A Feature Fusion Machine Learning Framework for Early Prediction of Mental Health Disorders Using Textual Data

Mental health disorders have become a major global concern, with increasing prevalence and limited access to early diagnosis, particularly in low- and middle-income countries. Traditional approaches to mental health assessment are often delayed, subjective, and resource-intensive, making early detection difficult. Recent advances in machine learning and natural language processing have shown potential for analyzing textual data from digital platforms to identify early signs of mental health conditions. However, many existing approaches rely on single-feature representations, limiting their ability to capture the full range of linguistic patterns associated with psychological distress. This study presents a feature fusion machine learning framework for the early prediction of mental health disorders using textual data. The proposed system utilizes a labeled dataset obtained from Kaggle and applies preprocessing techniques including text cleaning, normalization, and tokenization. Feature extraction is performed using term frequency-inverse document frequency (TF-IDF) and n-gram representations, which are then combined through a feature fusion mechanism to produce a richer and more comprehensive input representation. The system is implemented using ML.NET, providing a scalable and efficient environment for model development and evaluation. Model performance is assessed using accuracy, precision, recall, F1 score, and area under the curve (AUC). Experimental results show that the proposed feature fusion system achieves an accuracy of 91.2%, precision of 90.6%, recall of 90.9%, F1 score of 90.7%, and an AUC of 0.963, outperforming an existing convolutional neural network-based model that achieved an accuracy of 87.96% and an AUC of 0.951. The results demonstrate that combining multiple statistical feature representations improves classification performance for mental health prediction while maintaining computational efficiency, without requiring the large datasets or computational resources typically demanded by deep learning models. The findings provide an empirical foundation for subsequent research extending this framework toward multimodal and transformer-based architectures for mental health prediction.