Machine Learning-Based Mental Health Classification System: Design, Implementation, and Evaluation

Abstract

Mental disorders such as Bipolar Type-1, Bipolar Type-2, and Depression continue to affect millions of people worldwide, yet early and accurate diaganosis is challenging due to stigma, limited resources, and the subjectivity of self-reporting. Trying to bridge this gap, this project sought to develop a mental health diagnosis system that possesses the ability to classify individuals into Bipolar Type-1, Bipolar Type-2, Depression, or Normal states based on organized user input. Utilizing data acquired from an online repository, the system was designed with careful data cleaning, pre-processing, and class balancing using SMOTE for equal representation. Logistic Regression, Decision Tree, and Random Forest models were trained individually and then ensemble together using both hard and soft voting ensemble methods to obtain more stable predictions. The final ensemble model outperformed the individual models with accuracy up to 80%. This solution was deployed as a simple web app where users are able to answer a few guided questions and receive AI-generated feedback about their possible mental state instantly. The project demonstrates that the application of an ensemble of machine learning models will enhance early mental health screening and provide a supportive, accessible tool that will encourage individuals to seek professional help when needed.