MindBridge: A Transformer-Based Conversational Agent for Real-Time Depression and Anxiety Detection and Support in University Students

Abstract

Mental health challenges are becoming a major concern among university students who still have limited access to appropriate and affordable support due to barriers such as stigma, cost, and time limitations. Most of the digital tools available, such as Wysa and Woebot, are based on rigid rule-based logic and are incapable of understanding the contextual nuances of student-expressed emotional distress. For this system, a system-oriented, agile research methodology was used. Three public datasets the Mental Health Text Classification Dataset, the Reddit Depression Dataset, and the EmpatheticDialogues Dataset, were combined and balanced to create a training corpus of 24,000 labeled entries equally distributed across three emotional categories: depressive, anxious, and neutral. The Mental-BERT transformer model was trained on this corpus in five epochs with the help of Google Colab's GPU infrastructure. This classification model was then incorporated into a Flask web backend, together with the Groq API for conversational response generation, a PostgreSQL relational database for secure data persistence, and a web-based frontend consisting of a responsive chat interface with user authentication. After evaluating the fine-tuned Mental-BERT model on a test set of 3,600 samples that were not used during training, the overall classification accuracy was 96%. Also, the macro-averaged precision, recall, and F1-score were equally high at 96% across the three classes. We carried out a user study with 58 university students who gave a mean System Usability Scale score of 76.55 for MindBridge. That score belongs to the good usability category and is also above the industry average threshold of 68. More than 70% of participants graded the system as Good or Excellent. Empathy and conversational quality metrics showed that the respondents considered MindBridge a source of empathy and assistance, while the mean figure for perceived usefulness was 4.03 out of 5. All 16 automated system tests were successfully passed. Results show that using a domain-specific pre-trained transformer model alongside prompt-engineered large language model response generation and a hybrid safety architecture can lead to the development of a very accurate, usable, and empathetic mental health support system for university students.