Abstract
Anxiety disorders, emotional conditions, and stress are becoming an increasing tendency in modern society, which makes digital mental health technologies that provide support in these cases highly demanded. This paper focuses on proposing an AI-driven framework that combines Natural Language Processing (NLP) with privacy-protective systems to recognize emotions in the text of the user. The AES-256 encryption, Supabase authentication, and role-based access control are used to secure data security. It is presented as a multi-label problem of emotion classification, which is fine-tuned on the BERT-base-uncased and ModernBERT Large transformer. ModernBERT exhibited rapid convergence and enhanced situational sensitivity especially when it comes to sarcasm and mixed emotional states. Key contributions include:•Planning an emotion feedback loop offering contextual response and customized mindfulness teaching.•Performing a Class imbalance analysis of emotions and trade-offs in precision and recall.•Illustrating privacy-first architecture that is able to integrate with digital-therapy and clinician assisted decision software.