Leveraging large language models to identify engagement-driving features in vaping-related TikTok videos: a cross-sectional study (Preprint)

• DOI: 10.2196/preprints.76265
• Publication Date: 2025-04-21T04:50:07Z
• Authors: Zidian Xie; Nanda Kishore Kor...; Amisha Dubey; Luchuan Song; Chenliang Xu; Karen M. Wilson; AnaPaula Cupertino; Dongmei Li
• Keywords: Preprint; Cross-sectional study

<sec> <title>BACKGROUND</title> Electronic cigarette (e-cigarette) use is prevalent in youth and young adults. TikTok, a popular social media platform for adults, has been used to disseminate e-cigarette-related videos, primarily dominated by promotional videos. </sec> <title>OBJECTIVE</title> We aim identify key TikTok video features associated with high user engagement assist future design vaping prevention campaigns. <title>METHODS</title> collected 1,487 videos related metadata using the API (Application Programming Interface). applied large language models GPT-4 Video-LLaMA extract (e.g., promotion content, background, gender, lifestyle, talking, cartoon, tricks, containing emoji) from randomly selected hand-coded 25 check accuracy of two identifying these features. utilized generalized linear identity link functions significant (likes + shares comments)/views. <title>RESULTS</title> Compared model, model exhibited higher (83%-100 % vs. 24%-88 %) feature identification. Notably, backgrounds cars, private spaces, or shops demonstrated significantly than public spaces. Moreover, featuring smoking vaping, vape emojis, funny silly content heightened engagement. Conversely, e-cigarettes experienced lower <title>CONCLUSIONS</title> like background settings, adult presence, emojis substantially enhance These insights offer valuable guidance designing compelling campaigns improve