We analyzed 2762 tweets created between December 2018 and September 2021. URL clicks available on Twitter analytics were used as a proxy for user interest. Twitter analytics data for @AntibioticResis were collected from the in-built analytics toolset on the Twitter website. Data fields include the number of impressions (the total number of times a tweet was displayed on timelines), engagements (the total number of interactions with a tweet—likes, retweets, and replies), and the number of retweets, replies, likes, and URL clicks.

To understand the impact of user base characteristics on URL clicks, we examined the most common words that occurred in the user biographies of the @AntibioticResis account followers (retrieved on June 17, 2022). To obtain the common words, we removed hyperlinks, special characters, and stopwords from the biographies written in English (inferred from the Python langdetect package; Python Software Foundation). Single characters were not considered common words. After applying the above cleaning steps, we identified dominant roles (health practitioners and professionals, academics) and topics (AMR, infectious diseases, microbiology, and public health) from a word co-occurrence network. We then curated a list of words for each role or topic and counted the number of followers by role and topic (Tables S1 and S2 in Multimedia Appendix 1).

Follower categories suggest that there are two different types of attention to tweeted scientific articles: (1) academic attention from the followers classified into health practitioners, professionals, and academics and (2) general attention from the followers who do not provide roles. We acknowledge that some followers without role information may actually be health practitioners, professionals, or academics. We assume that this case is uncommon. We termed this academic attention and general attention, respectively. In particular, we focused on 12 WHO priority pathogens (critical: Acinetobacter baumannii, Pseudomonas aeruginosa, Enterobacteriaceae; high: Enterococcus faecium, Staphylococcus aureus, Helicobacter pylori, Campylobacter spp, Salmonella, Neisseria gonorrhoeae; medium: Streptococcus pneumoniae, Haemophilus influenzae, and Shigella spp) and 3 organisms of specific interest: Escherichia coli, Klebsiella pneumoniae, and Enterobacter spp.

Many organisms receive greater attention from the academic community than others, and this has the potential to impact engagement levels on Twitter. To quantify this attention for the selected pathogens, we obtained the total number of papers published per year as indexed on PubMed with a set of broad search terms to capture all possible cases. Search terms consisted of the full pathogen name or bacterial name with the abbreviated genus and a truncated wildcard to encompass terms related to resistance (eg, “Acinetobacter baumannii” OR “A. baumannii”) AND resist* AND Journal Article [Publication Type]). Where a whole genus was named, this was used as the search term along with the truncated wildcard search term (eg, “Enterobacter” AND resist*). Data collection was limited to papers published from 1990 to 2021. Academic attention was calculated from the publication shares for pathogens by year. To clarify the steps of measuring academic attention, we introduced the following mathematical notation: suppose n_i,2021 is the number of publications for pathogen i in the year 2021. The publication shares for i in 2021 p_i,2021 equal to . If a tweet in 2021 contains 2 pathogen names, i and j, the academic attention on this tweet is p_i,2021+p_j,2021. The minimum of academic attention is 0 if no pathogen name is listed in a tweet, and the maximum value in our data is 0.496 for the paper titled “Synthesis and characterization of biogenic selenium nanoparticles with antimicrobial properties made by Staphylococcus aureus, methicillin-resistant Staphylococcus aureus, Escherichia coli, and Pseudomonas aeruginosa.” The distribution of academic attention values can be seen in Figure S1 in Multimedia Appendix 1.

General attention is likely different from academic attention to the selected pathogens. We inferred the general attention on the Twitter platform from the number of tweets mentioning a pathogen name by using the Twitter Academic application programming interface (API). As tweets have a character limit, it was likely that bacterial names would be shortened, so tweets containing either the full bacterial name or bacterial names with an abbreviated genus were included (eg, “Pseudomonas aeruginosa” AND “P. aeruginosa”). Unlike PubMed, the Twitter API does not allow wildcard search, so we use a query (“full name” OR “abbreviated name”) (resist OR resistant OR resistance).

To identify candidate phrases attracting Twitter accounts to click a URL link, each title was converted into a list of the selected pathogens, organisms, and unigrams from the tweets as follows. First, all uppercase characters were converted to lowercase characters, and then stopwords, single-digit numbers, and special characters except hyphens joining 2 words were removed. Second, the WHO priority pathogens were identified (both full names and abbreviations; eg, Acinetobacter baumannii and A. baumannii) and considered independent phrases. Third, the remaining words in a title were stemmed from the Porter stemmer [13].

The URL clicks in our data were overdispersed and heavily right-skewed (Figure S2 in Multimedia Appendix 1). A total of 377 tweets (14%) did not obtain a single URL click, and the maximum number of URL clicks was 66 for the paper titled “The cost of antimicrobial resistance.” The average number and variance of URL clicks were 5.0 and 32.9, respectively. To explain overdispersed URL clicks, we used a negative binomial (NB) regression with identified phrases. The probability of having a count y in NB f(Y=y|λ_i,P)is written as

where λ_i is the mean of the distribution, x_i is the covariates of i, β is a vector for coefficients, and P is the dispersion power parameter, which is set to 1 in this paper. We estimated β for the following covariates to understand the impacts of various factors on URL clicks. The NegativeBinomialP function in the Statsmodels Python package was used for our analyses [14].

First, we constructed binary variables indicating whether a title contains a WHO priority pathogen or a common nonpathogen stem appearing in more than 50 tweets (65 common nonpathogen stems exist in total). x_ik is 1 if a pathogen or a stem k is mentioned in the tweet i, and 0 otherwise. Second, the number of retweets was included as a covariate to consider the size of potential audiences varying by the number of followers of retweeters. Third, the number of phrases in a title was added to the regression models, as long titles may discourage Twitter users from clicking a link. Fourth, we considered academic attention and general attention to the pathogens, as academic attention and general attention may also affect URL clicks in addition to the phrases shown in a title. For each pathogen, we calculated its publication and tweet shares over the selected pathogens for each year in order not to be affected by total publication and tweet counts over time. For each tweet, academic attention and general attention are calculated as the sum of the publication and tweet shares for the pathogens, respectively. Interestingly, academic attention and general attention are highly correlated (the Pearson correlation is 0.93), though they were measured from different sources. Using the covariates above, we built 3 different NB models explaining URL clicks. Model 1 described URL clicks per tweet in the context of the selected phrases, the number of retweets the post received, and the number of individual phrases contained within the title. Each of the following models used all the factors from model 1 with added considerations. Model 2 considered academic attention only, and model 3 considered general attention only.

To generate vector representations of paper titles, a pretrained SPECTER was used [15]. This SPECTER obtained embeddings of scientific articles based on their citations to convert paper titles into vectors. If 2 paper titles appeared close together in the SPECTER embedding space, their topics were likely to be similar. Leveraging these characteristics, for each WHO priority pathogen, the embedding vectors of paper titles containing the pathogen name were obtained.

To assess how engagement changes depending on paper type, tweets with ten or more URL clicks (N=380) were manually categorized according to the type of article included in the tweet. Categories used were review articles, original basic research, clinical, and others (including short communications, comments, errata to previously published works, and books). Paper types were expressed as a percentage of the total number of tweets per number of URL clicks.

No ethics review was sought because the study only explored publicly available data on social media and did not conduct any experiments on human subjects.

We aimed to characterize the followers of the @AntibioticResis Twitter account to better understand the users that most frequently engage with this content on social media. Firstly, we assessed the number of new followers gained per month over the time period from July 2016 to April 2022. Interestingly, it was noted that there was an obvious spike in new followers in November each year from 2016 to 2020 (Figure S3 in Multimedia Appendix 1).

To gain a better understanding of the professional backgrounds of the individuals who follow this Twitter bot, a co-occurrence network of frequent words that appear in more than 50 followers’ biographies was created. A node in the network is a keyword, and a link between 2 nodes is created if these nodes appear together in more than 50 followers’ biographies. Links have no weight, so all links have the same width. The length of a link does not represent any quantity of the network but derives instead from the network layout algorithm. The locations of nodes are determined by the ForceAtlas 2 algorithm in Gephi [16]. The ForceAtlas 2 algorithm [17] positions nodes by making them repulse each other based on their connections.

The co-occurrence network reveals that the primary group of individuals following the @AntibioticResis account are likely health practitioners and professionals or academic researchers (Figure 1). These roles tend to be associated with 4 areas: AMR, infectious diseases, microbiology, and public health. The words representing each group (health care professionals and academics) are shown in Table S1 in Multimedia Appendix 1, and the terms related to each topic are listed in Table S2 in Multimedia Appendix 1. A biography can have words across multiple roles and multiple topics. The numbers of followers by role and topic are provided in Figure 1.

The WHO has identified several pathogens and bacterial families (divided into critical priority, high priority, and medium priority) for which developing new treatments should be an international prerogative [18]. As priority pathogens, these organisms should receive increased attention as funders and governments recognize the need to support work on them, and the public is widely informed about their negative impacts on health.

To explore academic and general attention for the selected pathogens, we calculated relative publication and tweet counts compared to 2 baselines. The baseline for academic attention is the number of journal articles on PubMed with a query “Journal Article[Publication Type],” and the baseline for general attention is the number of English tweets from StoryWrangler, a comprehensive tool providing daily word usage since 2008 [19] (Figure 2; see Figure S4 in Multimedia Appendix 1 for raw publication and tweet counts for the selected pathogens).

The relative publication counts for A. baumannii and K. pneumoniae increased, suggesting rapidly growing academic attention to these pathogens. Other critical and high-priority pathogens and organisms of specific interest maintained certain levels of relative publication counts. Two medium-priority pathogens, S. pneumoniae and H. influenzae, lost academic attention. On the other hand, the relative tweet counts tend to increase for most of the selected pathogens, reflecting an increased general interest. This trend partly reflects increasing awareness about AMR as a result of global and national public health campaigns.

It was noted that some organisms had spikes in general public interest over time. For example, it was shown that the medium-priority organism Shigella had a substantial increase in interest from the general public on Twitter in 2015. This spike was shown to coincide with an increase in searches on Google for that organism (Figure S5 in Multimedia Appendix 1). This suggests that increases in the general interest of a particular organism on the internet and discussions surrounding it on social media may be closely related.

The WHO priority pathogen list is recognized as a key scale of the threat posed by each pathogen. We hypothesized that URL clicks would be correlated to the priority level of the pathogens on this list. To explore this, we used NB regressions. Model 1 estimated URL clicks based on the existence of the selected bacterial name, the number of retweets, and the number of phrases in a title. From this model, 3 critical priority pathogens—A. baumannii, P. aeruginosa, and Enterobacteriaceae—and one high-priority pathogen—S. aureus—and 3 pathogens of specific interest—E. coli, K. pneumoniae (also considered collectively under the term Enterobacteriaceae), and Enterobacter—were found to be positively associated with URL clicks (Table 1). This indicates that Twitter users’ interest in AMR partially aligns with the WHO’s priority pathogen list. Furthermore, several pathogens, including Campylobacter spp, N. gonorrhoeae, and Salmonella spp, showed no association with URL clicks. However, the significant coefficients may originate from the time-varying interests of academic researchers and the general audience. For this reason, we include academic attention (model 2) and general attention (model 3) to model 1. In model 2, academic attention is not significant, while adding it increases the rate ratios for A. baumannii, P. aeruginosa, Enterobacteriaceae, and K. pneumoniae, although with less statistical significance. In model 3, general attention is also not significant. The rate ratios and their significance are close to those in model 1 (see Tables S3, S4, and S5 in Multimedia Appendix 1 for the full regression results). These results suggest that the WHO critical priority pathogens and K. pneumoniae tend to attract URL clicks more than other selected pathogens, even though we consider both academic and general attention.

The mention of specific pathogens was expected to affect URL clicks. However, the impact of nonpathogen-specific words and phrases (stems) on URL clicks was unclear. We, therefore, examined common nonpathogen-related stems from the titles of the papers tweeted by the @AntibioticResis account. In model 1, a number of nonpathogen-related stems were found to have either a significant increase or significant decrease in URL clicks associated with their presence in an article title (see Table S3 in Multimedia Appendix 1 for all estimates). Those that significantly (P<.05) increased URL clicks were “gram-neg” (rate ratio: 1.47, P<.001), “antimicrobi” (1.44, P<.001), “urinari” (1.41, P=.048), “new” (1.35, P=.001), “bacteria” (1.35, P<.001), “use” (1.27, P=.005), “emerg” (1.25, P=.01), “treatment” (1.24, P<.001), “antibiot” (1.24, P<.001), “detect” (1.21, P=.02), “molecular” (1.19, P=.04), “multidrug-resist” (1.18, P=.02), and “resist” (1.13, P=.002). Nonpathogen stems that significantly (P<.05) reduced URL clicks were “activ” (0.81, P=.007), “agent” (0.73, P=.02), and “viru” (0.73, P=.045). The number of retweets was positively associated with URL clicks (1.24, P<.001), implying, as expected, that retweets enhance the visibility of tweets. On the other hand, the number of phrases in a paper was found to be negatively associated with URL clicks (0.93, P<.001). This suggests that papers with longer titles are less likely to encourage Twitter users to click a link.

Many factors affect the ability of an organism to garner greater web-based attention than others. We hypothesized that organisms that were positively correlated with URL clicks would be published under a greater range of topics, therefore attracting a greater number of readers. Using a UMAP (Uniform Manifold Approximation and Projection) 2-dimensional vector mapping [20], we created a visual representation of the diversity of paper topics for the selected organisms, whose names appear in more than 30 tweeted papers (Figure 3) and each organism individually (Figure S6 in Multimedia Appendix 1). The papers about E coli are broadly distributed in the embedding space as much as those about P aeruginosa, but E coli is not associated with either an increase or decrease in URL clicks, while P aeruginosa is at the level of 0.05 in our regressions. This naive comparison suggests that the extent to which papers about a pathogen are distributed across topics is unrelated to URL clicks. More analysis may be needed to determine whether topic diversity is a factor, and this will be the focus of future studies.

It is clear that the mention of particular pathogens has a significant effect on the level of user engagement a tweet will receive. However, to assess the impact of the type of paper on the number of URL clicks, we manually categorized the tweets with the highest number of URL clicks based on their general focus (Figure S7 in Multimedia Appendix 1). Almost half of the assessed tweets (48%) focused on the basic science of antibiotic resistance or resistant organisms. When the number of URL clicks was also considered, the tweets with over 30 URL clicks showed a more even distribution between article types, with reviews comprising 31% of all tweets with high URL clicks, as expected [21]. As URL clicks decreased, fundamental science papers became dominant, comprising 34% and 47% of papers with 20-29 URL clicks and 11-19 URL clicks, respectively.