Sentiment trajectory modeling in mental health discussions via transformer-based analysis.

Introduction: Social media platforms produce a constant stream of user-generated text, which holds promise as a source of useful information regarding population-level and individual mental health states. However, existing methods for sentiment analysis on social media texts process these texts as individual, static instances, failing to consider the underlying dynamics of emotional states, which are naturally evolving in nature. Methods: In this paper, we introduce a novel framework for Temporal Sentiment Progression Analysis, which uses domain-specific transformer architectures to reconstruct the entire emotional evolution of a user over a given period of time. This framework captures important aspects of emotional evolution, including variability, volatility, and key emotional inflection points in a user's emotional progression. Our approach performs per-comment sentiment and thematic classification using transformer models, followed by post hoc statistical analysis to examine temporal patterns in user discussions. The framework was evaluated on a dataset of approximately 4,000 Reddit comments collected from eight mental health-related subreddits, with additional synthetic samples used only during training to address class imbalance. Performance evaluation was conducted exclusively on authentic Reddit comments and further validated using external Reddit-based mental health datasets and comparative baseline experiments. Results: Our experimental results on a large-scale social media dataset reveal different sentiment progression archetypes, which are strongly correlated with self-reported mental health concerns. Our proposed framework achieves a high classification accuracy of 92%. Discussion: Our approach shifts from static sentiment analysis to dynamic sentiment progression analysis, which can help in understanding the evolution of emotional distress in a more nuanced manner and can be useful in the development of context-aware interventions for mitigating mental health concerns.