Clustering digital mental health perceptions using transformer-based models

Abstract

The rise in online mental health discussions underscores the need to understand diverse perspectives to inform targeted interventions. Addressing the granularity of existing research on mapping such perspectives, this study proposes a model combining contextual and sentiment analysis, keyword scoring, and clustering techniques to identify themes in online comments. Using transformer-based models (BERT, ALBERT, ELECTRA), the study achieved high-quality clustering of seven distinct mental health perspectives: stigma, empowerment, treatment approaches, recovery, social/environmental factors, advocacy, and cultural dimensions. ELECTRA outperformed others in clustering quality (silhouette score: 0.73; Davies-Bouldin Index: 0.34). The findings reveal cohesive, well-separated clusters that enhance understanding of digital mental health discourse. These insights provide a foundation for data-driven advocacy, tailored interventions, and broader awareness, addressing the complex dynamics of mental health narratives in online spaces. This study bridges a critical research gap by offering a systematic approach to analysing and interpreting mental health perspectives in digital environments.