A hybrid explainable machine learning and transformer-based framework for psychological stress detection in women’s social media narratives

The prevalence of psychological stress among women in digital social communities underscores the need for interpretable and accurate AI-driven diagnostic tools. This study proposes a hybrid stress detection framework that integrates lexical and semantic processing to classify stress-related content from female-oriented online platforms such as Reddit and X (formerly Twitter). The lexical stream applies TF-IDF vectorization coupled with a Random Forest classifier, enabling interpretability and effective keyword-based stress detection. Complementarily, the semantic stream utilizes a fine-tuned BERT model to capture deeper contextual and metaphorical cues often present in women’s stress narratives. Ensemble strategies—including late fusion (soft voting) and stacked generalization with logistic regression—were employed to combine the strengths of both streams. The model was evaluated on a manually curated and balanced dataset of 1000 labeled posts. Experimental results show that while individual models performed well (TF-IDF + RF: 97% accuracy, BERT: 99% accuracy), the hybrid ensemble model achieved the best performance with 99% accuracy and 1.00 ROC-AUC, outperforming state-of-the-art baselines. The proposed framework demonstrates strong potential for real-world deployment in mental health monitoring systems, particularly for gender-sensitive applications. It offers a scalable, explainable, and highly accurate solution, contributing to the growing field of affective computing and digital mental health.