JMIR Publications

ABSTRACT

Background:

Online youth forums offer valuable insights into how cultural identities, relationships, and social contexts shape young people’s mental health. Analyzing these data at scale requires a hybrid approach that combines computational methods with youth’s contextual experiences. Hybrid intelligence brings computational analysis together with researchers’ and youth’s perspectives, while intersectionality theory helps explain how identities and systems of power intersect in these contexts.

Objective:

This study aimed to identify youth mental health needs and challenges in a Dutch online peer support forum through an intersectionality-guided analysis that combined researchers, large language models (LLMs), and youth co-analysis, and to examine the practical contributions and limitations of this hybrid workflow.

Methods:

We collected and analyzed a corpus of 37,691 posts (including replies) from 16 sub-forums on De Kindertelefoon.nl, the largest and only publicly available Dutch youth forum. BERTopic produced 200 topic clusters; screening removed 36 low-signal clusters, leaving 164 for analysis. Researchers annotated 40 clusters across four dimensions: overarching topic, emotional stages, system touchpoints, and identity characteristics. Two non-overlapping sets of 20 clusters supported few-shot prompting and prompt testing with Microsoft Copilot (GPT-4o model). Two co-analysis workshops involved 22 youth aged 12–23 years from diverse backgrounds. Participants analyzed selected forum posts and compared their interpretations with LLM annotations. Researchers synthesized workshop recordings, facilitator notes, group artifacts, and LLM outputs to refine final annotations.

Results:

The screened 164 clusters with 22,314 posts were organized into 10 thematic groups, with 31 sub-themes. “Romantic relationships” was the most prevalent theme, accounting for 9,873 posts (44.2%), followed by “Identity, expression, and belonging” (3,626 posts; 16.2%) and “Feelings, coping, and care” (2,852 posts; 12.8%). Across themes, we identified intersectional themes of mental health challenges linked to youth’s cultural identities, the people around them, and the institutions and services. Youth contributions connected emotional experiences to specific actors and situations and revised overly formal and difficult language. Practical limitations of LLM annotation included row misalignment, failed large-batch processing, and relatively low alignment with the researcher’s annotation on emotional stages. Workshop observations supported the value of relationship building, use of accessible materials, and facilitated pair work.

Conclusions:

Youth mental health needs in this forum extended across identity, relationships, and institutional contexts. Combining computational analysis with youth participation supported contextual interpretation and identified opportunities for more inclusive digital support. The workflow provides exploratory methodological evidence, and further evaluation is needed to assess its accuracy and transferability in other datasets and populations.