Identifying Nonsuicidal Self-Injury Subgroups on Social Media: A Machine Learning Analysis of Post Features and Usage Patterns

Objective: Adolescent non-suicidal self-injury (NSSI) is widely discussed on social media, highlighting the need for early and targeted online interventions. This study developed a typology of social media users at risk for NSSI by jointly modelling post-level textual features and platform usage patterns using unsupervised machine learning.Methods: Public posts and metadata were collected from X, formerly Twitter using the Korean keyword ‘jahae’ (self-injury) over a 1-year period. After preprocessing, transformer-based embeddings were applied for natural language processing and clustering was conducted at the post and user levels.Results: Twelve content categories were identified from five domains: everyday interaction, sharing interests, community engagement, stress expression, and affective expression. These categories include family-, academic-, diet-, and sleep-related stress; suicidal ideation; internal and external negative affect; self-denigration; and self-harm positive expressions. For usage behavior, six indicators were examined follower-to-following ratio, user-reaction-to-response ratio, proportion of media posts, average post length, average posting time, and salient hashtags. Based on these textual and behavioral features, 1,224 users with NSSI-related posting activity were clustered into five subgroups: online communication seekers, negative emotion and stress expressers, online relationship immersers, internal self-injury ideators, and active self-injury posters. The subgroups showed distinct content signals and usage patterns.Conclusion: These findings provide a framework for characterizing users who display NSSI-related risk signals on social media and offer implications for prevention and intervention. The proposed typology may inform research and support professionals working with adolescents in digital contexts.