Does X Mark the Spot? Investigating discussions about cancer screening programs on X/Twitter through corpus analysis (Preprint)

BACKGROUND While cancer screening is proven to be effective in the early detection of the disease and early detection enables better treatment options, screening uptake has been declining. Research shows that online health information helps people to make health-related decisions. However, not all online health information is credible, and misinformation might play a role in people’s choice to take part in screening. OBJECTIVE The overall goal of this study was to analyze online discussions about cancer screening programs using corpus analysis. Specifically, we aimed to investigate the full dataset through corpus analysis and misinformation in a manually coded subset. This enabled us to study naturalistic discussions about cancer screening over time, what information people share, and how prevalent misinformation is in these discussions. We differentiated posts on Twitter/X for cervical, breast, colorectal, and general screening. METHODS We extracted a corpus of 55 403 tweets from 2011 until 2023 tweeted by 22 493 users from a database containing over 5.9 billion tweets. We used specific search strings corresponding to the different types of screening and screening in general to gather our corpus. The corpus consisted of tweets, timestamps, hashtags, and shared URLs. We used a machine learning classifier trained on another dataset of tweets about cancer screening to automatically code whether a tweet fell in the scope of the study. We manually coded a randomly drawn stratified subset of 1200 tweets representative of the full corpus regarding year and screening program for the presence of misinformation. RESULTS We found that tweets were not uniformly distributed across different screening programs and over time. Most tweets discussed population screening in general, and the volume of tweets increased around real-world events. The hashtags in the tweets predominantly focused on the screening programs that were discussed in those tweets. In our corpus, most shared URLs linked to other tweets (10 569) or news websites (2807). In our manually coded subset, information was shared in 679 tweets. Twenty-three of those tweets contained misinformation. Topics in those tweets showed criticism towards the programs and policies, suspicions about conflicts of interest and anti-vaccination beliefs regarding HPV. Most users used rhetorical questions, sarcasm, fear mongering or expressed anger. CONCLUSIONS Our findings reveal that cancer screening programs are actively debated across social media platforms. We observed that conversations tend to spike in response to real-world events, suggesting social media can serve as a valuable lens into public reactions to health policy changes. Link-sharing behavior was common, though we noted a tendency for sources to reference back to the same platform where discussions originated. Despite finding limited instances of misinformation in our sample, we caution that even modest amounts of inaccurate information may have meaningful consequences for public health messaging and screening uptake.