Depression is a prevalent mental disorder that significantly affects the well-being and normal functioning of individuals. The widespread use of social media platforms such as Twitter (then X), Reddit, and Facebook provides an unprecedented opportunity to study user-generated textual content in the context of initial identification of depressive symptomatology. This paper proposes a holistic, data-driven approach to automated depression detection that combines traditional machine learning and modern deep learning techniques. The multi-platform data were subjected to stringent pre-processing, including text normalization, tokenisation, lemmatisation, and random oversampling to correct noise, heterogeneity, and class imbalance, ensuring healthy and reliable model training. The proposed models were compared systematically to a set of other traditional machine learning, deep learning, and attention-based baselines in the same setting of experiments, with the assessment of performance checked with standard evaluation metrics and visual analysis to provide an unbiased and strong performance measurement. The GRU-Attention model was shown to exhibit strong capability for long-range contextual interactions and emotionally significant tokens, achieving higher results than most non-transformer-based models. To boost predictive stability, generalization, and robustness, the results of the most successful models were stacked using a simple averaging method, yielding a final accuracy of 94.74%, with equal precision, recall, and F1-score. Such a wide-scale analysis, including confusion matrices and AUC-ROC curves, validated the effectiveness of the framework in minimizing misclassifications and identifying subtle linguistic patterns. The proposed system is a scalable, interpretable, and non-invasive system to detect depression at an early stage, which has a vast potential for being applied in digital mental health monitoring and intervention approaches.
Arshad et al. (Thu,) studied this question.