Framing Pro-Anorexia Discourse on YouTube in South Korea: Social Network and Exponential Random Graph Model Analysis of Video Communities.

Background: YouTube, as a participatory platform, allows algorithmic curation and user engagement to shape health information flows. This dynamic amplifies and isolates harmful narratives, producing enclosed "refracted publics." Pro-anorexia (pro-ana) content exemplifies this, glamorizing extreme thinness as self-control and promoting disordered eating while distancing viewers from evidence-based health discourse. Despite concerns about public health consequences, few studies have examined how channel characteristics and framing strategies drive engagement and echo chamber formation. Objective: This study analyzed how pro-ana discourse circulates on YouTube and identified entry points for intervention. It (1) compared pro-ana, anti-pro-ana, and recovery frames by channel type (institutional vs individual) and subscriber scale (mega [≥1,000,000], meso [100,000-999,999], or micro [10,000-99,999]); (2) constructed video-level networks from overlapping commenters to trace discourse clustering over time; and (3) evaluated how frame and channel attributes shape network connectivity and frame-crossing. Methods: We collected 489 Korean-language YouTube videos (January 2020-August 2024) using pro-ana keywords and related-video crawling. Each was coded into 3 frames. Channels were classified by operator type and subscriber scale. A weighted commenter-overlap network was built, backbone-extracted, and analyzed through social network analysis. An exponential random graph model tested predictors of intervideo ties. Results: Of 489 videos, 369 (75.5%) promoted pro-ana content, mostly in micro- and meso-individual channels (361/369, 97.8%). Recovery frames appeared mainly in micro-individual channels (47/52, 90%). Anti-pro-ana frames clustered in institutional channels (46/68, 68%) but remained structurally peripheral. After backbone filtering, the commenter-overlap network comprised 435 videos and 906 edges. The network showed low density (0.96%) and moderate modularity (0.58) with 19 communities, indicating a fragmented, echo chamber-like structure. Longitudinally, density declined while modularity rose, consistent with echo-chamber intensification. The largest community (n=242) was predominantly pro-ana, while recovery-only (n=25) and anti-pro-ana (n=13) communities were peripheral. The exponential random graph model revealed that homophily was strong for anti-pro-ana (odds ratio [OR]=2.20; P<.001) and became significant for pro-ana with channel×frame interactions (OR=3.12; P<.001). Critically, recovery frames in meso-individual channels were associated with greater intervideo connectivity (OR=2.59; P<.001), and anti-pro-ana frames in those channels were also connective (OR=1.55; P=.003). Conclusions: Pro-ana discourse on YouTube is fragmented into "refracted publics." Meso-level individual channels using recovery narratives serve as critical bridges across polarized frames. Public health efforts should partner with mid-tier influencers to co-create emotionally resonant recovery content that disrupts echo chambers and fosters cross-frame dialogue.