MindCanvas: A Multimodal AI-Powered Mental Health Evaluation Application
Empower Minds: A Multimodal AI-Powered Mental Health Evaluation Application
Description
This project aims to revolutionzie youth mental health support by integrating advanced AI with psychoanalytic theary to enhance the capacity of educators and mental health professionals to intervene with childrens's condition at an early stage in an affordable, scalable and professional manner.
Problem & Motivation
The gap between accessible and affordable mental health services and the increasing prevalence of mental health issues among youth, especially ineducational settings, underscores the urgent need for innovative solutions. Our project is motivated by the desire to bridge this gap and offer early, insightful, and accessible psychological assessments based on the Kinetic-House-Tree-Person drawing theories through leveraging image classification and generative AI models.
Data Source & Data Science Approach
Our datasets include case studies from an authorized projective tests textbook and kids drawings from partnering schools. Two layers of models are appied in our application with the first one to perform a multi-label classification task and the second one to provide mental health analysis based on the labels identified from the first layer.
To tackle the multi-label classification problem, extensive exploratory data analysis (EDA) was conducted to optimize label selection. We took the ensemble approach to optimize the results from both CNN image classification models and the LLM (Large Language Model) to identify the labels. The second layer model is enabled by a customized GPT-4 assistant that retrieves knowledge from pre-uploaded documents and outputs text in the desired format based on the prompts we crafted. We used metrics Accuracy, Recall, Precision, F1-Score, and Hamming Loss for the multi-label classification model evaluation.
Key Learnings & Impact
Our project demonstrates the promising future of integrating AI in assistance with mental health assessment to improve accessibility, affordability, and personalization of care. In such a project, data collection and data quality played significant roles: our model performance can be largedly improved if given sufficient time to gather better quality data. In addition, the future opportunity of collaboration with school psychologists would make the product more feasible and practical.
Acknowledgements
We extend our gratitude to the educational institutions, mental health professionals, and AI research community for their support and insights. Special thanks to our mentors and advisors for their invaluable guidance throughout this project.
