DOI: 10.18178/wcse.2025.06.007
NeuroPulse: Transformer-Based Sentiment and Topic Modeling for Psychological Risk Assessment
Abstract— NeuroPulse is a Streamlit-based web platform that analyzes Reddit posts to assess psychological risk factors using natural language processing. The system integrates transformer-based sentiment analysis, neural topic modeling, and machine learning classification to identify and explain mental health-related content. We fine-tune a DistilBERT model for sentiment detection, with fallbacks to lexicon-based methods (TextBlob and rule-based heuristics) for robustness. We apply BERTopic – a BERT-based topic modeling technique – to uncover themes in posts, and use logistic regression to classify posts into mental health categories (stress, depression, bipolar, ADHD, anxiety). The platform provides real-time analysis with SHAP explainability, highlighting which words and features contribute to model predictions. Our results show that transformer models improve sentiment accuracy and topic coherence over baseline methods (TextBlob, NMF), and the combined framework offers interpretable insights into online mental health discussions. We present key visualizations including sentiment distributions, topic-category relationships, and SHAP explanations, demonstrating NeuroPulse’s capability to support mental health risk assessment from social media text.
Index Terms— mental health; social media; sentiment analysis; topic modeling; explainable AI; BERT
Gokul Srinath Seetha Ram, Rashmi Elavazhagan, Lan Yang
California State Polytechnic University, USA
Cite:Gokul Srinath Seetha Ram, Rashmi Elavazhagan, Lan Yang, "NeuroPulse: Transformer-Based Sentiment and Topic Modeling for Psychological Risk Assessment", 2025 the 15th International Workshop on Computer Science and Engineering (WCSE 2025), pp. 37-44, Jeju Island, South Korea, June 28-30, 2025.