Leveraging Emotions for Enhanced Mental Health Prediction Using Machine Learning Algorithms

Mental health disorders are a growing global concern, yet early detection remains difficult due to stigma and limited awareness. This study investigates the prediction of mental health issues using a text-based dataset of user-generated content labeled as poisonous or non-poisonous. Classical machine learning models, including Random Forest, Naïve Bayes, and LSTM, were evaluated on the baseline dataset. To improve predictive performance, an emotion-augmented dataset was introduced, incorporating selected emotion features to capture nuanced sentiment variations. A fine-tuned RoBERTa model was then developed, leveraging these emotion features to enhance classification accuracy. The proposed model achieved an F1-Score of 0.92, significantly outperforming both baseline models and traditional machine learning techniques. These findings highlight the potential of integrating emotion-aware features to improve mental health detection and offer a scalable, data-driven approach for early intervention, particularly in resource-constrained settings. By leveraging natural language processing and deep learning, this work contributes to advancing automated mental health assessment, aiding in timely and effective support for individuals at risk.