Insulin pricing and other major diabetes-related concerns in the USA: a study of 46 407 tweets between 2017 and 2019

Data collection

A tweet extraction engine was developed in April 2017 and since then has collected >11.7 million (status July 2019) diabetes-related tweets in English, via Twitter’s Streaming Application Programming Interface (API), based on a list of diabetes-related keywords, such as diabetes, hypoglycemia, hyperglycemia and insulin, from all over the world (see online supplementary materials S1 for the full list of keywords used). All data collected in this study were publicly posted on Twitter. Therefore, according to the privacy policy of Twitter, users agree to have this information available to the general public.13 In this study, we restricted the analysis to 46 407 US-based tweets, with personal and emotional content (see figure 1 for an overview over the workflow).

Request permission

For further details concerning data representation, data preprocessing, training of machine learning classifiers to filter tweets with personal content, from institutional tweets (organisations, advertisement, research news, etc) related to diabetes; detect and exclude jokes; predict the gender and type of diabetes of users as well as the geolocation algorithm to determine tweets in the USA, please refer to online supplementary material S2. For the present work, the statistical unit considered is a tweet and not a user profile. Thus a user may have several tweets that are included.

Identifying emotions

To identify emotions among diabetes-related tweets, we combined detailed Parrot’s classification of emotions with a dictionary of emotional keywords present in the two most common tools to assess psychological health in people with diabetes, namely the PAID questionnaire and the DDS.14 Parrot identified over 100 emotions and conceptualized them as a tree structured list (full list in online supplementary material S3) with the six primary emotions being joy, love, surprise, sadness, anger and fear. To capture as many tweets as possible containing emotions, we hypothesized that the synonyms of the words from Parrot’s classification should be included as well. Synonyms were identified using the WordNet database.15

Moreover, due to the nature of Twitter as a microblogging service (short messages), people use emoticons and emojis that express special meanings. Emoticons constitute a metacommunicative pictorial representation of a facial expression using punctuation marks and letters, such as ‘:-)’ or ‘:D’.16 Emojis are incorporated into sets of characters available in mobile phones, such as ‘

’ or ‘

’. We used Wolny’s categorisation of emojis and emoticons into emotional categories similar to the six primary emotions defined by Parrot.17 This allowed us to determine a subset of emotional tweets containing either emotional words or emojis/emoticons. Emojis reproduced with permission from Twitter under CC-BY 4.0: https://creativecommons.org/licenses/by/4.0/

Sentiment analysis

Sentiment analysis is commonly used to investigate the positive or negative opinion within a tweet. We used a widely accepted human-validated sentiment analysis tool Valence Aware Dictionary for Sentiment Reasoning (VADER).18–21 VADER was specifically developed for social media data and combines a lexicon and the processing of the sentence characteristics to determine sentence polarity. VADER computes sentiment and valence for each word level and provides positive, negative and neutral scores at the sentence level. We used the compound score as our main metric for the sentiment analysis (referred to as the SA score later in the text), which is unidimensional and a normalized measure of sentiment between −1 (most negative) and +1 (most positive).

Topic extraction

All tweets are represented via their word vector representations, with the consequence that words similar in semantics are also similar in the word vector space (see online supplementary material S2 for more details). We applied the unsupervised machine learning algorithm K-means, using cosine similarity as distance measure, to group the emotional tweets that are close to each other based on their word vector representations into topics/clusters.22 In order to define the right number of clusters k, the Silhouette score, measuring how on average each data point is closer to its cluster’s center than to any other cluster, was used as input parameter for the K-means algorithm.23 We obtained the highest score for k=30 clusters. All tweets were then assigned to one of the 30 clusters/topics. Each topic/cluster was then given a label by two authors (AA, GF) according to the 10 most contributing tweets (those closest to the topic center) and the most frequent words (top words) in the cluster.

Assessment of the mean income

We studied the associations between the topics of interest for people with or talking about diabetes and the mean income of the city, based on data from the 2017 American Community Survey.24 The tertiles for the mean household income were calculated: low income—(US$24 609, US$67 224), medium income—(US$67 225, US$86 758), high income—(US$86 759, US$394 259). Each geolocated tweet was associated with the mean income for its corresponding city and then assigned to its respective tertile: low income, medium (med) income, high income. Associations between topics and mean income tertile groups were assessed with χ² tests. For each topic, the p value was calculated using a χ² test between the binary variable if a tweet belongs to the corresponding topic and the tertile categories of the city-level mean income from the tweet’s origin.

Python (V.3.6) and the packages scikit-learn (machine learning algorithms and data preprocessing methods) and gensim (text processing, word representation) were used to implement the methods described above.25^,26 Algorithms related to the present study are open source under the following address: https://github.com/WDDS/Tweet-Diabetes-Classification.

Spatial distribution of diabetes-related tweets

This analysis was based on 167 743 geolocated, diabetes-related tweets from the USA (figure 2). The highest number of tweets were seen in California (N: 18 551) and Texas (N: 14 237), whereas Vermont (N: 197) and Wyoming (N: 131) had a low number of tweets. At the city level, New York City (N: 9663), Los Angeles (N: 5301) and Chicago (N: 4,884) had the highest number of tweets.

Request permission

Topics of interest

Among all US diabetes-related tweets, 46 407 were identified as emotional tweets (28% of tweets) of which 14 485 (31%) were written by men, 20 228 (44%) by women and 11 694 (25%) from unknown sex; 20 285 (44%) were predicted as from people with type 1 diabetes, 4375 (9%) from type 2 diabetes and 21 747 (47%) diabetes type unknown. Table 1 shows the detailed description of the 30 topics of interest about diabetes.

In most topics, an over-representation of women was observed, in particular in topics 10, 23 and 26 related to the importance of affordable insulin and the oral glucose tolerance test (OGTT). In contrast, men more frequently discussed topics around diabetes-related stories (topic 6) and ‘diabetic/insulin shock’ (topic 29). Discussions about ‘advocacy for affordable insulin’ (topic 10) and enjoying the exchange in the diabetes online community (topic 5) were almost exclusively dominated by people with type 1 diabetes. The only topics where people with type 2 diabetes had significantly more tweets than on average, is topic 19 related to the confusion between type 1 and type 2 diabetes or in topic 17 when they tweeted about the anniversary of their diagnosis. The Diabetes Social Media Advocacy (#DSMA) group on Twitter is a community of people exchanging about diabetes-related topics, in which the price of insulin is a central concern.

Primary emotions related to the topics of interest

Table 2 illustrates the sentiment and emotions distribution over the 30 topics.

In online supplementary material S4, the most common emotional words and emojis/emoticons have been listed for each topic. Overall, we found that the most positive topics were related to support and solidarity within the diabetes online community (topic 1) with 72% of emotions related to joy and about inspiring relatives living with diabetes (topic 2) with joy (36%) and love (31%) elements. In contrast, the most negative topics were related to ‘diabetes-related comorbidities’ (topic 30) of which 40% of the emotions were related to sadness, 19% to anger and 22% to fear; ‘diabetic/insulin shock’ (topic 29) with an over-representation of anger (45% of all emotions) and fear (46%) and ‘diabetes distress’ (topic 28) with sadness (24%), anger (27%) and fear (21%) emotions expressed. Topics 2 (‘inspiring relatives living with diabetes’) and 4 (‘diabetes awareness/support/donation’) showed a higher percentage of surprise emotions. The topics in which people with or talking about diabetes exchange, support each other and show solidarity (topics 1, 2, 3, 4, 5) contained joy and love emotions most frequently. Topic 9 was related to beauty products with diabetes-related names and was considered irrelevant for our analyses and disregarded.

Insulin pricing

Insulin pricing is a major topic of interest mentioned in 5 of the 30 topics (18% of tweets). A key topic associated with insulin pricing was related to the frustration of not being able to afford insulin (topic 25), in which many tweets expressed sadness (24%), anger (21%) and fear (20%). Topic 13 regrouped tweets in which people are looking for ‘glucose guardians’, a gender neutral term for a person who is giving money, often to buy insulin, in exchange for some favor or something in return. Topic 10 described the tweets of people who are fighting for affordable insulin, which are represented by the hashtag ‘#insulin4all’. The positive emotions in this group showed mutual support.

Associations between topics of interest and mean income

We observed that topics such as 10 ‘advocacy for affordable insulin’, 22: ‘insulin’ and 25: ‘frustration with insulin prices’ were positively associated with mean household city income such that cities with higher incomes were more likely to post tweets about these topics (p<0.001 for all) (see online supplementary material S5). Similarly, we found that topics 24: ‘misunderstandings of diabetes’ (p<0.001), 19: ‘confusion between type 1 and type 2 diabetes’ (p<0.05), 12: ‘life with type 1 diabetes’ (p<0.01) and 21: ‘glycemic instability’ (p<0.05) were positively associated with mean city income. Likewise, positive associations were found between topics 24 and 19 which are related to language use and the way people talk about type 1 and type 2 diabetes, with cities with higher incomes being more likely to talk about these subjects.

In contrast, we observed that topics 5: ‘diabetes social media advocacy’ (#DSMA) group enjoying online support (p<0.001), 18: ‘day-to-day stories about diabetes’ (p<0.001) and 26: ‘oral glucose tolerance test (OGTT)’ (p<0.001) were negatively associated with mean city income, such that cities with lower income were more likely to post these tweets. Furthermore, cities with medium mean income tweet about topic 15: ‘insulin and insulin pump complications’, were more frequently (p<0.001).

Insulin pricing

We found that insulin pricing was a major concern among tweets shared in the USA (18% of all tweets were related to insulin pricing) and is associated with both positive (joy, love) and negative (sadness, anger, fear) emotions. People frequently shared their frustration with insulin prices, access to insulin and identifying sources of insulin including ‘glucose guardians’ or donations, which represent major obstacles for people with diabetes.32^,33 Positive emotions are present when it comes to solidarity in the fight for affordable insulin in the community. We observed associations of topics addressing insulin pricing to be more frequent in cities with high mean incomes. This does not necessarily indicate that people living in cities with a high mean household income feel more concerned about insulin prices, but rather they probably have a greater ability to tweet around this issue. A large number of tweets geolocated in cities with a high mean household income included the hashtag ‘#insulin4all’, a campaign that unites the diabetes community around the access to treatment for everyone.34

It is known that there are key challenges for a global and fair access to insulin.35 With respect to insulin pricing, we are the first to exhibit and quantify, on a large sample of people with or talking about diabetes, the extent of the crisis in the USA based on social media data. In addition, we have also been able to highlight the different emotions and fears associated with the crisis around insulin pricing.

Strengths and limitations

This study has numerous strengths. First, a major advantage of using social media data is that information is expressed spontaneously, on a large scale, and in real-time, in what can be considered as an open digital space with flat role hierarchy for information sharing and development of online communities. This potentially minimizes biases such as responder bias that you would observe in traditional and observational studies. We evaluated tweets related to diabetes from a large number of people with a large variability in their profiles. The methodologies developed in this study present an innovative way to concentrate on relevant (personal, emotional) geolocated tweets (USA), to identify topics of interest and emotions shared within topics. This approach is able to capture trends in the online diabetes community as well as socioeconomic factors that can be associated with social media data at the ecological level. This new way of capturing data supplements the detection of topics which are less medically oriented.

There are, however, several limitations to consider. First, diabetes-related concerns expressed on Twitter may not be representative of all people with diabetes. However, it has been previously suggested that it can be partially offset by the large variability in the social media profiles, a key strength in digital epidemiology.36 While we did observe large variability in our Twitter profiles, we found an over-representation of people with type 1 diabetes and women in our study when compared with known diabetes epidemiology literature.37 The greater representation of type 1 diabetes may be explained by the younger demographics of Twitter users.38 Alternatively, type 1 diabetes may have more involved care, more devices, more challenging medication and more frustrations to report on Twitter as compared with type 2 diabetes. Regardless, our results should be interpreted in the context of the Twitter population only. Second, the precision of our filter classifiers (personal content, jokes), gender and type of diabetes classifiers, is not perfectly accurate, which means that we cannot guarantee that 100% of tweets are posted by actual people with diabetes and it was often impossible to define the sex or type of diabetes. Third, we were unable to account for several clinical and environmental factors that may help to tease out these associations. The label provided by the researchers for each topic is not exclusive. By refining the tweets in each topic, more subtopics could be defined. This could be a future direction to investigate. Fourth, the geolocation of tweets was partially based on locations the users provided, which might not be their true location. Fifth, emotion detection is still a challenge in the machine learning field due to the occurrence of sarcasm and irony. It is one of the open research questions. Last, causal inference between the mean household income per city and the topics of interest of people residing in the corresponding city cannot be made as it is subject to ecological fallacy.

A perspective of our work is to extend our analyses to include more countries and languages.