Abstract: The prevalent global mental health illnesses have become a pressing societal challenge, garnering substantial attention from both academic researchers and industry practitioners. Despite demonstrating promising diagnostic performance, current deep learning applications in mental health monitoring face fundamental limitations in clinical translation, particularly concerning their opaque decision-making processes and inability to elucidate underlying pathological mechanisms. To address these critical gaps, we introduce the Dual-level Rule Learning (DRL) framework - a novel paradigm for trustworthy mental health monitoring that simultaneously addresses decision-level and recognition-level explainability through psychologically-grounded symbolic reasoning. Our framework innovatively combines domain expertise from established psychological concepts ($e.g.$, PANAS scales) with LLMs' assistance for generating human-interpretable logical predicates that bridge symptom patterns with underlying etiological factors. Specifically, the cognition module leverages LLMs' contextual understanding to identify potential psychopathological factors during model training, while the decision module synthesizes these insights into generalizable logical rules through predicate aggregation, thereby ensuring both monitoring transparency and etiological plausibility. Comprehensive experimental results demonstrate the effectiveness and trustworthiness of our proposed framework.