AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

2

Abstract This capstone project explores the application of machine learning to predict early mental health risks using a large-scale structured survey dataset. We analyzed over 260,000 responses to identify behavioral, psychological, and demographic indicators associated with mental health vulnerability. After extensive data cleaning, unsupervised clustering was used to generate a three-level risk target (Low, Medium, High). We trained and compared three models: Logistic Regression, a Tabular Neural Network (TNN), and a Soft Voting Ensemble that combined both. Logistic Regression offered interpretability, while TNN captured complex, non-linear relationships. The ensemble improved prediction stability by averaging probabilities. The TNN achieved the highest performance, with 79% accuracy, a macro F1-score of 0.78, and an AUC of 0.94. Our results demonstrate that key features, such as stress, coping ability, and family history, can effectively predict mental health risk. A Gradio interface was developed for usability testing, with future goals including model explainability, real-time deployment, and ethical safeguards. This work supports proactive mental health screening in educational and workplace settings.

Keywords: mental health, machine learning, logistic regression, TabNet, soft voting ensemble, risk prediction, psychological assessment, structured survey data

202