a biomedical model of LBP could address the needs of patients more holistically, help with improving LBP outcomes, and increase patient satisfaction. These findings have the potential to assist health care providers and clinicians with developing more effective, personalized therapies for LBP. There is also the potential to use social media to identify any major changes in community beliefs and needs regarding LBP that can be addressed in a timelier manner.