Advancing Mental Health AI through Synthetic Data Generation and Curriculum Training

Overview

This project addresses the increasing demand for mental health support using advanced AI technology, focusing on:

1. Generating synthetic datasets for AI model training
2. Developing a context-aware AI model for efficient screening and support

Background

• Rising number of mental health help seekers
• Limited availability of medical practitioners
• Need for efficient screening of distress candidates

Challenges

1. Lack of large, diverse datasets
2. Developing AI models responsive to context-aware situations

Our Solution

Synthetic Dataset Generation

• Utilizing OpenAI's GPT-4 and Nemotron models
• Validation from medical experts

Context-Aware AI Model

• Curriculum-inspired AI summarizer model
• Extracts relevant diagnostic features from input text

Methodology

1. Synthetic data generation
2. Real-world data collection from Reddit forums
3. Medical expert evaluation and annotation
4. Fine-tuning classifier and summarizer models
5. Comparative analysis of model performance

Key Findings

• Fine-tuned models trained on merged datasets (synthetic + annotated) perform better
• Summarizer model improved classification accuracy by 5% for real-world data

Impact

This project is a step towards developing an AI assistant to:

• Screen large volumes of submissions from distress individuals
• Facilitate connections between needy individuals and medical experts

Resources

Models and pruned datasets are freely available for the research community.

Keywords

Generative AI, GPT, Synthetic Data, Domain Expert, Curriculum-based, Summarizer model, BERT, Context-Aware, Mental Health, Distress

---

We welcome contributions and feedback from the community to further improve this important work in mental health support.