Mental well-being is a cornerstone of sustainable societal development. Online platforms offer an unprecedented lens into individuals’ emotional states, providing opportunities to design intelligent systems for early detection and support of mental health issues. However, current ML/DL models lack transparency, and lexicon-based approaches often fail to capture nuanced user intent. We present TRUST-MH, a Transparent and Responsible User-level Semantic Tagging framework for Mental Health assessment, combining structured knowledge bases (DepressionFeature Ontology, UMLS) and commonsense reasoning (COMET) with deep learning architectures. TRUST-MH enhances both interpretability and accuracy, outperforming leading baselines (like BERT and MentalBERT) on benchmark datasets (CLEF e-Risk and PRIMATE). Most importantly, its recommendations align with socially grounded principles, producing insights that are interpretable by mental health professionals (MHPs) and empowering data-driven policies for promoting psychological well-being. TRUST-MH thus advances the broader vision of recommender systems designed not just for personalization, but for sustainability and social good.