Abstract:

We propose a real-time system for monitoring population-level mental health trends by analyzing social media text with transformer-based models. The pipeline continuously ingests posts from platforms such as Twitter and Reddit, applies preprocessing and normalization, and classifies content for indicators of depression and anxiety using fine-tuned transformer encoders (BERT, RoBERTa, DistilBERT). An ensemble of these models achieves high performance (90% accuracy, F1 0.89), surpassing recent baselines. The system maintains low inference latency ( 120 ms per post), making it suitable for continuous, high-volume monitoring. We benchmark against traditional baselines (TF-IDF + logistic regression, CNN) and provide error analysis highlighting common misclassifications. Ethical considerations, including privacy, consent, bias, and potential harms, are explicitly addressed to ensure responsible use. While current work is limited to English text, our framework is extensible to multilingual and multimodal data (e.g., text + images), with future directions including temporal modeling and clinical validation. These results demonstrate the feasibility of transformer ensembles for efficient, real-time mental health surveillance on social media.